The AI Enthusiast — Enterprise & Mobile Edition · Issue No. 09

01 — GREG MILLER

Claude Code Is Powerful — But Stateless

Every session starts from scratch. Decisions, error patterns, and domain insights vanish when you close the terminal.

1

Invisible costs

No per-session breakdown until the bill arrives

2

Context resets

Every decision, error pattern, and domain insight vanishes when you close the terminal

3

Agents go out of scope

Subagents touch files they shouldn't and use expensive models for cheap tasks

4

No guardrails

Force pushes, hard resets, and recursive deletes are one prompt away

5

Tribal knowledge

Each engineer's Claude knows different things about the codebase

02 — GREG MILLER

Context Rot Is Real

Output quality degrades well before the hard limit — and developers consistently overestimate AI performance.

60%

Output quality degrades at ~60% context utilization — well before the hard limit

Blake Crosley, 50 sessions study

65%

of developers report AI misses context during refactoring

Qodo 2025 Report

39pt gap

METR Paradox: Devs were 19% slower with AI, yet believed they were 20% faster

METR July 2025

+154%

PR size increase with 90% AI adoption. Also: +9% bugs, +91% review time

Google DORA 2025

3.8%

Only 3.8% of developers report both low hallucinations AND high confidence shipping AI code

Qodo 2025 Report

03 — GREG MILLER

Trellis

Governance Middleware for Claude Code — Hooks, Structured Knowledge, and Agent Ownership Rules

Greg Miller POS-Luxottica / POS-DevOps

H

Session Hooks

Fire at every key session event — start, tool calls, compaction, end

K

Persistent Knowledge

Shared team context: domain files, ADRs, error patterns, agent history

A

Agent Ownership

Frontmatter-declared paths and model tiers enforced at hook level

$

Cost Reporting

Per-session token and cost breakdowns — no surprise bills

/

Slash Commands

knowledge-sync, audit, standup, ADR, PR — all without leaving Claude

T

195 Automated Tests

62 hook + 9 cost + 12 contract + 112 integration tests

Azure DevOps dev.azure.com/POS-Luxottica/POS-DevOps/_git/trellis

04 — GREG MILLER

Hook Lifecycle

Session Start

Injects knowledge index, agent roster, process reminders

Session Start

Warns about knowledge files not updated in 90+ days

Before Every Tool Call

Blocks destructive commands

Enforces agent model tiers and file ownership boundaries

After Tool Calls

Prompts error logging; suggests push + PR after commits

After Subagents Finish

Prompts appending learnings to agent knowledge file

Before Context Compaction

Prompts graduating in-flight findings

Captures context before it's compressed away

Session End

Writes per-session cost report; warns if no knowledge was recorded

05 — GREG MILLER

Slash Commands

Knowledge & Maintenance

/knowledge-sync

Push local knowledge to org hub as PR

/knowledge-pull

Pull latest org knowledge from hub

/knowledge-merge

Resolve merge conflicts in knowledge files

/trellis-update

Update hooks and knowledge without leaving Claude

/trellis-audit

On-demand audit for documentation drift

/checkpoint

Capture session findings into knowledge files

Workflow & Reporting

/deep-dive

Structured architecture analysis

/standup

Generate standup from git + ADO activity

/adr

Record an architecture decision

/pr

Branch check, push, and PR creation

/token-report

On-demand cost report for current session

06 — GREG MILLER

Controlling the Spend

40-60%

Model routing + caching

60-80%

Full optimization (prompt + cache + cascade)

40-70%

Context reduction (20K → 8K tokens)

-60%

Specialized sub-agents (3-5 tools vs 20-30)

50%+

On-premise AI factory over 3 years (Deloitte)

Trellis-Specific Controls: Per-session cost reports · Model tier enforcement · Agent budget controls · Automatic cost warnings at session end

07 — IRV CASSIO

Community & AI Literacy

AI Literacy Day — March 27, 2026

ailiteracyday.org EDSAFE AI Alliance 150+ organizations

▸Cincinnati AI Catalyst event — Friday March 27, 6:30-8:00 PM EDT
The Hidden Chapter Bookstore, Fort Thomas, KY
Hosted by Kendra Ramirez and Helen Todd

▸YMCA of Greater Cincinnati — Friday March 27
“Made to Ask” (9-9:45 AM) · “Made to Think” (10-10:45 AM)

▸“The AI Doc” — Opens nationwide March 27
AMC West Chester 18 and other Cincinnati-area theaters

Cincinnati AI Catalyst

▸cincycatalyst.ai — Largest AI meetup in SW Ohio (150+ monthly)

▸Monthly at UC Digital Futures Building

▸Co-hosts: Kendra Ramirez & Helen Todd

▸Helen Todd — Creativity Squared podcast

▸“Friday Features” weekly newsletter by Carl Fraik

Cincy AI Week 2026

June 9-11 · Union Hall, Cincinnati

Leadership Roundtables, AI Summit, Women in AI & Tech, AI for Business

cincyaiweek.com

08 — IRV CASSIO

From the Field: Developer Tools

Four tools for individual developer control.

