Skip to content

docs: add skill review guidance to CLAUDE.md#460

Open
vanceingalls wants to merge 1 commit intomainfrom
vance/skill-review-guidance
Open

docs: add skill review guidance to CLAUDE.md#460
vanceingalls wants to merge 1 commit intomainfrom
vance/skill-review-guidance

Conversation

@vanceingalls
Copy link
Copy Markdown
Collaborator

@vanceingalls vanceingalls commented Apr 23, 2026

Summary

  • Adds a "Reviewing skill changes" section to CLAUDE.md with guidance for reviewing skills/ diffs
  • Frames review through three lenses: where agents misinterpret, silently skip, or produce wrong output on first attempt
  • Reminds reviewers that agents have no memory between compositions and must survive in 300+ line instruction files

Why

Skill review requires a different mindset than code review — instructions that look correct to a human reader can still fail when an agent processes them. This codifies the review heuristics we've been applying manually.

Test plan

  • Read the added section for clarity and completeness
  • Verified CLAUDE.md renders correctly

🤖 Generated with Claude Code

Copy link
Copy Markdown
Collaborator Author

This stack of pull requests is managed by Graphite. Learn more about stacking.

Comment thread CLAUDE.md
Comment on lines +72 to +88
### Reviewing skill changes

When reviewing a changeset that touches `skills/`, review as an agent-instruction
designer who has shipped many skills and watched agents fail on each one. Focus on:

- **Where will agents misinterpret?** Ambiguous vocabulary, instructions that assume
context the agent won't have 200 lines later, field names that don't match across
files (e.g., a picker emits "Canvas" but an eval checks for "bg").
- **Where will agents silently skip?** Soft suggestions agents will deprioritize,
gates that lack enforcement, steps with no verification that they ran.
- **Where will agents produce wrong output on first attempt?** Missing examples
(agents need to see the shape, not just read the rules), under-specified formats,
instructions that require multi-step reasoning to combine.

Assume the agent reads the full SKILL.md before authoring and has no memory
between compositions. Consider: will this instruction survive in a 300+ line
file? Will the agent connect Step 0's context to a quality check 250 lines later?
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also add a ci check that each skill change should have a SbS HTML attached comparing prev skill/s vs updated skill/s

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants