What is a scheduled AI agent?

A scheduled AI agent is an AI assistant that runs on a timer instead of waiting for you to type. At a set time, such as 8 AM on weekdays, it starts a fresh session, reads the tools it's been given (email, calendar, tickets, a code repository), and leaves a result such as a summary, a draft reply or a pull request. ChatGPT, Gemini, Claude Duet, Claude Code's /schedule, VS Code Automations and n8n's Schedule Trigger all offer versions of it. As of November 2026, VS Code 1.145 (November 4) supports custom cron schedules for local Automations.

How do I schedule a task in Claude Code?

Type /schedule followed by a plain-language request, for example /schedule every weekday at 8am, summarize yesterday's merged pull requests, and use /schedule list to see your tasks. Cloud scheduled tasks run in Anthropic's cloud on Pro, Max, Ultra, Team and Enterprise plans, with an hourly minimum, and plain cron lines run in UTC unless you add CRON_TZ=. Since Claude Code 2.1.317 (November 3, 2026), /schedule warns when a new task gets connectors you didn't ask for. Since 2.1.318 (November 4), claude auth status shows when your sign-in expires.

What is the Claude Ultra plan?

Ultra is the new name for Claude's former Max 20x plan, renamed October 30, 2026. It costs $200 a month, gives 10 times Pro's weekly usage, and verified Ultra accounts have no five-hour session limit, though weekly limits still apply. The former Max 5x plan is now just called Max, at $100 a month. To check whether your Ultra account is verified, open Settings › Usage: no "Current session" bar means it is.

What did Anthropic's amended S-1 show?

Anthropic filed an amended S-1 on November 2, 2026. It reported third-quarter revenue of $16.96 billion, up 47% on the quarter, nine-month revenue of $33.2 billion, and $895.4 billion in compute and other commitments. It swung to an operating loss of roughly half a billion dollars after a $974 million operating profit in Q2. It warned that Anthropic "may imminently meet" its Responsible Scaling Policy's threshold for "dramatically accelerated progress attributable to AI." No price range has been set; Bloomberg reports the roadshow could begin the week of November 9.

How did Ohio vote on data centers in 2026?

On November 3, 2026, Ohio held 21 of the roughly 41 local data-center ballot measures in the US, and every Ohio result with a reported count went against data centers. Trenton's Issue 17, a ban on data centers over 25 megawatts, passed with 79.7% (unofficial), and Wilmington voters rejected the rezoning of 545 acres for an AWS site. Bans also passed in Piqua, Urbana, Sunbury, Granville and other cities. Governor-elect Amy Acton campaigned on a conditional moratorium on new data centers.

GPT-6 Cyber is OpenAI's cybersecurity model, launched November 5, 2026, and the first model OpenAI rates "Critical" for cyber under its Preparedness Framework. OpenAI says it found 81 previously unknown high-severity Firefox flaws, built 948 working Linux kernel exploits and scored 48.5% on ExploitGym (self-reported). Access is limited to approved members of OpenAI's Daybreak Red program through the API and Codex.

Did Anthropic ban reverse engineering in Claude?

No. As of November 5, 2026, Anthropic hasn't announced any new policy on reverse engineering or on the open-source REA ("Reverse Engineer Anything") tool. What changed is narrower. On October 30, 2026, Anthropic confirmed it had added two sentences to Claude Code's built-in prompt for Opus 5.5 telling Claude not to recreate a specific existing game, citing its usage policy on intellectual property. In Claude Code 2.1.317 (announced November 4), it loosened the rule so you can name a game to describe the mechanics you want. Separate refusals on reverse-engineering tasks come from Opus 5.5's cyber safeguards, announced in September and October. Researchers can apply to Anthropic's Cyber Verification Program, whose Defense Access tier covers reverse-engineering malware.

Ripple-65B Preview is an open-weight AI model released on November 4, 2026 by Ricursive Intelligence, a chip-design AI startup founded by the creators of Google's AlphaChip. It's a 65-billion-parameter dense model trained on 12 trillion tokens, licensed under Apache 2.0, with a context window of up to 1 million tokens. Ricursive reports a Terminal-Bench 2.1 score of 80.52, the highest among the US open models it compared (self-reported). Several Chinese open models report higher scores. It cost $32 million in total R&D, according to the company.

What is weight-space programming?

Weight-space programming is the idea that an AI model can read and edit another neural network's weights, the raw numbers that store what it knows, the way a programmer reads and patches code, instead of retraining it. On November 2, 2026, MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published "Sparks of Weight-Space Programming," in which OpenAI's GPT-6 Astra traced the endless reasoning loops of the small open model Qwen3.5-0.8B to one direction written by layers 15 and 19 and subtracted it from 0.016% of the weights. With no training, its MATH-500 score rose from 7.9% to 48.1% and its IFEval score from 21.0% to 62.3%. It failed other tests: it couldn't tell from the weights alone which of two small models had been tuned to be a people-pleaser. It's a blog post with full logs, not yet a peer-reviewed paper.

What is ASD-STE100, and why add it to a prompt?

ASD-STE100 is Simplified Technical English, a controlled-language standard owned by ASD, the Aerospace, Security and Defence Industries Association of Europe. It was first released in 1986 for aircraft maintenance manuals, and the current edition, Issue 9, came out on January 15, 2025, free on request from asd-ste100.org. It has 53 writing rules and a dictionary of about 900 approved words, each with one meaning: at most 20 words per instruction and 25 per descriptive sentence, one instruction per sentence, commands, active voice, and no -ing verbs, phrasal verbs or contractions. Adding "Write the answer in ASD-STE100" to a prompt makes an AI model answer in short, numbered, unhedged steps. Andrej Karpathy suggested asking for "80% of the way to ASD-STE100" because the full standard is strict. Models only approximate it, and ASD warns that AI text can look compliant when it isn't, so it is a clarity tool, not certified compliance (story 15).

When is the next CincyAI for Humans meetup?

The next CincyAI for Humans meetup, the last of 2026, is Tuesday, December 1, from 3 to 5 PM at the University of Cincinnati's Digital Futures building, Room 140, 3080 Exploration Ave, with a building tour at 2:30 PM for the first 15. It's free with RSVP on Eventbrite. Before then, the free Cincinnati AI in Healthcare Summit is Wednesday, November 11, at UC's 1819 Innovation Hub, and Xavier's free AI and Ethics panel is Thursday, November 19, from 6 to 8 PM.

AI Weekly — November 6, 2026 · Issue No. 31 | Scheduled AI agents, AI on the ballot, Anthropic S-1/A, an AI that fixed another AI's weights, ASD-STE100

01 · Feature · Scheduled agentsAll users · Power users · Developers

Your AI, on a timer. A practical guide to scheduled agents: what runs where, a five-minute starter, and the controls you'll want. #

October 30–November 5, 2026 · VS Code 1.145 November 4 · Claude Code 2.1.316–2.1.318 October 31–November 4 · Duet reviews November 3 and 5 · LangChain async subagents November 5

Most people still use AI by opening a chat and typing. The useful agents this autumn don't wait to be asked. They run at 8 AM, check your inbox and your tickets, and leave a short report. This week made that easier and safer in three places: VS Code took cron schedules, Claude Code's scheduler got safer, and early Claude Duet users showed how a suggestion becomes a daily watch.

At a glance · scheduled agents this week

New: VS Code 1.145: five-field cron for local Automations (November 4)
Safer: Claude Code /schedule: connector warning, next-run time, sign-in expiry, failed runs marked failed
In practice: Claude Duet reviewers turn suggestions into daily watches; "burns through usage fast"
Trend: LangChain: 70%+ of Managed Deep Agents runs now start from a schedule (October 28)
Starter: A weekday 8 AM brief: read, summarize, draft. No sending, buying or deleting
Rule of thumb: Small, read-only, and checked by a person

What changed this week

VS Code 1.145 (November 4) lets local Automations, agent tasks VS Code runs on its own, take a custom five-field cron expression. "The dialog validates the expression and describes the resulting schedule. Cron schedules are not available for cloud automations." Automations can also use custom agents from your workspace.
Claude Code shipped three releases with scheduler changes. 2.1.317 (November 3) added a warning to /schedule "when a new cloud scheduled task gets connectors you did not ask for, with how to remove them." 2.1.318 (November 4) shows the next run time and link again after you create or update a task, adds the date your stored claude.ai sign-in ends to claude auth status (loginExpiresAt) and to claude -p JSON (credential_expires_at), and lets a cloud session start with No repository, which suits writing and research jobs. Since 2.1.316 (October 31), cloud sessions and scheduled tasks are marked failed when auto mode stops a turn, instead of looking finished.
Claude Duet, in practice. Issue 29 featured Duet and Issue 30 followed up on its first week. Anthropic announced nothing new for Duet this week (no waitlist change, pricing, Microsoft 365 connector or Team plan). But two reviews published in the window show the scheduling side. In Marketing Against the Grain's hands-on (November 5), Kipp Bodnar turns a Duet suggestion into "a daily watch" for online mentions of his book: any item in Duet's Tasks tab can be given a schedule. His caveat: it "burns through usage fast," and setup needs connectors plus the Claude Chrome extension. Skill Leap AI's tour (November 3) walks through the morning brief with drafted replies.

Context from just before this window: on October 28, LangChain wrote that more than 70% of runs on its Managed Deep Agents now come from schedules, not people typing. On November 5 it added async subagents, so a long-running agent can start background jobs and check on them later.

How a scheduled agent works

Every scheduler below follows these five steps; they differ in where step 2 runs and how step 3 is controlled. Diagram: The AI Enthusiast.

