CLAUDE.md
Composed from the answers you gave and the bundles you picked. Your own paragraphs stay where you put them.
# acme-web ## Git <!-- skills:base:pushes=ask --> - Never push without explicit approval, even after a commit was approved.
The Claude Code
setup I
use.
CLAUDE.md · rules · skills · hooksCLAUDE.md · .claude/rules · .claude/skills · .claude/hooks
This is the Claude Code setup I use on my own projects: a CLAUDE.md, some rules, skills and hooks. It is split into bundles so you can take the parts that fit and leave the rest. The web builder and the CLI write the same files.
What init writesStart hereStart here3 min readPhilosophy4 min read
~/acme-web
$ pnpx @patrity/skills init --yes --profile nuxt-app.claude/rules/browser-testing.md.claude/skills/browser-testing/SKILL.md.claude/rules/docs.md.claude/skills/handover/SKILL.md.claude/skills/wiki-parity/SKILL.md.claude/skills/wiki-parity/_systems.example.json.claude/skills/wiki-parity/wiki-parity.test.ts.claude/rules/backend.md.claude/rules/cli.md.. 18 moreCLAUDE.md.claude/settings.json.claude/settings.local.json.gitignoreSetup written.
initnuxt-app
One run into an empty project, and what it wrote.
pnpx @patrity/skills init · --yes --profile nuxt-app
In your repo
What the run writes
A file, a directory and a record. The paths are already relative to your project root, so there is no wrapper directory to dig anything out of.
.claude/
Rules scoped to paths, skills with their scripts, hooks, and the settings that wire them up.
.claude/ rules/web-nuxt.md skills/nuxt-docs/SKILL.md hooks/protect-env.sh settings.json skills.lock.json
Lockfile
Every file the tool wrote, hashed. That is how the next run can take a bundle back out without touching your edits.
{
"version": 1,
"projectName": "acme-web",
"answers": { "pm": "pnpm" },
"bundles": {
"nuxt": { "sha": "8f3c1d" } The choices
Why the setup looks like this
These are my choices, and each one is either a question you can answer differently or a bundle you can leave out.
Rules and skills stay apart
A rule is a glob and a constraint, and it loads on its own when Claude touches a matching file. A skill is the procedure, and it gets invoked when it is needed. Keeping them apart is what keeps my CLAUDE.md short enough that Claude still reads it. /skill/nuxt
Hooks fail closed
I moved the checks I keep forgetting into hooks, so the harness runs them rather than me. Delete a tracked hook script and the wiring exits 2, which is the behaviour I want from a check I rely on. /skill/quality-hooks
Docs in three tiers, one of them tested
Handovers, a wiki and frozen specs. The wiki is the tier that rots on me, so a registry names every system and a test fails on a system with no page, or a page nobody registered. /skill/docs-discipline
UI work is checked in a browser
Typecheck, lint and unit tests all pass on a component that never mounted, so I drive the change with playwright-cli: snapshot, act on refs, assert with eval, then read a screenshot. /skill/browser-testing
Memory and process are questions
These are the habits most likely to be wrong for your project, so the wizard asks instead of assuming. full process is brainstorm, spec, plan, TDD, then a two-stage review, and none is on the list too. Memory is off unless you turn it on. /skill/iterative-spec-design
The registry
The bundles
Listed in the order the CLI installs them. Every one is also a zip you can take on its own, with the CLAUDE.md lines that make it make sense.
AI Models
Lets Claude call the models you run yourself or reach through a gateway, for speech, vision, embeddings, reranking, decisions, images, video and music, from one CSV of endpoints.
aiinferenceaudioimagesembeddings
Browser Testing
Prove UI changes in a real browser with playwright-cli. Install-if-missing, the snapshot→ref→act workflow, the test-account convention and evidence discipline.
testinge2eplaywrightbrowser
Doc Fetcher
A generator for per-library documentation skills. Configure a template and Claude Code fetches a library's real docs from GitHub instead of recalling them.
docsgeneratorpython
Docs Discipline
The three-tier docs model (handovers, living wiki, frozen specs) with a handover skill and a mechanical wiki-parity test.
docshandoverswikiprocess
Iterative Spec Design
Design a system section by section with an approval checkpoint after each one, so a wrong assumption is caught in section two instead of in the finished spec.
processplanningspecs
Nuxt
Nuxt 4 framework docs fetcher and the rule that makes Claude Code check it before guessing a composable.
nuxtvuedocs
Nuxt UI
Nuxt UI v4 component docs and template fetchers plus the rule that makes Claude Code check them before touching a .vue file.
nuxt-uivueuidocs
Quality Hooks
Fail-closed Claude Code hooks: protect .env and credential files, lint after every edit, and flag an unwritten convention before context is compacted.
hooksqualitysafety
Read-only DB Access
A dedicated read-only role, a one-statement runner wrapped in BEGIN READ ONLY, and the rule that keeps destructive SQL out of Claude's hands.
databasepostgressafety
Remote Ops Quoting
The ssh argument-flattening trap that silently runs your command on the wrong machine, the two forms that work, and the base64 hop for anything with nested quotes or SQL.
opssshproxmoxshell
From the lab
Recent posts
October 15, 2026
My First Local AI Model Was an Autocomplete Server. Its Last Customer Was an AI Agent.
I evicted my self-hosted autocomplete model to make room for Cloudflare's Clef, a local Jev alternative, and found its only remaining user was an AI agent.
September 14, 2026I Benchmarked Nine Models to Spec a Four-GPU Node. Every Single Number Crowned the Wrong Winner.
I have to spec a two-to-four card inference node, so one Blackwell went on the bench first. Every throughput test picked one winner. Then a benchmark that actually compiles the code scored it 28.9% against 64.4%.
Take what you want
Take the whole thing, or just the bundle you came for.