Skip to content

Route the RL skills on their descriptions - #72

Open
jjroelofs wants to merge 1 commit into
1.xfrom
fix/71-skill-audit
Open

jjroelofs wants to merge 1 commit into
1.xfrom
fix/71-skill-audit

Conversation

@jjroelofs

Copy link
Copy Markdown
Contributor

Fixes #71

Summary

Prompt-audit fixes for the agent skills that drush rl:setup-ai installs, applied identically to each .claude and .agents copy:

  • rl (medium): removed the triggers: list, which Claude Code ignores, and replaced the description with one that names the drush rl:* command groups (list, inspect, analyze, create, update, delete, settings) and says when to use the skill: experiments, A/B or multivariate tests, variants, bandits, conversion rates, or which variant is winning.
  • rl_page_title and rl_menu_link: same treatment, since rl:setup-ai installs them alongside rl and they had the same keys. Their descriptions now name their rl:page-title:* and rl:menu-link:* commands and when to use them.
  • All three (low): version: 1.0.0 moves to metadata.version, the place the Agent Skills specification allows.

RlSetupCommands copies and compares the files and never parses the frontmatter, so no code change is needed. Every command named in the descriptions exists in src/Drush/Commands or the submodule's commands class.

Left out on purpose

  • The submodule skill names rl_page_title and rl_menu_link contain underscores, which the specification does not allow. Renaming them changes the installed directory names that RlSetupCommands derives from the module name, so it belongs in its own change (noted in Agent skill: route on the description and drop non-standard frontmatter keys #71).
  • The skill bodies have pre-existing markdownlint findings (long table rows, missing blank lines around two headings). This repository has no markdownlint configuration and this PR leaves the bodies untouched.

Testing

  • skills-ref validate passes for both rl skill directories (it failed before on "Unexpected fields in frontmatter: triggers, version"). The submodule skills now only fail on the underscore name described above.
  • Each .claude/.agents pair is byte-identical (cmp), as the skills pair check requires.
  • markdownlint-cli2 reports no new findings; the remaining ones are the same pre-existing body findings, shifted up by the shorter frontmatter.

The rl, rl_page_title and rl_menu_link skills kept their routing words
in a `triggers:` list, which Claude Code does not read: it chooses a
skill from the description alone, and current Claude models follow that
text literally. Nothing in this repository reads the key either.
Remove it and rewrite each description to say what the skill does and
when to use it.

Move the non-standard top-level `version` key under `metadata`, as the
Agent Skills specification requires. RlSetupCommands only copies and
compares the files, so no code reads it.

Fixes #71

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.

Agent skill: route on the description and drop non-standard frontmatter keys

1 participant