Skip to content

agent-skills.md

Latest commit

 

History

History
70 lines (50 loc) · 2.2 KB

agent-skills.md

File metadata and controls

70 lines (50 loc) · 2.2 KB
Raw

Agent Skills

This repository exposes a small set of repo-local slash-command workflows for agent sessions. They are prompts, not Workshop actions. They do not replace the role bootstrap, Workshop confinement, or verification matrix in AGENTS.md.

Slash-command definitions live in .claude/commands/:

  • /tdd: .claude/commands/tdd.md
  • /grill-with-docs: .claude/commands/grill-with-docs.md
  • /diagnose: .claude/commands/diagnose.md

Shared Rules

These workflows must preserve these repository rules:

  • Run agent-role <role> and context before planning, editing, reviewing, or running checks.

  • Run project shell commands through Workshop, usually with:

    tools/workshop-shell -c 'workshop run <project-alias> -- <action>'

  • Follow the Router, Architect, Implementer, and Reviewer boundaries in .agents/roles/.

  • Use named Workshop actions from .agents/commands.md whenever possible.

  • Preserve generated-artifact, secret-handling, and Git-push boundaries from AGENTS.md.

/tdd

Use /tdd for a narrow red-green-refactor loop.

Expected behavior:

  • identify the smallest testable behavior
  • write or update one focused failing test first
  • run focused tests through Workshop
  • implement only enough to pass
  • refactor only after the focused test passes
  • finish with the relevant verification from .agents/workflows/verification.md

/grill-with-docs

Use /grill-with-docs to challenge a proposal or explanation against the repository's documented constraints.

Expected behavior:

  • read AGENTS.md and the relevant linked docs first
  • distinguish documented facts from assumptions
  • ask pointed questions where the proposal conflicts with policy or lacks acceptance criteria
  • require Workshop-routed verification for claims involving project commands
  • recommend the smallest next role, question, check, or change

/diagnose

Use /diagnose for failure triage.

Expected behavior:

  • reproduce with the narrowest Workshop-routed command
  • capture the exact command and failure
  • classify the likely affected area
  • inspect only the files needed to explain the failure
  • propose the narrowest next check, fix, or role handoff
  • avoid broad rewrites and unrelated cleanup