ai-stack — a shared brain for AI coding assistants

One set of rules, skills and agents that Claude Code, Codex, Gemini CLI, GitHub Copilot CLI and OpenCode read from the same place, for every account on a machine.

📖 Setup & usage · 🇪🇸 Léeme en castellano

Why

Most people configure each assistant separately: one CLAUDE.md here, an AGENTS.md there, skills duplicated in three places. Change a rule and you change it four times — or forget three.

This puts the rules, the skills and the agents in one directory, symlinked into every assistant and readable by every user on the machine. Update it once; everyone gets it.

What's inside

174 skills	Invocable playbooks: security auditing, SEO, marketing, frontend, context engineering, documentation, translation…
223 agents	Specialist subagents in Claude subagent format, with adapters for other assistants
2 protocols	@protocol-enforce opens a task; @protocol-end closes and documents it
3 hooks	Shared memory injection · anti-sycophancy enforcement · documentation-debt tracking

Catalogs: skills/SKILLS_CATALOG.md · agents/AGENTS_CATALOG.md

The parts worth stealing

Even if you don't adopt the whole thing:

• scripts/hooks/anti-sycophancy.js — detects praise, agreement without evidence and filler sign-offs. On Claude Code it hooks into Stop and rejects the reply, forcing a rewrite. It distinguishes "you're right because the log shows X" (fine) from "you're right" alone (not fine). No dependencies, --self-test included.
• scripts/hooks/doc-debt.js — answers "what did I touch and never document?" by comparing each project's last-close mark against file mtimes. Computed from disk state rather than a session event, so every assistant sees the same debt — even the ones without hooks.
• skills/protocol-end/ — a close-out protocol that edits existing files instead of rewriting them, and verifies what it wrote before reporting success.

Requirements

Node.js 18+ (standard library only, no npm install) and Bash. That's it.

Install

git clone https://github.com/bootablearg/ai-stack-agents-and-skills.git /opt/ai-stack
ln -s /opt/ai-stack/AGENTS.md ~/.claude/CLAUDE.md
ln -s /opt/ai-stack/skills ~/.claude/skills
ln -s /opt/ai-stack/agents ~/.claude/agents

Full instructions — multi-user setup, hooks, the other four assistants — in SETUP.md (castellano).

Design rules

1. Knowledge only. No execution state, no session logs, no credentials. If it describes what happened, it belongs in your project, not here.
2. No absolute paths, no hardcoded users. Scripts resolve their own location; the rest uses $AISTACK_HOME, $WORKSPACE, $HOME.
3. Everything portable. Skills are plain markdown; any assistant can read them as context even without native skill support.

License

MIT — see LICENSE. Bundled third-party skills keep their own licenses and copyright notices; attribution in NOTICE.md. Note that vendor/trailofbits-skills/ is CC BY-SA 4.0: if you build on that folder specifically, share-alike applies to your derivative.