agent-standards

Shared coding standards for AI agents.

Usage

Add as a git submodule:

git submodule add git@github.com:block65/agent-standards.git agent-standards

Instructions live in AGENTS.md, which is tool-neutral and read by every agent that follows the convention. CLAUDE.md holds one import line and nothing else, so Claude Code reads the same document as everything else and there is only one file to maintain:

@AGENTS.md

Example

For a rust crate, in your project's AGENTS.md:

**Always:** Read `agent-standards/index.md`, then follow all "Always load" standards listed there.
**Before writing or modifying Rust code:** STOP. Read `agent-standards/lang/rust.md` and follow it.

See examples/ for both files.

index.md is the barrel, so a new "Always load" entry needs no AGENTS.md edit in consuming repos. Consumers track latest, not a pin: set the submodule branch to main and sync with git submodule update --remote agent-standards, wired into the repo's bootstrap/sync task so it runs on every pull. Pin a SHA only when a repo explicitly opts out of latest. Only hardcode task-specific standards.

Assumptions

These standards assume the compend CLI (from the block65/compend repo) is on PATH. Standards docs defer library and language facts to it rather than copying them (e.g. lang/rust.md → compend get rust-book <topic>). Where it is not installed, follow the standards alone; do not substitute web fetches for compend lookups.

Code review runs through the pear CLI, which carries its own pear-review skill (pear skill install). The reviewer reads the diff in a browser; the agent attaches to the review and services comments as they arrive. See pear's own README for usage: it is a tool, not a standard, so nothing here restates it.