claude-code-setup

My complete Claude Code setup — skills, hooks, agents, commands, and an LLM-maintained wiki — installable with one command.

Three systems in one repo:

1. 🧠 LLM Wiki — a personal knowledge base your agent maintains for you (Karpathy's raw → wiki → schema model). The agent ingests sources into cited pages, logs every session, updates project statuses as you work, and lints itself weekly. You browse it in Obsidian.
2. 🔎 Deep Research pipeline — three portable skills that turn a research question into a source-backed, cited brief: advanced-web-search (Google-operator discovery in a real browser), crawl4ai (URL → LLM-ready markdown), deep-research (the orchestrator + evidence gates).
3. 🛠 Senior engineering setup — SENIOR-LOOPS.md (seven senior-engineer build loops), planner + reviewer subagents, /review /boot /evolve commands, always-on core invariants, a desktop-toast hook, and a context-usage status line.

Install (one command)

macOS / Linux / Git Bash:

git clone https://github.com/keysersoose/claude-code-setup && cd claude-code-setup && ./install.sh

Windows PowerShell:

git clone https://github.com/keysersoose/claude-code-setup; cd claude-code-setup; powershell -ExecutionPolicy Bypass -File install.ps1

Agent-native: clone the repo, open Claude Code inside it, and say "install this setup for me" — the agent reads this README and does everything, including the manual steps below.

The installer copies skills, hooks, agents, commands, and rules into ~/.claude/. It never touches your settings.json or CLAUDE.md — those two merges are the manual steps:

1. Merge settings.example.json into ~/.claude/settings.json (hook wiring, status line, and the LLM_WIKI_PATH / LLM_WIKI_OWNER env vars).
2. Paste CLAUDE-SNIPPET.md into ~/.claude/CLAUDE.md with your wiki path filled in.
3. Create your wiki: cp -r wiki-template ~/my-wiki && cd ~/my-wiki && git init && git add -A && git commit -m "wiki: birth"

What you get

Path	What it is
skills/wiki-ingest, skills/wiki-lint	The two LLM-wiki operations: source → cited page, and the weekly maintenance pass
skills/deep-research	Research orchestrator: scope → discovery → triage → capture → evidence matrix → agent-written brief
skills/advanced-web-search	Google-operator search rendered in real Chrome (via OpenCLI), deduplicated to clean JSON
skills/crawl4ai	Raw capture: Crawl4AI turns URLs into LLM-ready markdown with quality checks
hooks/wiki-log-marker.mjs	SessionEnd/PreCompact: one-line session marker appended to the wiki daily log + auto-commit
hooks/wiki-catchup.mjs	SessionStart: detects debt left by previous sessions and asks the new one to pay it down
hooks/heavy-task-wiki.mjs	PostToolUse: after a push/merge/deploy, reminds the agent to update the wiki now
hooks/precompact-session.mjs	PreCompact: exports the whole chat to markdown + scaffolds a recap stub before compaction
hooks/claude-toast.ps1	Windows desktop toast when Claude finishes or needs attention
agents/architect.md	Planning-only subagent — outputs PLAN/CHANGE/RISK/ORDER/VERIFY, refuses to write code
agents/reviewer.md	Bug/security-focused review subagent with a hard verdict format
commands/review.md	/review — pre-commit pipeline: diff → bugs/security/perf → SHIP IT / NEEDS WORK / BLOCKED
commands/boot.md, commands/evolve.md	A self-learning loop: load learned rules each session, promote/prune them weekly
rules/core-invariants.md	Always-on rules: never commit secrets, grep before creating, smallest change wins
statusline/	Status line (PowerShell + bash): model, context-usage bar, tokens, project
SENIOR-LOOPS.md	Seven senior-engineer build loops: spec → plan/blast-radius → test-first → security → critic → definition-of-done
wiki-template/	The empty wiki skeleton (see CONVENTIONS.md for the full schema)

Requirements

• Claude Code — the skills are plain markdown and port to other agents too (each research skill ships an agents/openai.yaml)
• Node.js (for the hooks)
• For the research skills only: Python 3.10+, OpenCLI (search), and pip install crawl4ai (capture). Everything degrades gracefully if missing.

Design principles

• The wiki is the LLM's domain. You browse it; the agent maintains it. Thin index, cited pages, append-only logs, delete-when-wrong.
• Search results are leads, not evidence. The research skills enforce direct source inspection, labeled confidence, counter-evidence, and gaps in every brief.
• Hooks guarantee the floor; the model writes the meaning. Even a crashed session leaves a committed one-line trace in the wiki log.
• Smallest change wins. The invariants and loops exist to make the agent behave like a senior engineer, not a keen intern.

Credits

• The LLM-wiki model (raw → wiki → schema, ingest one at a time, qmd at scale) comes from Andrej Karpathy's public notes on LLM-maintained wikis.
• The deep-research skills are derived from Nous Research's Hermes agent skills — see NOTICE.

License

MIT — see LICENSE.