Agent config for ontology-driven LLM development. One source, minimal surface, secure by default.

One file. One source. No drift.

Rules here are organizational scar tissue — grown from real incidents, not upfront design. A rule gets added the second a mistake repeats. Nothing is added speculatively.

• One source. Edit AGENTS.md only. build.sh generates all tool-specific aliases.
• Minimal surface. Short rules shape judgment. Long rules get ignored — and the middle ones get ignored most. Research shows LLMs reliably attend to the beginning and end of long contexts, with significant degradation in between (Liu et al., 2023 — "Lost in the Middle"). Keep AGENTS.md short and front-load what matters.
• Ontology-grounded. Models, schemas, and knowledge graphs are first-class.
• TDD. Every function ships with a test. Red → green → refactor, always.
• Secure. No hardcoded paths, tokens, or secrets. Ever.

ai-config/
├── AGENTS.md              ← single source of truth — edit this only
├── build.sh               ← generates all tool folders + wires git hooks
├── claude/
│   ├── CLAUDE.md          ← generated
│   ├── settings.json      ← model, hooks, statusline config
│   ├── agents/            ← 8 specialist sub-agents
│   ├── hooks/             ← session management, context injection, statusline
│   └── commands/          ← /-check, /-ship, /-review, /-discover, /-session, /-sync
├── cursor/
│   ├── user-rules.md      ← generated — paste into Settings → Rules for AI
│   └── rules/
│       └── global.mdc     ← generated — drop into .cursor/rules/ per project
├── gemini/
│   └── GEMINI.md          ← generated
└── codex/
    └── AGENTS.md          ← generated

Each tool folder is built exhaustively. You only need the generated config file for your tool — the rest (agents, hooks, commands) is Claude Code-specific.

The tool folders exist purely for convenience. All you actually need is AGENTS.md from the root — drop it into whatever your agent reads. The per-tool generated files just save you a copy-paste or a config lookup.

1. Clone and build:

git clone https://github.com/PR0CK0/ai-config.git ~/ai-config
cd ~/ai-config
./build.sh        # generates all tool folders + wires pre-push hook

2. Claude Code — rules only (drop-in):

cp ~/ai-config/claude/CLAUDE.md ~/.claude/CLAUDE.md

Claude Code — full setup (agents, hooks, slash commands, statusline):

cp ~/ai-config/claude/CLAUDE.md ~/.claude/CLAUDE.md
cp ~/ai-config/claude/settings.json ~/.claude/settings.json
cp -r ~/ai-config/claude/agents ~/.claude/agents
cp -r ~/ai-config/claude/hooks ~/.claude/hooks
cp -r ~/ai-config/claude/commands ~/.claude/commands
chmod +x ~/.claude/hooks/statusline.sh

3. Cursor — global rules: Paste cursor/user-rules.md content into Cursor → Settings → Rules for AI.

Cursor — per-project:

mkdir -p .cursor/rules && cp ~/ai-config/cursor/rules/global.mdc .cursor/rules/

4. Gemini CLI:

mkdir -p ~/.gemini && cp ~/ai-config/gemini/GEMINI.md ~/.gemini/GEMINI.md

5. Codex CLI:

cp ~/ai-config/codex/AGENTS.md ~/AGENTS.md   # global
# or copy to project root for per-project

Per-project pointer (any tool):

echo "READ ~/ai-config/AGENTS.md BEFORE ANYTHING" > AGENTS.md

Symlinks instead of copies:

./build.sh --symlink

build.sh wires .githooks/pre-push automatically — if AGENTS.md changes, the hook rebuilds all aliases and blocks the push until generated files are committed.

MIT