Agent Skill Description Optimizer

Skill Description Optimizer (no API key)

[](https://docs.claude.com/en/docs/claude-code) [](https://github.com/netopsengineer/agent-skill-description-optimizer/actions/workflows/release.yml) [](https://opensource.org/licenses/MIT)

Optimizes the `description:` frontmatter of a Claude Code / Agent **skill** - the text that decides whether Claude invokes the skill. Both halves of the loop (evaluation **and** improvement) run entirely through the `claude -p` CLI, so **no `ANTHROPIC_API_KEY` is required** - it reuses the auth the CLI already has.

Why this exists

Skill-creator's own `run_eval.py` already evaluates triggering with `claude -p` (no key). The *only* part of its optimization loop that needs an API key is the **improver** (`improve_description.py` / `run_loop.py`), which calls the `anthropic` Python SDK. This project replaces that piece: **both** the evaluation and the improvement go through `claude -p`.

Relationship to skill-creator

This is a re-implementation of skill-creator's optimization loop, intended to be usable in its place without an API key. Where it matches and where it deliberately differs:

Behavior	This tool	skill-creator
Trigger detection (stream tool-call intent)	Equivalent logic (see note)	run_eval.py