S

/statusline

Real-Time Session Awareness

▸Customizable status bar: model, context %, rate limits, cost

▸Color-coded thresholds — green/yellow/red at a glance

▸Shell script fed JSON from Claude Code

H

SSH MCP (Hive)

Secure Remote Agent Access

▸MCP server gives Claude SSH access to remote servers

▸Command allowlist with glob matching blocks danger

▸Full audit log — credentials never exposed to LLM

D

Claude Dispatch

Multiply Your Context

▸Turns your session into an orchestrator

▸Fans work to background workers with fresh context

▸Bidirectional communication for clarifications

P

/playground

Rapid Prototyping

▸One HTML file, zero dependencies — open in any browser

▸Interactive controls that update in real time

▸Config explorers, dashboards, presentations, prototypes

09 — IRV CASSIO

Interactive Playgrounds

Self-contained, single-file HTML applications that Claude builds on demand — explore, configure, and share without any setup.

P

What Is a Playground?

▸One HTML file, zero dependencies — open in any browser

▸Interactive controls: sliders, toggles, dropdowns that update in real time

▸Claude generates the entire thing from a single prompt

▸Works offline — share via email, Slack, or file drop

U

Use Cases

▸Config explorers — visualize settings before committing

▸Data dashboards — quick charts from JSON or CSV

▸Presentations — slide decks like this one, fully self-contained

▸Prototypes — UI mockups stakeholders can click through

UI Playground — interactive controls for theming and layout

UI Playground — live theming & layout controls

CLAUDE.md Explorer — structured config visualization

CLAUDE.md Explorer — structured config viewer

Try it: Ask Claude Code /playground — describe what you want to explore and get a working interactive app in seconds.

10 — IRV CASSIO

The Agent Ecosystem

Four pillars of Claude’s autonomous capabilities.

🖥

Claude Cowork GA Jan 2026

▸Desktop agent for knowledge workers

▸Built-in VM, browser automation, 11 open-source plugins

▸Full Windows parity

📱

Claude Dispatch NEW Mar 2026

▸Phone-to-desktop tasking

▸Scheduled recurring tasks, persistent memory

▸Zero setup — scan a QR code

🖱

Computer Use NEW Mar 2026

▸Click, scroll, type, navigate apps directly

▸Permission-first — asks before touching new applications

💻

Claude Code 350K+ DAU

▸1M+ merged PRs, Agent SDK, Skills, Channels

▸Code Review & Code Security shipping

11 — IRV CASSIO

Claude Dispatch

Launched March 17, 2026 · Research Preview — Phone-to-desktop tasking for Claude Code.

📱

Phone-to-desktop tasking

Assign work from your phone, Claude executes on your desktop

🔄

Async by design

Send a task and walk away, return to completed work

🧠

Memory across sessions

Remembers your preferences, project structures, workflows

📅

Scheduled tasks

Recurring automations like weekly reports built in

⚡

Zero setup

Scan a QR code, start dispatching in under 2 minutes

How It Works

1

Open Cowork

Launch Claude Desktop, enter Cowork mode

2

Scan QR

Click Dispatch, scan QR with your phone

3

Send Tasks

Assign work from anywhere via phone

4

Return

Come back to completed work on desktop

Note: Desktop must stay on — Dispatch is a remote control, not cloud compute. macOS only for now.

The Stack That Remembers

✓Context rot is measurable and starts at 60% utilization

✓Trellis: hooks + knowledge + agent ownership = consistent Claude behavior

✓Statusline, SSH MCP, and Dispatch give individual developers control

✓Playgrounds turn Claude into a rapid prototyping engine

✓Claude isn’t catching up — it’s building the infrastructure everyone else adopts

✓AI literacy is a community effort — get involved locally

irvcassio.com cincycatalyst.ai ailiteracyday.org creativitysquared.com

“Structure is what turns a powerful tool into a reliable teammate.”