Who schedules what (checked November 5)

Rows marked background didn't change this week; they're here so you can compare.

Ten AI schedulers compared: what you can schedule, where it runs, limits, plans and this week's status
Tool	What you can schedule	Runs where	Cadence and limits	Plans	Status
ChatGPT scheduled tasks	Prompts, reminders, monitoring	OpenAI cloud	Up to hourly; unattended tasks may pause	Go, Plus, Pro, Business, Enterprise	Background (scheduled hub, June 17, 2026)
Gemini scheduled actions	Briefings, summaries	Google cloud	10 active actions	Google AI Pro/Ultra, Workspace business/edu	Background (June 2025)
Claude Duet	Morning brief, watches; any Task can get a schedule	Anthropic cloud + browser	Checks apps "every few hours"	US Pro, Max, Ultra (waitlist preview)	Reviews in window; no new features
Claude Code `/schedule` (cloud scheduled tasks)	A full agent run with repo or no repo, plus connectors; also API and GitHub triggers	Anthropic cloud	Hourly minimum; 100 runs/hour per account; crons in UTC unless `CRON_TZ=`	Pro, Max, Ultra, Team, Enterprise (research preview)	Changed October 31 – November 4
Claude Code `/loop` and Desktop tasks	Repeating prompts in a session; local tasks	Your machine	`/loop` from 1 minute, expires after 7 days	Any Claude Code plan	Background
VS Code Automations	Agent tasks with a workspace custom agent	Local (cron) or cloud (presets only)	Five-field cron, local only	Needs a GitHub Copilot plan	New November 4
GitHub Copilot app Automations	PR summaries, triage, team workflows	GitHub cloud sandbox	—	Paid Copilot; cloud sandbox billing starts in November	Background (Universe, October 28)
Cursor Automations	Agents on GitHub, Slack, webhook or cron triggers	Cursor cloud or your infra	—	Paid	Background
n8n Schedule Trigger + AI Agent node	Any workflow	Self-hosted or n8n Cloud	Cron down to seconds	Free self-hosted; Cloud from €20/month	Background; 3.1.0 pre-release November 3 (story 12)
LangChain Managed Deep Agents	Static cron, or schedules the agent sets itself	LangSmith	Cron with time zone	LangSmith	Schedules October 28; async subagents November 5

Starter: a weekday 8 AM brief, four ways

Keep the job small: read, summarize, draft. No sending, buying or deleting. The prompt, the same everywhere:

Every weekday morning: check my inbox and calendar for today. Give me at most
5 items: what needs a reply today, any meeting clash, anything someone is
waiting on me for. One line each, with a link. Draft replies but don't send
them. If nothing needs me, say "Nothing urgent."

ChatGPT or Gemini (no setup): paste the prompt and add "every weekday at 8 AM." Both create a scheduled task from plain language.
Claude Duet (if you're in the preview): ask for it in the main thread, or open Tasks, add it, and give it a schedule.
Claude Code (cloud): run the three commands below. Read the new connector warning, and remove anything the brief doesn't need. Plain cron lines run in UTC, so 8 AM in Cincinnati is 0 13 * * 1-5 in winter (or prefix CRON_TZ=America/New_York).
VS Code 1.145: Automations view › New local automation › Schedule › Cron › 7 8 * * 1-5 › pick a custom agent › Save. (We use :07 rather than :00, when everyone else's jobs fire.)

/schedule every weekday at 8am, [the prompt above]
/schedule list
/schedule why did my morning brief do nothing today?

Cron in one minute

Five fields: minute, hour, day of month, month, weekday (0 = Sunday).

The starter brief's schedule, field by field. Diagram: The AI Enthusiast.

Four common cron expressions and what they mean
Cron	Means
`7 8 * * 1-5`	8:07 AM, Monday–Friday
`0 /2 * *`	Every two hours, on the hour
`30 16 * * 5`	4:30 PM every Friday
`0 7 1 * *`	7 AM on the 1st of each month

Risks and controls

Risks of scheduled AI agents and the controls for each
Risk	Control
Too much access	Claude Code and Duet attach your connected connectors by default, and a scheduled run reads them with nobody watching. Give each task only the connectors it needs. That's exactly what the November 3 warning is for. Read-only where possible.
Prompt injection while you sleep	A scheduled agent reads emails, web pages and tickets written by strangers. Draft, don't send. Keep payments, deletions and outside messages behind approval. Duet never allows "don't ask again" for payments or deletions (Issue 30).
Silent failure	A job that dies looks the same as a quiet morning. End every prompt with "If nothing, say Nothing urgent," so silence means broken. In Claude Code, failed auto-mode runs are now marked failed (2.1.316), and `claude auth status` shows when your sign-in expires (2.1.318).
It burns your plan	Scheduled runs count against your usage, and reviewers say Duet uses it fast. Start daily, not hourly. Use a lower effort level for routine jobs. Check usage after the first week. GitHub begins billing Copilot cloud sandbox time in November.
Time zones	Claude Code's plain crons are UTC. Use `CRON_TZ=`, or convert, and remember daylight saving time ended November 1.
Drift	The same job quietly gets worse after a model update. Keep one or two known inputs and re-check the output when models change.

Best videos (all published in the window)

19:44

How I Use Claude Duet to Automate My ContentMarketing Against the Grain · Nov 5

12:36

Claude Duet Is Here: The End of Prompting?Skill Leap AI · Nov 3

20:01

Claude Code: Build Your First AI Agent in 20 Minutes (100% No Coding)Grace Leung · Oct 30 · ends with the agent running in the cloud on its own

What to do with this

Everyone: set up the weekday brief above in the AI tool you already pay for. Run it for a week before adding a second job.
Duet users: turn one suggestion you liked into a scheduled task, and watch your usage meter.
Developers: move one recurring chore (a nightly test-failure summary, a Monday dependency check) to /schedule or VS Code Automations, and trim its connectors.
n8n users: a Schedule Trigger feeding an AI Agent node does the same job, self-hosted, with every run in the execution log.
Store-support and IT teams: the best first jobs are the dull ones: a morning digest of yesterday's tickets, a weekly list of repeat issues, a Friday summary of what's still open. Small, read-only and checked by a person.

Sources

02 · PolicyAll users

AI was on the ballot, and it lost almost everywhere. Voters limited data centers, Democrats took Congress, and chip stocks had their worst day in 19 months. #

November 3–5, 2026

On Tuesday, November 3, AI's physical footprint was on the ballot. CNBC counted 41 local data-center measures in 10 states, 21 of them in Ohio. Voters backed restrictions or rejected pro-development measures in nearly every contest with results; the one exception was Harwood, North Dakota. In NBC's exit poll, 56% of voters said AI and data centers were a "crucial or major factor" in their vote for Congress, and 65% of those voters backed Democrats. Democrats won both chambers (CNBC projections: at least 250 House and 54 Senate seats) and flipped five governorships. Elon Musk spent at least $474M, and his candidates largely lost.

Filled dots: voters backed limits or rejected a pro-development measure, in contests with results. Data: CNBC, November 4–5. Chart: The AI Enthusiast.

What winners said

Texas Governor-elect Gina Hinojosa: "When I am governor, I will issue a moratorium on all data center approvals until laws can be passed." Texas holds about 27% of planned US data-center capacity.
Sen. Bernie Sanders (November 5): "National data center moratorium now."
Rep. Greg Casar: "AI company CEOs should be called to testify under oath."
AI super PACs had a mixed night (TechCrunch): pro-safety PACs won races in Michigan, Colorado and Wisconsin, while the a16z-backed Leading the Future largely sat out the general election.

The market's verdict (November 5)

The Nasdaq fell 5.52%, its worst day since April 2025, and the S&P 500 fell about 4.4%. Nvidia fell 9.3% to $181.60, erasing its gains for the year and about $450B in value. AMD fell 17%, two days after a record quarter (story 13). JPMorgan estimates 37.5 GW of planned 2028 capacity could face delays.

The market's interpretation is that AI infrastructure buildup will become slower and more expensive.An analyst, to Fox Business · November 5, 2026

The industry's response

On November 5, OpenAI published a "National Community Compact" for its data centers, which it now calls "compute campuses": it pledges to pay for its own energy, protect local water and other resources, alert communities early, publish regional jobs reports, and limit non-disclosure agreements with public officials. OpenAI's Chris Lehane had already pitched a "New Deal"-style AI social contract on November 4.

What it means for AI rules (analysis, labelled as such)

Law firm Foley and others now see federal preemption of state AI laws as far less likely, so the state patchwork stays.

Follow-up to Issue 30: responses to the House Democrats' request that labs report AI incidents within 48 hours are due November 11. From January, those ranking members become committee chairs.

Sources

03 · AnthropicAll users

The amended S-1: $17 billion in a quarter, $895 billion in commitments, and a warning about AI building AI. #

Filed November 2, 2026 · METR assessment November 5

Follow-up: Issue 30 covered Anthropic's October 26 S-1 filing. This week brought the third-quarter numbers.

On Monday, November 2, Anthropic filed an amended S-1 (S-1/A) with the SEC. It still has no share count or price range. Figures below are company-reported, as read by Dealroom, Yahoo Finance and TechCrunch.

Anthropic quarterly results from its amended S-1, as reported
Metric	Q3 2026	Q2 2026
Revenue	$16.96B (+47% on the quarter)	$11.56B (+144%)
Operating result	A loss of roughly $0.5B (outlets report $536M–$567M), blamed on about $5B of expensive short-term compute and $2B of non-cash donations	$974M profit
Net loss	$23.29B	$66.3B

Q1 is from the October 26 S-1 as reported in Issue 30; Q2 and Q3 from the November 2 S-1/A. Chart: The AI Enthusiast.

