# Meeting-action skill: author-applied examples

A concrete meeting-actions skill was authored from the pinned skill-creator instructions. Its description scopes the task to committed actions in meeting records. The draft requires source quotes, literal `unknown` values for missing fields, and no meeting actions for unrelated inputs.

The author applied the draft to three supplied fixtures in the same conversation and saved the actual authored responses. The output-check script reads those files; it does not generate answers. Four assertions passed: named owners and dates, unknown missing fields, an unrelated recipe producing no meeting actions, and exact source spans for all actions. The local quick_validate.py check also passed. Its filename and hash are recorded in the manifest.

These examples are transparent author-applied samples. There was no nested model invocation, independent grader, automatic selection test, or no-skill control. They show the draft and its intended output contract; they do not establish reliable trigger selection or improvement.

## Comparison plan — not run

1. Start a fresh agent context for each of the three inputs. Preserve exact prompts, full output and disclosed runtime/model metadata.
2. In the with-skill arm, install the complete draft folder and record whether it was selected automatically. If explicitly invoking it, label that separately.
3. Run the identical user requests in fresh contexts without this skill. Do not show the expected outputs to either arm.
4. Score both outputs against the same predeclared checks, preferably with a grader unaware of the arm. Preserve failures, retries and sample counts.
5. Report measured differences only after both arms actually run. Repeat with additional unseen records before making a broader claim.

For the unrelated recipe, a normal agent may fulfill the recipe request using other capabilities. The narrow meeting-actions extraction result is simply not applicable; this saved example does not test the host agent's routing.

## Provenance

Pinned source: [anthropics/skills at 8a1541c4a3ffa5a20a5a91de0dcf3f0bab1d1ef4](https://github.com/anthropics/skills/blob/8a1541c4a3ffa5a20a5a91de0dcf3f0bab1d1ef4/skills/skill-creator/SKILL.md). Exact source hashes and references are in sources.json.

Executed on 2026-10-04 by Codex in this conversation using Node.js 24.13.0 and Python 3.12.3 on Linux. The exact model version was not exposed. Original command timestamps and stream hashes are in commands.jsonl. These records do not establish independent human adoption, customer outcomes or a compatibility certification.
