agent-rules

A public, work-free operating manual for running coding agents (Claude Code, Codex CLI, OpenCode, and similar) — dogfooded configs, rules, and workflows kept in sync across machines.

It works as a config hub: core.md holds the shared, tool-agnostic rules; map says where each tool reads its global instructions; setup/install.sh symlinks each installed tool's load-point back at the hub. Edit the hub once, every agent gets it. A new machine is git clone + ./setup/install.sh. Start at AGENTS.md, then docs/setup.md.

What's here

path	what
core.md	the shared base rules every agent tool loads
map / workflow-map / prompt-map	tool → load-point → hub-file; workflow → skill catalogs; prompt template → runtime prompt dir
AGENTS.md	repo-local rules for working in this checkout (CLAUDE.md symlinks to it)
check-privacy.sh	pre-commit privacy guard; identity patterns live outside the repo
docs/	the manual: setup, memory & recall, compaction, model & quota, drift, custom providers
rules/	composable rule snippets
workflows/	own generic skills (council, drain-memory, handoff, wrap, your-voice)
setup/	install.sh, doctor.sh, example configs (macOS/zsh)
tools/recall/	a semantic memory-recall CLI

Quick start

Clone into a dot folder (e.g. ~/.agent-rules), not a project folder — every tool's global load-point points back at this checkout.

git clone git@github.com:Jetemple/agent-rules.git ~/.agent-rules   # or HTTPS
cd ~/.agent-rules
# read docs/setup.md first — prereqs (Homebrew, Python 3.13, llama.cpp) and the order
./setup/install.sh --dry-run    # preview
./setup/install.sh              # create symlinks
./setup/doctor.sh               # verify

The installer also registers every skill declared in workflow-map by linking it into each listed catalog (agents, codex, claude). For machine-private skills or source overrides, create ~/.config/agent-rules/workflow-map (same three-column format) — a private entry with the same name replaces the public one:

wrap ~/.private/wrap agents,codex,claude

Runtime-specific slash-command prompt templates (Pi's /goal, for now) are registered in prompt-map and linked as <name>.md into the runtime's prompt directory (pi → ~/.pi/agent/prompts). They stay separate from workflow-map because their invocation and interpolation syntax is runtime-specific. A pre-existing real file that is byte-identical to the repo copy is adopted (replaced by the link); a differing one is left alone and reported.

Dependencies (not republished here)

• Superpowers — a separate skills marketplace; several workflows here assume it is installed. This repo links it as a dependency, doesn't republish its skills.
• embeddinggemma-300M (GGUF) — the embedding model tools/recall uses, served by llama.cpp (llama serve). Pull it per step 3 of docs/setup.md.

Platform

macOS / zsh. Windows is not supported yet.

License

MIT — see LICENSE.