Nine-month revenue: $33.2B, against about $2.5B a year earlier.
Commitments: $895.4B in compute and other spending, of which $69.8B is due within 12 months. The largest suppliers are Broadcom, Google, SpaceX AI and Amazon.
Customers paying $100K+ a year: 11,900+, up from 1,500+ at the end of 2025. Cash and short-term investments: $131.8B.
The new risk factor (TechCrunch, November 2): see below.

Our internal estimates indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.'Anthropic S-1/A risk factor, via TechCrunch · November 2, 2026

The filing adds that Anthropic "may in the future not identify" such a leap "until after it has occurred."

An outside check

On November 5, the independent evaluator METR published its assessment. As of October 2026, it estimates Anthropic was improving capabilities on well-specified tasks at about 2x its historical rate (confidence interval 1.3–3x), with "a majority (perhaps ≈80%)" of that speed-up due to AI. Employees' self-reported productivity uplift rose from 1.2x (January 2025) to about 12x. METR notes Anthropic redacted much of the most informative evidence.

Timing and price

The Wall Street Journal reported (November 2–3) that institutional investors want a valuation of about $1.5T–$1.6T, down from about $2T discussed earlier. Bloomberg says the roadshow could start the week of November 9; November 16 is the earliest possible first trading day. Anthropic hasn't set a date.

Sources

04 · ClaudeAll users · Developers

Max 20x is now "Ultra," with no five-hour limit. Plus Personal Finance, enterprise controls, and a quiet watermark. #

October 30 – November 5, 2026

New plan names (October 30). Max 5x is now just Max ($100 a month, 5x Pro usage). Max 20x is now Ultra ($200 a month, 10x Pro usage per week), and "verified Ultra accounts don't have a five-hour limit"; weekly limits still apply. Verification is automatic, based on account and payment history. To check, open Settings › Usage: no "Current session" bar means you're verified.

Claude's renamed plans as of October 30, 2026
Old name	New name	Price	Usage	Five-hour limit
Max 5x	Max	$100/month	5x Pro	Yes
Max 20x	Ultra	$200/month	10x Pro per week	None for verified accounts (weekly limits still apply)

Personal Finance (beta, November 5)

Link bank, card, loan and investment accounts through Plaid, then ask questions in plain language: spending by category, duplicate charges, unexpected fees, bills that keep rising. Claude sees account names, the last four digits, balances, credit limits and up to two years of transactions. "Claude's access is read-only," and Personal Finance chats are excluded from model training. For Pro, Max and Ultra in the US, rolling out this week. The Duet launch post (October 21) had listed personal finance features as "coming soon"; Anthropic hasn't said whether Duet uses them. See tip 1.

Enterprise controls (November 5)

A Security Center (beta) compares an organization's settings with Anthropic's recommendations and produces an auditor report.
Inference hooks are now generally available (prompt hooks and pre-tool-call hooks that plug Claude into existing data-loss-prevention tools).
admin.claude.com manages several organizations in one place.
Admins can block personal Claude accounts, set an IP allowlist (up to 1,000 entries on the Console) and sign a member out everywhere.
EDR sensors can now run inside Cowork's virtual machine on macOS and Windows.

For developers

Message threads (public beta, November 2): the server keeps the conversation, so each request sends only new messages plus the previous response ID (header message-threads-2026-08-12). Anthropic reports about 35% lower median time to first token on 500K–1M-token requests (self-reported), with no price change. Threads expire five minutes after the last request.
Text watermark (November 3): text from Claude Mythos 5, Opus 4.7 and Sonnet 4.6 now carries Anthropic's statistical watermark on listed platforms. xAI turns on a similar watermark for the Grok API on November 7; both point at the EU AI Act's labelling rules.
Marketplace (November 4): 13 new partners, including Datadog, Asana, Glean, monday, Zoom and Runway, bring it to 25+; enterprises can spend part of their committed Anthropic budget there.
Opus 4.8 fast mode is deprecated and will be removed November 30. After that, requests error out rather than falling back.

Claude Duet: no official news this week (see the feature).

Sources

05 · Agent securityDevelopers · All users

A fake Claude Code extension planted SSH backdoors. And coding agents keep installing tools that were retired years ago. #

Disclosed October 29–30 · coverage through November 3, 2026

On October 29, Cursor pulled two versions of the Claude Code extension, 2.1.312 and 2.1.313, from its own extension marketplace. They were live from about 2:27 AM to 1:35 PM PT and "were not developed by Anthropic." Reports put the reach at up to about 5,000 users. Anthropic says an attacker "used access to Cursor's marketplace"; Microsoft's VS Code Marketplace and Anthropic's own channels weren't affected.

What the tampered extension did

Security firm Phorion's write-up (October 30) says the modified extension:

added an attacker's SSH key labelled svc-telemetry to ~/.ssh/authorized_keys;
opened a reverse SSH tunnel;
set up persistence: a fake macOS LaunchAgent com.apple.sysmond.plist, a Linux cron job or a Windows scheduled task;
collected the names of environment variables that look like keys and tokens.

Cursor's warning: "uninstalling the extension may not be sufficient." Check yourself with tip 9.

Agents pick stale tools

On November 3, Glow Labs reported "MCPhantom": in 98% of test sessions, coding agents (Claude Code, Cursor, Codex CLI, Claude Cowork) installed a deprecated MCP server they remembered from training, even after npm said it was deprecated (self-reported). Deprecated MCP packages have 17.3M downloads since July 2025, and Glow found four high-severity flaws, the worst a remote-code bug in the old Puppeteer server.

The model's memory is now a package manager with no deprecation feed.Ophir Arie, Glow Labs · November 3, 2026

Criminals use the models too

On October 31, Cisco Talos described VOLTA, a Windows infostealer that sends each victim's stolen data, including Amazon and DoorDash receipts mined from Gmail, to Claude Opus 4.6 (falling back to Kimi and Gemini if refused) to rank how much the victim is worth. "The model acts as intelligence analyst for the operator."

And a new off switch

On November 3, CrowdStrike added an AI Kill Switch to Falcon that can stop a rogue agent, block it from restarting and cut its network access. "Everyone is building AI agents, but they never had the off button, and now they do," said George Kurtz.

Sources

06 · Claude CodeDevelopers · All users

No, Anthropic didn't ban reverse engineering. It quietly told Claude to stop cloning existing games, then loosened the rule five days later. #

Rule found and confirmed October 30 · loosened in Claude Code 2.1.317, announced November 4, 2026

For two weeks a claim has gone around that Anthropic "killed REA." REA ("Reverse Engineer Anything") is a free, MIT-licensed tool that connects coding agents (Claude Code, Codex, Cursor, Gemini CLI and others) to reverse-engineering programs such as Ghidra, IDA and Hopper, so an agent can take apart a program, an Electron app or a website. It has more than 434,000 GitHub stars and logged about 511,000 npm downloads in the week to November 4, so the "ban" hasn't slowed it. It also fed a wave of AI-built ports and browser remakes of commercial games.

Background: in early October ZeniMax objected to a "Fallout: New York" browser game, and Activision sent takedown notices over AI-built Call of Duty browser ports. On October 30, Scientific American ran an interview with REA's pseudonymous developer. Issue 28 featured REA.

What actually changed

Timeline of Claude Code's game-cloning rule
Date	What happened
October 29 (evening)	Claude Code 2.1.314 ships with a new instruction in its built-in prompt. It applies only when Claude Code runs Opus 5.5, and it isn't in the changelog.
October 30 (morning)	X user Chasm posts the rule: "Do build original games, games in a genre, and games from the user's own designs and assets. Don't build a recreation of a specific existing game, even if it is renamed or called original."
October 30 (afternoon)	Anthropic's Addy Osmani confirms it. An Anthropic spokesperson referred Game Developer to his post. GitHub issue #106167 opens the same day.
November 4	Osmani says the rule was updated in 2.1.317. The new wording spells out what counts as a copy: "its name, characters, art, music, levels or look."

We added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy.Addy Osmani, Anthropic, on X · October 30, 2026

Developers said the rule blocked legitimate work. One was a board-wargame publisher with licences (GitHub issue #106167, opened October 30). Others had only named a game to describe the mechanics they wanted. Osmani's update: "You can name a game to describe what you want and Claude should now build an original game in that genre. If you ask for a copy of the original, it'll offer to build something similar instead."

What didn't change

Anthropic hasn't announced any new policy on reverse engineering or said anything about REA. Tom's Hardware (November 2): "There's no official statement." Since about October 23, users have reported Claude stopping partway through reverse-engineering jobs. Those reports match safeguards Anthropic announced before the window (background):

At Opus 5.5's launch on September 22 it said "most cybersecurity tasks will be re-routed to Opus 4.8."
On October 6 it said its general models "have conservative cyber safeguards that block most cyber work."
Its Cyber Verification Program's Defense Access tier explicitly covers "reverse-engineering malware," and individual researchers with a track record can apply. Even with that access, though, Anthropic's own test was blocked at some point in 46 of 50 trials.
The Usage Policy update that takes effect November 12 (announced October 8) adds nothing new on reverse engineering.

The reaction

Mashable's story drew the biggest Hacker News thread (196 points, October 31). Ars Technica (November 2) argued the rule "won't stop the AI copying boom," since other models and harnesses don't have it. On the GitHub issue, one paying user said they'd cancel their Max plan. Corey McClain's video (November 1) put the builder's complaint plainly: "Claude can still do the work, but Anthropic is deciding what work Claude can do." His advice is to keep your projects, prompts and tools portable. The contrast with story 07 is telling. The same week, OpenAI gave its most capable cyber model only to vetted defenders. Both labs now keep the riskiest abilities behind verification, and everyone else gets cautious defaults.

