Live inventory · 20 Aug 2026

The working
AI stack

Every model, skill, gate, and hook actually installed on this machine — read straight from ~/.claude, not from a wishlist. Two of the four layers are fully built. Two are empty.

24
Skills installed
4
Models on tap
1/5
MCP servers live
0
Hooks wired
0
Gates configured
01 — Request lifecycle

How one prompt moves through the stack

Six stages from keystroke to answer. The two amber stages are the ones you have not built yet, which is why nothing currently stops a bad tool call or fires on completion.

agentic loop — re-plan until the task is actually done Prompt CLI / VS Code Gate permissions not configured Model Opus / Sonnet Fable / Haiku Skills 24 loaded + memory Tools Bash / Edit MCP servers Hooks 9 events free none wired Result + deploy 01 02 03 04 05 06 07
Built & running Skills / context layer Available but unconfigured Agentic loop
02 — Model routing

Four models, one job each

Your default is set globally in ~/.claude/settings.json. Everything else is per-session or per-subagent.

Fable 5 default 1M context

Currently the machine-wide default. The [1m] suffix opts into the million-token window, which is why whole manuscripts and long audits fit in one pass.

claude-fable-5[1m]

Opus 5 planning

Architecture, plans, and any judgement call that is expensive to get wrong. Slowest and priciest, so it earns its place on decisions rather than execution.

claude-opus-5

Sonnet 5 execution

The workhorse. Once Opus has decided what to build, Sonnet builds it at a fraction of the cost.

claude-sonnet-5

Haiku 4.5 fast lane

Cheap, quick classification and lookups. Good for the mechanical passes inside a larger job.

claude-haiku-4-5-20251001

Groq · Gemini · Codestral not on this Mac

Your Windows box ran a multi-model MCP server exposing free Groq, Gemini, and Codestral endpoints for research-grade work. That server did not migrate.

Routing policy manual

The Opus-plans/Sonnet-executes split is a habit, not a rule. Nothing in config enforces it, so every session starts on the global default until you switch by hand.

03 — Skills

24 skills in three packs

Skills load on demand by description match, so an unused skill costs nothing until its trigger fires. This is the deepest layer of your stack by a wide margin.

SEO & Content

11 skills
seo-technicalseo-pageseo-local seo-schemaseo-sitemapseo-cluster seo-contentseo-geoseo-sxo improve-aeo-geohumanizer

Cloudflare Platform

11 skills
cloudflarewranglerworkers-best-practices durable-objectsagents-sdkcloudflare-one cloudflare-one-migrationscloudflare-email-service sandbox-stablesandbox-nextsandbox-migrate-to-next

Web Delivery

2 skills
web-perfturnstile-spin

The shape of this is lopsided

Twenty-two of twenty-four skills serve client web work. Your books, the KDP build pipeline, and your actual day job in sales have zero skill coverage. The Windows machine had tampa-market-scan and nine Tampa delivery skills that did not come across.

Local scripts 2 files

Two helper scripts sit outside the skill system in ~/.claude/scripts/readability_stats.py and audit-jsonld.py. Neither is reachable from a skill, so they only run when invoked by hand.

04 — Tools & MCP

One server live, four waiting on OAuth

MCP servers extend the tool surface past the built-ins. Four of yours are connected accounts that need authorising before their tools appear.

chrome-devtools connected

Full browser control: navigate, snapshot, screenshot, run scripts, trace performance, audit with Lighthouse. This is what read 21st.dev to build this page.

stdio · npx chrome-devtools-mcp@latest

Gmail connected

Search, read, draft, send, label. The one connector already carrying your account.

claude.ai connector

Google Drive needs auth

Worth fixing first. If your manuscripts are backed up to Drive, this is the path that gets them onto this Mac without touching the Windows laptop.

Google Calendar needs auth

Unauthorised. Scheduling and availability tooling stays dark until it is connected.

Apollo.io needs auth

Your prospecting database. On Windows this fed the Tampa cold-call pipeline through a custom wrapper.

multi-model not migrated

A Python MCP server exposing groq_chat, gemini_chat, and codestral_chat. Still on the old machine.

05 — Gates & hooks

Nine events you can hook. You use none.

Hooks are shell commands the harness runs at fixed points in the lifecycle. They are how you get deterministic behaviour instead of asking the model to remember. Your settings.json holds two keys: theme and model.

SessionStart
Fires when a session opens. Load project state, print a dashboard, warm a cache.
unwired
UserPromptSubmit
Fires on every prompt before the model sees it. Inject context or block on a keyword.
unwired
PreToolUse
Fires before a tool runs, and can veto it. This is the real gate — the one that stops a destructive Bash command before it executes.
highest value
PostToolUse
Fires after a tool succeeds. Auto-format edited files, run the linter, rebuild.
unwired
Notification
Fires when the CLI wants your attention. Route to desktop notifications or Slack.
unwired
Stop
Fires when a response finishes. Commit, deploy, or announce completion.
unwired
SubagentStop
Fires when a spawned subagent finishes, separately from the main thread.
unwired
PreCompact
Fires before context is summarised. Persist anything that must survive the squeeze.
unwired
SessionEnd
Fires on session close. Flush logs, write a summary, tidy scratch files.
unwired

Permission gates are empty too

Both project entries in ~/.claude.json carry allowedTools: []. Nothing is pre-approved, so routine reads prompt you every time, and nothing is denied either.

The cheapest fix available

A PreToolUse hook plus a short allowlist removes most permission prompts and adds a real safety net, in one edit. The /fewer-permission-prompts command can generate the allowlist from your own transcript history.

06 — What to build next

Ranked by payoff

01

Rotate the leaked API keys

Your Windows backup carried Groq, Apollo, FAL, and HeyGen keys in plaintext inside memory files, and they are now sitting in ~/.claude/projects/ on this Mac. Rotate all four at the provider.

02

Wire PreToolUse and an allowlist

The single highest-leverage change. Turns an unguarded stack into a gated one and cuts prompt fatigue at the same time. Start with /fewer-permission-prompts.

03

Authorise Google Drive

The likely route to recovering six manuscripts without going back to the Windows laptop. Everything in the book catalogue is currently memory-only: Claude knows the titles, word counts, and build decisions, but not one manuscript file is on this machine.

04

Rebuild the KDP toolchain

Typst 0.15 and Pandoc 3.9 drove the book build system on Windows. Neither is installed here. brew install typst pandoc restores it.

05

Port the sales skills

tampa-market-scan and nine Tampa delivery skills exist on the old machine. They are the only skills that serve your actual job, and none of them made the move.