potiuk opened a new issue, #80: URL: https://github.com/apache/airflow-steward/issues/80
Skill `description` and `when_to_use` frontmatter fields are loaded into the Claude Code system prompt **every turn** for routing, even when the skill body itself is not invoked. Verbose phrasing in those fields is therefore a permanent steady-state token tax across every session. ### Worked example — [apache/airflow#66444](https://github.com/apache/airflow/pull/66444) Trimming `maintainer-review`, `pr-triage`, and `pr-stats` frontmatter saved roughly **500 tokens/turn**: ``` maintainer-review: ~420 → ~170 tokens pr-triage: ~300 → ~125 tokens pr-stats: ~170 → ~105 tokens ``` The `description` / `when_to_use` fields only need what the *router* uses: what the skill does, how it picks targets where that matters, trigger phrases. Anything else (file-path cross-refs, mutation enumerations, history footnotes, "successor to X" notes) belongs in the SKILL.md body — which loads only on invocation. ### Proposed - Add a "caveman" (or similarly named — terse-by-design) **write-skill** that all skill authoring and editing flows pipe through. - The skill enforces lean phrasing in frontmatter: drop the categories above, keep what the router needs. - Update existing skill-authoring guidance and any agent prompts to reference the new write-skill so it gets used by default. - Sweep existing skills once, applying the new style. **Why:** every token in skill frontmatter is paid on every turn across every session; lean phrasing is a multiplicative win. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