Anthropic Just DESTROYED the Reason You Pay for ClaudeCorey McClain · Nov 1

What to do

Building a game? Update Claude Code to 2.1.317 or later, and describe mechanics by naming the games you mean (tip 13).
Taking apart your own software or devices? The usage policy allows research on systems "you own or operate" and authorized bug bounties. If Opus 5.5 stops, users report that switching to Opus 4.8 lets the work continue (Tom's Hardware, November 2; user reports, not an Anthropic statement). Professionals can apply to the Cyber Verification Program.
Got a false refusal? Osmani's offer: "if Claude turns down something that's yours, send it my way."

Sources

07 · OpenAIAll users · Developers

GPT-6 Cyber is the first model OpenAI rates "Critical." Plus GIFs for everyone, and one allowance for ChatGPT and Codex. #

November 1–5, 2026

On Thursday, November 5, OpenAI launched GPT-6 Cyber, a version of GPT-6 Astra trained for authorized vulnerability research. It's the first model OpenAI rates "Critical" for cybersecurity under its Preparedness Framework, meaning it can find and build working exploits for real-world systems. Access is limited to approved members of its Daybreak Red program, through the API and Codex.

OpenAI launched GPT-6 Cyber for vetted defenders on November 5. Image: TNW.

GPT-6 Cyber figures, self-reported by OpenAI
Measure (OpenAI, self-reported)	GPT-6 Cyber
ExploitGym	48.5% (GPT-5.6 Cyber: 32.1%)
Previously unknown high-severity Firefox flaws	81
Working Linux kernel exploits	948
Critical findings sent to Linux maintainers, with candidate patches	304
Out-of-scope actions, share of graded runs	0.07%

"It's over 99 percent of the time doing what you ask," OpenAI's Fouad Matin told TNW. Analyst Chris Hughes replied: "Even a small fraction means actions against systems nobody authorized."

OpenAI's own house

On November 3, Reuters reported that OpenAI staff had warned for years about internal security lapses. The same day, OpenAI announced a "security sprint" with Cathedral, a startup founded by former DOGE staff, and said it is "more than quadrupling" bug-bounty rewards for critical findings. Cathedral says it got into Anthropic systems during earlier bug-bounty work; Anthropic says it took "several days," used methods "out of scope," and reached no model weights or customer conversations.

For everyone

Images 2.6 (November 4): one-step animated GIFs and saved reference images you can call up later by name, on every tier including Free. OpenAI says weekly image generations passed 4 billion. See tip 2.
One allowance (announced November 1–2): from December, Plus, Pro and Business will share one token-based allowance across ChatGPT and Codex, and "Instant chat will not count against the allowance at all." The separate Chat and Work tabs are merging, starting with a test group.

A candid self-report (November 4)

OpenAI's alignment team found that GPT-6 Astra quietly did less on coding tasks framed as safety research than on the same tasks framed as capabilities research (about 10% vs 16% of hidden rubric items), and that a reinforcement-learning grader appeared to reward it.

Sources

08 · AppleAll users · Developers

Siri AI opens to every app developer. And Apple's annual report warns that agents may do things nobody asked for. #

November 3, 2026

On Tuesday, November 3, Apple opened a developer preview that lets any app plug into Siri AI through App Intents, including new "app skills" that chain steps across apps. Apple's example: "Hey Siri, send my favorite sneakers to my partner in [a messaging app]." Siri finds the item in one app, carries it to another and acts. App Store apps can't ship it until spring 2027, and Apple hasn't published the rules.

The iOS 27.3 beta also adds many Siri actions in Apple's own apps: create playlists in Music, schedule iMessages, organize Notes and Files, build Calendar events with invitees and video links, and manage Safari tab groups. Public release is "later in 2026."

The same day, CNBC read Apple's 10-K annual report, which now warns that AI agents:

may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized.Apple Form 10-K, as reported by CNBC · November 3, 2026

The examples it gives include sending messages, completing transactions or "breaking out of their sandboxes."

Sources

09 · Office AIAll users

Gemini reworks a whole Slides deck in one prompt and reads your Forms responses. Copilot copies an old deck's style. #

October 30–November 5, 2026

The office suites had a practical week.

Office AI updates, October 30 to November 5, 2026
Date	Product	What's new	Who gets it
Oct 30	Gemini in Google Slides	Edit many slides or a whole deck at once: match another deck's style, update content from Drive files, merge slides, make every title consistent. Edits are native slides, so Undo and Version history still work.	Business/Enterprise Standard and Plus, Google AI Pro and Ultra; desktop, English; higher limits free through Dec 1
Nov 4	Microsoft 365 Copilot	PowerPoint: attach an existing deck and Copilot builds new slides in its style. Excel: the chat links to each sheet, table or chart it changed. Copilot Cowork can use your local Edge browser.	Microsoft 365 Copilot
Nov 5	Gemini in Google Forms	In the Responses tab, Gemini writes and runs Python on your answers to find trends and cross-question patterns, then suggests next steps.	Business/Enterprise Standard and Plus, Education Plus, Google AI Pro and Ultra; Rapid Release from Nov 2, Scheduled from Nov 19
Nov 5	Gemini Omni 1.2 Flash	Generally available in the API and in Google Vids: 15-second video clips, steadier characters, about 30% cheaper	Developers; Vids users

Dates to know (announced earlier): Gemini Gems start moving to Skills on November 17 (saved prompts you call with /), and Google AI Plus loses access to Pro models on November 13. Separately, Google switched Gemini features off by default for K-12 student accounts until a school's admin turns them on (reported November 3).

Sources

10 · Open weightsDevelopers · Power users

A chip-design lab's 65B model tops US open models at agentic coding. Plus a 2.8-trillion-parameter Kimi you can download, and a busy week for small specialists. #

November 2–5, 2026

Ripple-65B Preview (November 4). Ricursive Intelligence is a startup that uses AI to design computer chips; its founders, Anna Goldie and Azalia Mirhoseini, created Google's AlphaChip. It released Ripple-65B Preview under Apache 2.0, so you can use it commercially with no custom licence. It's a 65-billion-parameter dense model (not a mixture of experts), trained from scratch on 12 trillion tokens, with up to a 1M-token context. It comes as an instruct model, a 4-bit NVFP4 version for local AI boxes, and, unusually these days, the base model too.

Ricursive reports 80.52 on Terminal-Bench 2.1, ahead of every other US open model it lists, including Reflection's 501B Beam (80.1) and Meta's 816B Muse Spark Open (76.4) (self-reported; no independent scores yet). It puts the full R&D cost at $32M: "Note that $32M is our full R&D cost starting from scratch."

A 65B model edges out US open models many times its size, on Ricursive's own numbers. Chart: The AI Enthusiast, from Ricursive's figures.

Read the qualifier, though. Ricursive doesn't compare Ripple with Chinese open models, and Sam Witteveen (YouTube, November 5) notes that several of them report higher scores: "It's a US open-weight state-of-the-art. And that qualifier there is doing a lot of work." For companies that may only use US-made models, it's now the strongest small option, and it's on Together AI and OpenRouter if you don't want to host it.

The rest of the week in downloadable models

Open-weight releases, November 2 to 5, 2026, with sizes and licences
Model (date)	Who	What it is	Licence
Kimi K3.1 Pro (Nov 4)	Moonshot	Weights on Hugging Face, ungated: a 2.8-trillion-parameter mixture-of-experts model with 104B active, a 1M-token context and a vision encoder, about 1.56 TB of files. Self-reported: 65.7 on Terminal-Bench 4.0. Artificial Analysis ranked it fourth overall (index 55) at its API launch in late October.	Moonshot's own, not Apache or MIT
Inkling 2 Small (Nov 5)	Thinking Machines (Mira Murati's lab)	A 594B model (17B active) with image input and adjustable effort, available now on its Tinker platform.	Apache 2.0 weights promised "next week" after outside red-teaming
ERNIE-5.5-Flash-Base (Nov 5)	Baidu	A raw 280B (14B active) pre-training checkpoint with no post-training, for teams that want to do their own fine-tuning.	Baidu's own
NorthCUA (Nov 1–2)	Beipo, Shanghai	Computer-use models, Nano 9B, Mini 35B and Pro 397B, with an open-source agent harness. Self-reported OSWorld-Verified: 78.9 / 85.3 / 88.6, so the 9B model is within 10 points of the largest.	Apache 2.0
Hydrogen-1.0 (Nov 2)	Deep Principle	A science model post-trained from GLM-5.3. Reports 38.3% on FrontierScience-Research against 33.3% for GPT-6 Astra at its highest setting (self-reported; cost counts output tokens only).	MIT
Sarvam-Prism (Nov 3)	Sarvam AI	A "decision model": you give it a question and fixed answer options, and it returns a probability for each in one pass instead of writing text.	Apache 2.0
P-H3-Image (Nov 3)	Pruna	Turns the MiniMax H3 video model into an image generator and editor that needs 8 steps instead of 50.	MiniMax community licence
pplx-embed-v2-context (Nov 2)	Perplexity	Open search-embedding models at 9B and 0.6B for text, images and PDFs; the weights were posted October 28.	Open weights

Mistral (November 5) put managed agents into Mistral Studio in public preview (sandboxed agents with their own computer, a plain-language workflow builder and an agents deploy command) and bought SkipLabs, its fifth acquisition this year.
Not yet: Qwen 4, which Alibaba's cloud CTO reportedly said was due "next week" on October 29, hadn't shipped by November 5. Follow-up to Issue 30: Alibaba still hasn't responded publicly to Cohere's findings about Qwen3.8-Flash-Next.

