Project context

crab’d reads the context your repository already keeps for coding agents — its AGENTS.md / CLAUDE.md instructions and its skills — and folds them into each run. The agent that answers a mention, reviews a PR, or implements an issue then follows the same conventions your local agents do, with no duplication into config.

Both are on by default and controlled by the context config section.

Instruction files

crab’d loads AGENTS.md, then CLAUDE.md, from the checkout root and appends them to the system prompt:

• Both are honored. AGENTS.md is the cross-tool convention; CLAUDE.md keeps parity for repos coming from claude-code-action. If both exist with the same content (a common symlink/copy), it’s included once; if they differ, each is included under a labeled heading.
• crab’d’s own rules stay above them. Project instructions are appended after the base prompt and your prompt.instructions, and the prompt states that core instructions win on conflict. This keeps a repo-authored file from silently overriding crab’d’s guardrails.
• Capped at 40k characters combined, so a large file can’t crowd out the diff and conversation.

Only the checkout root is read (not nested AGENTS.md files) — the agent can still open nested files itself with its file tools while it works.

# .crabd.yml — opt out if you'd rather steer purely from config
context:
  instruction_files: false

Skills

crab’d discovers skills under .agents/skills/ and .claude/skills/. For each skill it reads the name and description from the SKILL.md frontmatter and lists them in the prompt:

## Available skills
- **run-tests** — Use when the user wants to run the suite. (`.claude/skills/run-tests/SKILL.md`)

This is progressive disclosure: only the one-line description is in the prompt. When a task matches a skill, the agent reads that skill’s SKILL.md itself (with its file tools) for the full instructions — the body is never preloaded, so many skills cost almost no context.

• A skill in both roots is listed once (.agents/skills wins).
• A skill with no description is skipped — without one, the agent has no basis to pick it.

context:
  skills: false

Note

This is different from the crab’d skills page, which lists Claude Code skills for humans building with crab’d (configure-crabd, add-crabd-to-repo, …). Those help you in your editor; the skills here are your repo’s skills, surfaced to the crab’d agent at run time.

Layering

context.instruction_files and context.skills are booleans, so they follow the usual highest-layer-wins rule. An org can turn the whole behavior off for every repo by setting them in the org config and locking the paths:

# org .crabd.yml
context:
  skills: false
governance:
  locked: [context.skills]