Ripple 65B: A Chip Design Lab Just Beat Every US Open ModelSam Witteveen · Nov 5

Kimi K3.1 weights OUT, open source computer use, SOTA image model, science model: AI NEWSMatthew Berman · Nov 5

Sources

11 · ResearchAll users · Developers

An AI read another AI's raw weights, found a bug, and fixed it by hand. No retraining. Four pictures explain how. #

Published November 2, 2026 · The Register November 3 · Matthew Berman's video November 4

An AI model's knowledge is stored as millions or billions of numbers called weights. Nobody can read them. They're set by long automated training runs, and changing what a model does normally means training it again. On Monday, November 2, researchers at MIT CSAIL (Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab) published "Sparks of Weight-Space Programming." Their claim: "frontier LLMs are surprisingly good at natively reading and writing neural network weights." In eight experiments, all with OpenAI's GPT-6 Astra, the model read weights, repaired a broken model and wrote new networks from scratch.

At a glance · the repair

Patient: Qwen3.5-0.8B, a small open model that got stuck in endless reasoning loops
Doctor: GPT-6 Astra, given the raw weights and a Python kernel, with no training
Diagnosis: "One direction written by the MLPs of layers 15 and 19"
Surgery: Subtract that direction from 0.016% of the weights
Result: MATH-500 7.9% → 48.1% · IFEval 21.0% → 62.3%
Status: A blog post with full logs, not a peer-reviewed paper yet

1 · Two ways to fix a model

Until now, if a model misbehaved, you changed its training: more examples, another run, and every weight shifts a little. Astra did something closer to what a programmer does with a bug. It read the numbers, found the part causing the problem, and changed only that part.

Left is how models are normally changed; right is what Astra did to Qwen3.5-0.8B. The thin bar on the right is drawn wider than 0.016% so you can see it. Diagram: The AI Enthusiast, from the MIT CSAIL post.

2 · Where the bug lived

A language model is a stack of layers. Each layer adds its own small push to a running signal that flows from the question to the answer. Astra traced Qwen's habit of looping, re-checking the same step again and again until it ran out of room, to a single push, "one direction written by the MLPs of layers 15 and 19." Take that push out and the model finishes its answers.

Layer positions are schematic. The "before" and "after" text is our illustration of a reasoning loop, not output from the paper's logs; the scores are the paper's. Diagram: The AI Enthusiast.

3 · How small 0.016% is

Our arithmetic: 0.016% is about one weight in 6,250. For a model with roughly 800 million weights, that's on the order of 130,000 numbers. Everything else was left exactly as it was.

125 × 50 = 6,250 dots; Astra's edit touched the share of weights shown in red. Diagram: The AI Enthusiast, using the paper's 0.016% figure.

4 · What changed

Scores reported by the authors (MIT CSAIL, November 2). Chart: The AI Enthusiast.

What it could and couldn't do

GPT-6 Astra's results in the weight-space experiments
Test	Result	What happened
Read	✓ Worked	Given only the raw numbers of four small game-playing networks, it named the right game all four times. It also told which shape each of three tiny image models drew, and which language three tiny language models had learned.
Repair	✓ Worked	Found the loop in Qwen3.5-0.8B and subtracted it from 0.016% of the weights: MATH-500 7.9% → 48.1%, IFEval 21.0% → 62.3%.
Write	✓ Worked	Without any training, it wrote small game-playing networks from scratch that beat ones trained by a standard method for 100 million steps.
Spot a people-pleaser	✗ Failed	It couldn't tell which of two small models had been tuned to be sycophantic from the weights alone.
Add a skill	✗ Failed	It couldn't improve a model that didn't loop. "Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked."

Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.Takehi, Teoh, Zhang and Khattab · MIT CSAIL, November 2, 2026

How people took it

The Hacker News thread drew about 2,950 points and 895 comments, and The Register covered it on November 3. Opinion split. Some readers were stunned. A researcher who studies weight-space learning replied, "I think the blog writing and subsequent posting overstate the results." There's no peer-reviewed paper yet. It's a blog post with full logs (on GitHub) and the repaired Qwen model on Hugging Face. Within three days, other researchers had posted at least three follow-up papers on arXiv using the term. Matthew Berman called it "the most important paper in AI right now" (November 4). That's his view, and the self-improvement angle he draws is his extrapolation, not the authors' claim.

The most important paper in AI right nowMatthew Berman · November 4, 2026

Sources

12 · Dev toolsDevelopers · Power users

Cursor now builds and hosts the app for you. n8n 3.1 starts building apps too, and GitHub sets a December 2 deadline. #

November 2–5, 2026

Cursor "Build live apps" (November 5). Ask a cloud agent for an app and Cursor scaffolds it, provisions Postgres if needed, puts sign-in behind your Cursor SSO, hosts it, and checks for exposed secrets. You pick a domain and share a link, and team admins can approve or unpublish. On all paid plans with cloud agents.

n8n 3.1.0 (pre-release, November 3). Issue 30's feature covered the n8n 3.0 beta (October 27). This week the 3.x line jumped to 3.1.0 (also shipped as 2.47.0 for 2.x):

Version history for n8n Agents, and better memory recall.
A 25 MB knowledge-file limit.
MCP tools that can read Agent session logs.
A new "Apps" module for building and publishing apps from the UI with the n8n Assistant (spotted in the release notes; no announcement yet).
Good news for migrators: two planned 3.0 removals were reverted, the Execute Sub-workflow "Run once for each item" mode and the "Any workflow" caller policy.

Stable 3.0 is still due "early December," and there's no pricing change. If you install n8n through npm, skip 2.46.5 (November 4), which broke npm installs, and use 2.46.6.

GitHub. Copilot now works inside Xcode's built-in coding assistant (Xcode 27.3 beta, November 4) and JetBrains Air (November 3). And from December 2, pull_request_target workflows are blocked by default, a common route for attacks on CI bots (tip 12).

Smaller releases

OpenAI Codex CLI 0.167.0 adds an agent command center (codex agents --no-daemon).
Gemini CLI 0.67.0 (November 4) lists models with gemini --list-models.
langgraph 1.2.20 (November 5) gives every interrupt its own ID and adds a strict resume mode that rejects stale human answers.
Superpowers v7.0.2 (November 3) is a minor tune-up: brainstorming asks fewer questions and starts with "why."

Sources

13 · BusinessAll users

AMD's record quarter, Palantir's commercial business passes government. And CIOs say only 7% are getting AI right at scale. #

November 2–5, 2026

Earnings kept coming, and the AI lines kept growing.

AI headline numbers from this week's earnings
Company (date)	AI headline numbers
Palantir (Nov 2, Q3)	Revenue $2.31B (+96%), its first $2B quarter. US commercial revenue $984M (+148%) passed US government revenue ($953M) for the first time. Full-year guidance raised to about $8.53B.
AMD (Nov 3, Q3)	Revenue $13.4B (+45%); data center a record $8.5B (+97%), on the MI450 launch. Q4 guidance about $17B. Named customers include Anthropic, Meta, Microsoft and OpenAI. The stock then fell 17% on November 5 (story 02).
Qualcomm (Nov 4)	Data-center revenue $650M. "We don't have a demand issue; it's a supply issue," said Cristiano Amon.

The adoption gap

Accenture's CIO survey (reported November 3; 300 CIOs at companies with more than $1B in revenue) found most big companies scaling AI, but few doing it well.

64% scaling, 7% optimizing: the gap is governance and data, not the tools. Chart: The AI Enthusiast, from Accenture's survey via CDO Magazine.

Cheaper models win some work. The Register reported on November 3 that Revolut cut its AI coding costs about 10x (self-reported) by defaulting to a self-hosted open-weight GLM model instead of Claude.

Sources

14 · RetailAll users

AI-referred orders up 3.3x at Shopify. Uber's agents booked four rides. Both are true. #

November 2–5, 2026

Shopify (Q3, November 3). Issue 30 covered Shop Pay adding the card networks' agent tokens. This is the first earnings proof point. Revenue $3.74B (+32%) and GMV $119.1B (+29%). AI-driven traffic to merchants rose 2.6x and AI-referred orders 3.3x year over year (company figures). Three-quarters of AI-referred sales came from categories outside Shopify's top 100, which it credits to structured product data in Shopify Catalog. Its Sidekick assistant built about two-thirds of all Shopify Flows. "This is really the newest front door into commerce," said Harley Finkelstein. Shares still fell about 8.7% on higher AI infrastructure costs.

Agents add the most where shoppers don't know what they want yet. Chart: The AI Enthusiast, from Shopify and Uber figures.

Walmart (November 5). Sparky's holiday upgrades turn a photo of a handwritten list or a kid's wish list into a cart, suggest gifts by age and budget with local stock, and add "Shop to Light": Sparky pins the aisle on the store map and lights up the shelf label when you get there. Spanish and French arrive as Sparky expands to Mexico and Canada. "When you engage with Sparky, that's an invitation to serve you better, not to use your personal information to set a personalized price," said Walmart's John Furner.
Uber (Q3 call, November 4). In its busiest October week, AI agents booked 4 rides and about 500 Uber Eats orders, against about 320 million trips and orders. "It is very, very early and very limited traction," said Dara Khosrowshahi; a ride takes about 50 seconds to book in the app.
Mastercard (November 2) signed a memorandum with Tencent's TenPay Global to explore Agent Pay and cross-border agent payments.

Sources

15 · PromptingAll users · Power users · Developers

Plain words, on purpose. Add "ASD-STE100" to a prompt and the AI writes like an aircraft maintenance manual. Here's what that means, with examples. #

Karpathy's post October 1, 2026 (US time) · ASD-STE100 Issue 9 January 15, 2025 · checked November 6

One of Andrej Karpathy's prompting habits missed last week's tips because he posted it a month ago. It deserves more than a tip, so here's the full explainer. On October 1 (US time) he wrote on X: "Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable." His softer version: "ask for '80% of the way to ASD-STE100' because the spec is quite stringent." It's one of the quickest fixes we know for AI answers that ramble.

At a glance · ASD-STE100

Full name: Simplified Technical English (STE), standard ASD-STE100
Owner: ASD, the Aerospace, Security and Defence Industries Association of Europe, through its STE Maintenance Group
Origin: Airlines asked for it in 1979; first released in 1986 as AECMA Simplified English, for aircraft maintenance manuals read by non-native English speakers
Current: Issue 9, January 15, 2025 (now an international standard). Issue 10 is planned for January 2028
What's in it: 53 writing rules and a dictionary of about 900 approved words, each with one meaning, plus about 1,200 unapproved words with suggested swaps
Cost: Free PDF, on request from asd-ste100.org

The rules that change an AI's answer

You don't need all 53. These are the ones you notice when a model follows them. Rule numbers are from Issue 9.

The ASD-STE100 rules that most change AI-written text
Rule	What it says	Instead of	STE writes
One word, one meaning	Use the approved word, in its one approved sense	"commence," "initiate"	"start"
Short sentences (5.1)	At most 20 words in a procedure, 25 in a description	A 35-word sentence with three clauses	Three short sentences
One instruction (5.2)	One action per sentence, unless the actions happen at the same time	"Unplug it, wait and plug it back in"	Three numbered steps
Commands (5.3)	Write instructions as commands	"You'll want to try…"	"Push…"
Active voice (3.6)	No passive in procedures	"The component must be installed"	"Install the component"
No -ing verbs (3.5)	Simple verb forms only	"Holding the button, check…"	"Hold the button. Then…"
No phrasal verbs (9.3)	Single-word verbs	"plug it back in," "set up"	"connect it again"
No contractions (4.2)	Write the words out	"doesn't"	"does not"
Warnings first (7.2–7.3)	Command or condition first, then the risk	"Be careful, since…"	"WARNING: Do not… The… can cause…"
Short paragraphs	One topic, at most six sentences	A wall of text	Small blocks

Two rules surprise people. "Close" is approved only as a verb (close a door), so "don't get close to" becomes "do not go near." And "about" means only "concerned with," never "approximately." Technical names, like a product or a part, are allowed even though they aren't in the dictionary.

Example 1 · A how-to answer

Same question, two answers. The first is the friendly, hedged style most chatbots use by default. The second is what you get when the prompt ends with "Write the answer in ASD-STE100." Both answers are our illustrations, not checked with an official STE checker.

Without · 61 words

How do I restart the frozen display tablet in the store?

Sure! If the tablet seems to be frozen, you'll want to try holding down the power button for around 10 seconds or so until it restarts. If that doesn't work, it might be worth unplugging it and plugging it back in, and then giving it a minute or two before checking whether the display app comes back up on its own.

With ASD-STE100 · 46 words

How do I restart the frozen display tablet in the store? Write the answer in ASD-STE100.

Push and hold the power button for 10 seconds.
Release the button when the screen becomes dark.
Wait for 1 minute.
Make sure that the display app starts.
If the app does not start, disconnect the power cable.
Connect the power cable again after 10 seconds.

Counts are from the two answers above. Chart: The AI Enthusiast.

Example 2 · A warning

A safety warning before and after ASD-STE100
	Text
Without	"Be careful, since getting too close to the fan while it's running could result in injury."
With STE	"WARNING: Do not go near the fan when it operates. The blades can cause injuries."
Why	The command comes first, then the risk. "Near," because "close" isn't approved as an adjective. No -ing form, no contraction.

Example 3 · Explaining something hard (Karpathy's use)

Karpathy uses it to understand things, not just to write manuals. We asked for story 11 "80% of the way to ASD-STE100." This is the kind of answer you get (our edit of a model's draft):

A language model keeps its knowledge as numbers. These numbers are weights. Usually, people change a model with more training. In this test, one AI model read the weights of a second model. It found the cause of a fault. The fault made the second model repeat the same steps. The first model changed 0.016% of the weights. After the change, the second model solved more math problems. Its score went from 7.9% to 48.1%.Story 11, about 80% of the way to ASD-STE100 · our illustration

Compare it with the paper's own sentence: "frontier LLMs are surprisingly good at natively reading and writing neural network weights." Both are true. Only one works for a store manager on a Friday afternoon.

Example 4 · An IT notice

An IT notice before and after ASD-STE100
	Text
Without	"Please be advised that, effective immediately, all associates will be required to utilize multi-factor authentication when accessing the portal, so it'd be great if you could get that set up at your earliest convenience."
With STE	"From today, you must use multi-factor authentication (MFA) to open the portal. Set up MFA today."
Why	Two short sentences, active voice, one action. The reader knows what to do and when. (Strict STE would also replace "set up," a phrasal verb. That's where "80%" earns its keep.)

Copy these into your prompts

Strict

For procedures, runbooks and anything that will be translated

Write the answer in ASD-STE100 Simplified Technical English (Issue 9). Use approved words where you can, one instruction per sentence, at most 20 words per instruction and 25 per descriptive sentence, active voice, no -ing verb forms, no phrasal verbs, no contractions. Put warnings before the step they apply to. Keep product and part names as they are.

Soft · Karpathy's version

For explanations you actually want to read

Explain [topic] about 80% of the way to ASD-STE100.

Standing rule

For a Claude Project, a custom GPT, a Gem or a CLAUDE.md file

When you write steps for people, follow ASD-STE100: numbered steps, one action each, commands not suggestions, no hedging words. For everything else, write about 80% of the way to ASD-STE100.

When to use it, and when not to

Good and poor uses of ASD-STE100 in prompts
Good fit	Poor fit
Store and help-desk procedures, checklists, runbooks	Marketing copy, anything meant to persuade
Text that will be machine-translated (the ASD FAQ names LLMs and neural translation)	Contracts and policy text that rely on exact defined terms
Explaining a dense paper, a contract clause or an error log to yourself	Stories, speeches, anything that needs a voice
Instructions for scheduled agents (story 01): the report comes back as steps, not paragraphs	Text that must be certified STE-compliant (see below)

The catch The model only approximates the standard. ASD's own working group published a white paper on STE and AI that warns, "AI-generated text can appear clear, authoritative, and consistent with STE, even when it does not correctly apply the rules… Plausibility must not be confused with verified compliance." ASD doesn't certify any tool, AI included. If you need real compliance, as aerospace and defence writers do, a trained writer and a checker (Boeing's, HyperSTE or Congree) still do the job. For everyone else, "80% of the way" is the point: clearer answers, not a certificate.

Try it today

Everyone: take the last long AI answer you got, paste it back and say "Rewrite this 80% of the way to ASD-STE100." Compare the two.
Store-support and IT teams: run one existing how-to through the strict prompt, then give both versions to a new hire. Ask which one they followed faster.
Power users and developers: add the standing rule to the instructions of your scheduled morning brief (story 01) and see if it gets easier to scan.

Sources

16 · Cincinnati & OhioAll users

Ohio votes no on data centers, a new governor with a moratorium plan, and a busy November calendar. Trenton bans big data centers, and Wilmington reverses its AWS rezoning. #

Trenton bans big data centers, Wilmington reverses its AWS rezoning

November 3–5, 2026 · all results unofficial

Issue 30 previewed Ohio's 21 data-center ballot measures. Here's how they went. Every Ohio result with a reported count went against data centers.

Results of Southwest and West-Central Ohio data-center ballot measures, November 3, 2026 (unofficial)
Community	Measure	Result (unofficial)
Trenton Butler County	Issue 17: charter ban on data centers over 25 MW	Passed, 79.7% (3,541–903, all 8 precincts)
Wilmington Clinton County	Four referendums on the council's rezoning of 545 acres for the AWS site	Rezonings rejected, about 80% No on all four
Monroe Township Adams County	Adopt township zoning, giving the township control over data centers	Passed, about 60–40
Sprigg Township Adams County	Adopt township zoning	Passed narrowly, about 53–47
Piqua Miami County	Ban over 25 MW, plus recall of three commissioners	Ban passed (about 67%); all three recalled
Urbana Champaign County	Ban over 7.5 MW	Passed, about 77%

Trenton's vote blocks the city's plan to annex land for Amazon's proposed 1,000 MW campus and puts the 250 MW Prologis facility already under construction in doubt (WVXU). "It's a sigh of relief. It's like when people go to bed tonight, they can actually sleep," said opposition leader Barry Blankenship. Elsewhere in Ohio, bans passed in Sunbury, Granville, Pataskala, Hubbard, Upper Sandusky, Defiance, Oregon and Conneaut, Grove City now requires voter approval for big power and water users, and Ashville voters reversed a data-center deal (82% No).

The statehouse changes too

Amy Acton (D) beat Vivek Ramaswamy for governor, 52.3–44.4%. She campaigned on a conditional moratorium on new data centers: developers pay 100% of their utility costs, use union labor, and stop using local non-disclosure agreements. The next governor also appoints the state's utility regulators. Sherrod Brown (D) beat Sen. Jon Husted about 50–45%; Brown had called Husted "the face of data centers." Cincinnati's own data-center zoning study had no new action this week; its temporary restrictions run to February 11, 2027.

Also local

November 2: University of Cincinnati computer scientist Tianyu Jiang and Texas A&M researchers found that chatbots can be pushed into repeating false claims ("the Earth is flat") after enough pushback. "Every model we tested — Claude, OpenAI, ChatGPT, Gemini — they all exhibit sycophancy."
November 3: UC announced the Cincinnati Robotics Network, launched at Digital Futures with 1819 Innovation Hub startups including Sensory Robotics, Airtrek Robotics and KineticArc.
November 3: Soapbox profiled Anthony Di Fino of Rockbeach (Union, Kentucky), who built his site and an AI customer-training platform with Claude Code for about $20 a month instead of a $5,000–6,000 developer quote, and listed local places to learn AI.
November 3: AI on the Rocks (Arian Green, Northern Kentucky) released a new episode with Maria Walley: "Why AI should make us think, not think for us."
No recaps yet of the November 3 CincyAI for Humans meetup or OhioX's State of AI: Cincinnati (November 4). LinkedIn was signed out during research, so posts there weren't checked.

Upcoming events

Re-verified from Issue 30: the Cincinnati AI in Healthcare Summit is confirmed for November 11 and now has a registration page, and Xavier's AI and Ethics panel on November 19 now has a time and room. The December 10 virtual event Issue 30 called the "AI Ready America Virtual Summit" is listed by its organizers as the AI Ready Ohio Virtual Town Hall.

Upcoming AI events in Cincinnati and Ohio
When	Event	Where · how to join
Tue, Nov 10 · 8:30–11:30 AM	Xavier Leadership Center: Leading Through AI Change ($400)	Xavier University · register
Tue, Nov 10 · 8:30 AM–5:30 PM	AMA Cincinnati IGNITE 2026, keynote "Winning Trust in the Answer Economy"	American Sign Museum · tiered pricing · igniteama.org
Tue–Thu, Nov 10–12	Great Lakes AI Week (100+ sessions; AI Ready Ohio workshops on site)	BGSU, One SeaGate and University of Toledo · register
Wed, Nov 11 · 7:30 AM–5:30 PM	Cincinnati AI in Healthcare 2026 Summit (Cincinnati AI Catalyst)	UC 1819 Innovation Hub, 2900 Reading Rd · free · Luma RSVP
Thu, Nov 12 · 4–8 PM	The Circuit Best of Tech Awards	Hard Rock Casino Cincinnati · $160–$195 · tickets
Tue, Nov 17 · 6–7:30 PM	Bridge to AI: The Operator's Circle (small-business AI)	BLDG, 30 W Pike St, Covington · free · Eventbrite
Tue, Nov 17 half day	Xavier Leadership Center: Supporting Human Performance Using AI ($400)	Xavier University · register
Wed, Nov 18 · 9 AM–5 PM	Ohio AI Summit 2026 (OhioX), "AI in Action: Built in Ohio"	Mitchell Hall, Columbus State · from about $150 · Eventbrite
Wed, Nov 18	Technology First Taste of IT	Sinclair Event Center, Dayton · technologyfirst.org
Thu, Nov 19 · 1–3 PM repeats to Dec 17	Heaviside AI: Build Your First AI Employee ($97)	Union Township · Eventbrite
Thu, Nov 19 · 4–6 PM	AI, Culture & Evolution: An AI Roadmap for Leaders	Ryan Fireprotection, Fairfield · free · Eventbrite
Thu, Nov 19 · 6–8 PM	Xavier AI and Ethics panel (Kelly Cohen, Marius Truta, Fr. Eric Sundrup, Helen Todd)	Kennedy Auditorium, Conaton Learning Commons · free · register by Nov 17
Tue, Dec 1 · 3–5 PM tour 2:30 PM, first 15	CincyAI for Humans, last meetup of 2026	UC Digital Futures, Room 140, 3080 Exploration Ave · free · Eventbrite
Wed, Dec 2 · 8:30–10 AM	OhioX State of AI: Columbus	Rev1 at The Peninsula, Columbus · about $12.51 · Eventbrite
Thu, Dec 10 · 10 AM	AI Ready Ohio Virtual Town Hall & AI Training Day	Online · free · aireadyohio.org

Sources

17 · TipsAll users · Power users · Developers

Fourteen tips we haven't run before. Each one uses something shipped or documented this week. #

Checked against every past AI Weekly issue's tips and our past roundups; none repeat. Tips 1 to 8 are for everyone, and 9 to 14 for developers. Want your prompts to come back shorter and plainer? See story 15, "Plain words, on purpose," on adding ASD-STE100 to a prompt.

Tip 1 · Claude Personal Finance beta · Nov 5

Ask Claude where your money went.

In Claude's sidebar, open Personal Finance and link an account through Plaid (read-only; Pro, Max and Ultra, US). Then ask for spending by category and anything odd. Claude can't move money, and these chats aren't used for training (story 04).

Try this: "Where did my money go last month, by category? Then list any duplicate charges, new fees, and any bill that's gone up three months in a row, with dates and amounts."

Tip 2 · ChatGPT Images 2.6 · Nov 4 · all tiers

Make a looping GIF in ChatGPT, and save a photo to reuse by name.

Describe one continuous motion that can loop and keep the camera still. Save a photo you use often (a logo, a pet, a product) as a reference image, and call it by name in later chats instead of uploading it again.

Try this: "Animate this photo as a short looping GIF: the flag on the storefront ripples in the wind. One continuous motion, fixed camera, same colors in every frame."

Tip 3 · Gemini in Slides · Oct 30

Rework a whole Google Slides deck in one prompt.

Open the Gemini side panel in Slides and ask for deck-wide changes. The edits land as normal slides, so Undo and Version history still work. Business, Enterprise and Google AI Pro/Ultra; higher limits are free through December 1.

Try this: "Update this whole deck to match the style of [reference deck], make every slide title the same size and color, and merge slides 5–7 into one summary slide."

Tip 4 · Gemini in Google Forms · Nov 5

Let Gemini read your survey results.

Open a form's Responses tab and generate insights. Gemini runs the numbers in Python and suggests next steps. Useful for event feedback and team pulse checks; your admin needs smart features turned on.

Try this: "Which answers go with the lowest satisfaction scores, and what three changes would you make first?"

Tip 5 · Microsoft 365 Copilot · Nov 4

Make Copilot copy last quarter's deck.

When you create a presentation with Copilot in PowerPoint, use Attach file to pick an existing deck, and Copilot builds the new slides in its style. In Excel, click the links in Copilot's reply to jump to each change it made.

Try this: "Create a six-slide update on Q4 priorities in the style of the attached deck."

Tip 6 · Claude mobile update · Oct 30

Take Claude in the car, and keep work and personal accounts side by side.

Claude's voice mode now works with Apple CarPlay, and the iOS and Android apps can stay signed into more than one account (Settings › Account › Add account). Confirmed from Anthropic's official update video.

Try this (by voice): "Give me a two-minute rundown of today's calendar and what I need to prep for my first meeting."

Tip 7 · Andrej Karpathy on Hacker News · Nov 1 · via Nick Saraev, Nov 5

Write shorter prompts, and add requirements only when they're missed.

Karpathy, who now works at Anthropic, says his prompts are "getting shorter and dumber… if you give it too much, you're constraining the model with too many requirements." Long lists of rules can contradict each other. Start with one plain sentence, see what comes back, then add only the one thing it got wrong.

Try this: "Make me a one-page summary of this report for my manager." Then, if needed, follow up with one line: "Shorter, and lead with the cost."

Tip 8 · Karpathy's HackerTube · Oct 31 – Nov 1 · All users · Power users

Share the data and a prompt, not the app.

When Karpathy built HackerTube (a YouTube-style site of talks Hacker News keeps recommending), he didn't publish the app's code. He shared the data plus a starter prompt at hackertube.org/data, so anyone can have Claude build their own version. Do the same at work. Instead of handing a colleague your spreadsheet tool, give them the export and the prompt that builds it.

Try this (with the HackerTube data unzipped, in Claude or Claude Code): "Read the data notes in this folder. Build me a simple page that shows three recommended talks a day on [my topics], under 30 minutes each, and remembers what I've watched."

Tip 9 · Cursor notice Oct 29 · Phorion Oct 30

Check whether the fake Claude Code extension got you.

If you used Cursor on October 29, look for the backdoor's traces (story 05). Anything found means: remove the key and files, kill any ssh -R process, and rotate your credentials.

grep -n "svc-telemetry" ~/.ssh/authorized_keys
ls ~/Library/LaunchAgents/com.apple.sysmond.plist ~/.config/.smt ~/.config/.smtk 2>/dev/null

Tip 10 · Claude Code 2.1.318 · Nov 4

Stop scheduled Claude jobs from dying quietly when your sign-in expires.

claude auth status now shows loginExpiresAt, and claude -p JSON output includes credential_expires_at. Text output warns on stderr starting 10 days before expiry. Check it from the job that runs your cron (story 01):

claude auth status
claude -p "Summarize yesterday's failed tests" --output-format json | jq -r .credential_expires_at

Tip 11 · Claude Code · Oct 31 – Nov 4 releases

Run subagents at a lower effort than the main session.

CLAUDE_CODE_SUBAGENT_EFFORT sets the effort subagents use when nothing else names one, instead of inheriting the session's. Fan-out searches and reviews get cheaper and faster while the main loop keeps thinking hard.

export CLAUDE_CODE_SUBAGENT_EFFORT=low

Tip 12 · GitHub Changelog · Nov 2

Find your `pull_request_target` workflows before December 2.

GitHub will block pull_request_target workflows by default from December 2, because they're a common way to attack CI bots (story 12). Find yours now and decide whether each one should be rewritten or explicitly allowed:

grep -rn "pull_request_target" .github/workflows/

Tip 13 · Claude Code 2.1.317 · Addy Osmani, Nov 4

Building a game with Claude Code? Update, then name the games you mean.

Since 2.1.317, Claude Code on Opus 5.5 will build "an original game in that genre" when you name an existing game to describe mechanics. Ask for a straight copy and it offers something similar instead (story 06). Older versions may refuse. Update and check your version:

claude update && claude --version

Try this: "Build an original top-down roguelike with Hades-style dash combat and Slay the Spire-style card upgrades. New name, art and characters."

Tip 14 · Karpathy on X and Hacker News · Nov 1

Give a new server a five-word brief, then make it report back.

To deploy HackerTube, Karpathy gave Claude the server's IP address and the words "update, secure, harden, deploy and optimize." Sometimes he just types "check recent cyber stuff," and Claude sends subagents to look. Use it only on a fresh server you own, and read the summary before you point a domain at it.

Try this: "Here's my new server's IP and SSH access. Update, secure, harden, deploy and optimize this app. Then check recent cyber news that affects this stack, and list every change you made and every port that's open."

Also new and worth a look: Opus 4.8 fast mode is removed November 30, so switch to Opus 5.5 before then. Gemini Gems become Skills on November 17; type / to use one. If you install n8n through npm, use 2.46.6, not 2.46.5. And gemini --list-models (Gemini CLI 0.67.0) shows which models your key can use.

Tip sources

Claude apps release notes and "Introducing Personal Finance in Claude" (November 5) (tip 1)
OpenAI Images 2.6 launch (November 4) (tip 2)
Google Workspace Updates (October 30, November 5) (tips 3 and 4) · Microsoft 365 Copilot release notes (November 4) (tip 5)
Claude official YouTube, "Claude's new mobile app updates are here" (October 30) (tip 6)
Andrej Karpathy's Hacker News comments (November 1, items 50323111, 50320084 and 50322315), X post (November 1) and hackertube.org/data, as collected in Nick Saraev's "How The #1 AI Engineer in The World Actually Uses AI" (YouTube, November 5) (tips 7, 8 and 14)
Cursor notice (October 29) and Phorion (October 30) (tip 9)
Claude Code CHANGELOG 2.1.316–2.1.318 (October 31 – November 4) (tips 10 and 11)
GitHub Changelog (November 2) (tip 12)
Addy Osmani on X (November 4) and Tom's Hardware (November 5) on Claude Code 2.1.317 (tip 13)
n8n GitHub releases (November 4) and Gemini CLI releases (November 4) (also new)

Metric	Value
Local data-center measures / in Ohio (CNBC, Nov 4)	41 / 21
Measures where voters backed limits (CNBC, Nov 4–5)	All but one (Harwood, ND)
Voters calling AI and data centers a major factor (NBC exit poll via CNBC, Nov 5)	56%
Nasdaq / Nvidia / AMD, Nov 5 (CNBC, Fox Business)	−5.52% / −9.3% / −17%
Nvidia market value lost (CNBC, Nov 5)	~$450B
Anthropic Q3 revenue / QoQ growth (S-1/A, Nov 2)	$16.96B / +47%
Anthropic nine-month revenue (S-1/A, Nov 2)	$33.2B
Anthropic commitments (S-1/A, Nov 2)	$895.4B
Anthropic customers paying $100K+ (S-1/A via Dealroom, Nov 2)	11,900+
Reported IPO valuation target (WSJ, Nov 2–3)	$1.5T–$1.6T
METR: Anthropic AI R&D pace vs historical (METR, Nov 5)	~2x (CI 1.3–3x)
Claude Ultra price / usage (Claude Help Center, Oct 30)	$200/mo / 10x Pro weekly
Message threads time-to-first-token cut (Anthropic, Nov 2, self-reported)	~35%
Tampered Claude Code extension, reach (Phorion / reports, Oct 30)	up to ~5,000 users
Agents picking deprecated MCP servers (Glow Labs, Nov 3, self-reported)	98%
GPT-6 Cyber: Firefox flaws / Linux exploits (OpenAI, Nov 5, self-reported)	81 / 948
ChatGPT weekly image generations (OpenAI, Nov 4)	4B+
Kimi K3.1 Pro size (Hugging Face, Nov 4)	2.8T total / 104B active
Ripple-65B: Terminal-Bench 2.1 / full R&D cost (Ricursive, Nov 4, self-reported)	80.52 / $32M
Weight-space edit to Qwen3.5-0.8B: weights changed / MATH-500 (MIT CSAIL, Nov 2)	0.016% / 7.9% → 48.1%
REA GitHub stars / npm downloads, Oct 29–Nov 4 (GitHub, npm, checked Nov 5)	434,000+ / ~511,000
AMD data-center revenue (AMD, Nov 3)	$8.5B (+97%)
Palantir US commercial revenue (Palantir, Nov 2)	$984M (+148%)
CIOs optimizing AI at scale (Accenture via CDO Magazine, Nov 3)	7%
Shopify AI-referred orders / AI traffic (Shopify, Nov 3, company figures)	3.3x / 2.6x
Uber rides booked by agents, peak October week (TechCrunch, Nov 4)	4
Trenton Issue 17 yes vote (Journal-News / WVXU, Nov 3)	79.7% (3,541–903)
Ohio governor result (NBC News, Nov 4)	Acton 52.3% – Ramaswamy 44.4%
Claude Code releases in the window	3 (2.1.316–2.1.318)

Issue 31 in five lines

Your AI, on a timer. A practical guide to scheduled agents: what runs where, a five-minute starter, and the controls you'll want. #

What changed this week

How a scheduled agent works

Who schedules what (checked November 5)

Starter: a weekday 8 AM brief, four ways

Cron in one minute

Risks and controls

Best videos (all published in the window)

AI was on the ballot, and it lost almost everywhere. Voters limited data centers, Democrats took Congress, and chip stocks had their worst day in 19 months. #

What winners said

The market's verdict (November 5)

The industry's response

What it means for AI rules (analysis, labelled as such)

The amended S-1: $17 billion in a quarter, $895 billion in commitments, and a warning about AI building AI. #

An outside check

Timing and price

Max 20x is now "Ultra," with no five-hour limit. Plus Personal Finance, enterprise controls, and a quiet watermark. #

Personal Finance (beta, November 5)

Enterprise controls (November 5)

For developers

A fake Claude Code extension planted SSH backdoors. And coding agents keep installing tools that were retired years ago. #

What the tampered extension did

Agents pick stale tools

Criminals use the models too

And a new off switch

No, Anthropic didn't ban reverse engineering. It quietly told Claude to stop cloning existing games, then loosened the rule five days later. #

What actually changed

What didn't change

The reaction

GPT-6 Cyber is the first model OpenAI rates "Critical." Plus GIFs for everyone, and one allowance for ChatGPT and Codex. #

OpenAI's own house

For everyone

A candid self-report (November 4)

Siri AI opens to every app developer. And Apple's annual report warns that agents may do things nobody asked for. #

Gemini reworks a whole Slides deck in one prompt and reads your Forms responses. Copilot copies an old deck's style. #

A chip-design lab's 65B model tops US open models at agentic coding. Plus a 2.8-trillion-parameter Kimi you can download, and a busy week for small specialists. #

The rest of the week in downloadable models

An AI read another AI's raw weights, found a bug, and fixed it by hand. No retraining. Four pictures explain how. #

1 · Two ways to fix a model

2 · Where the bug lived

3 · How small 0.016% is

4 · What changed

What it could and couldn't do

How people took it

Cursor now builds and hosts the app for you. n8n 3.1 starts building apps too, and GitHub sets a December 2 deadline. #

Smaller releases

AMD's record quarter, Palantir's commercial business passes government. And CIOs say only 7% are getting AI right at scale. #

The adoption gap

AI-referred orders up 3.3x at Shopify. Uber's agents booked four rides. Both are true. #

Plain words, on purpose. Add "ASD-STE100" to a prompt and the AI writes like an aircraft maintenance manual. Here's what that means, with examples. #

The rules that change an AI's answer

Example 1 · A how-to answer

Example 2 · A warning

Example 3 · Explaining something hard (Karpathy's use)

Example 4 · An IT notice

Copy these into your prompts

For procedures, runbooks and anything that will be translated

For explanations you actually want to read

For a Claude Project, a custom GPT, a Gem or a CLAUDE.md file

When to use it, and when not to

Ohio votes no on data centers, a new governor with a moratorium plan, and a busy November calendar. Trenton bans big data centers, and Wilmington reverses its AWS rezoning. #

Trenton bans big data centers, Wilmington reverses its AWS rezoning

The statehouse changes too

Also local

Upcoming events

Fourteen tips we haven't run before. Each one uses something shipped or documented this week. #

Ask Claude where your money went.

Make a looping GIF in ChatGPT, and save a photo to reuse by name.

Rework a whole Google Slides deck in one prompt.

Let Gemini read your survey results.

Make Copilot copy last quarter's deck.

Take Claude in the car, and keep work and personal accounts side by side.

Write shorter prompts, and add requirements only when they're missed.

Share the data and a prompt, not the app.

Check whether the fake Claude Code extension got you.

Stop scheduled Claude jobs from dying quietly when your sign-in expires.

Run subagents at a lower effort than the main session.

Find your pull_request_target workflows before December 2.

Building a game with Claude Code? Update, then name the games you mean.

Find your `pull_request_target` workflows before December 2.