meta-skill-author

Write a description with what the skill does, when to use it, and its key capabilities, keeping the key use case first. Then run the description optimization loop: build about 20 realistic queries split 60/40 train/test, iterate up to 5 times, and select by test score to avoid overfitting.

Use the two-stage port method: first generalize the skill into a harness-neutral ports/generic.md with a capability-contract.yaml, then apply the target platform's adapter profile to produce the platform-specific deliverable. Each adapter maps required capabilities to native, partial, or absent support.

It targets Claude Code, Cursor, GitHub Copilot, OpenAI Codex, and Perplexity, with a dedicated adapter profile for each. The portable core uses open-standard frontmatter fields, while platform-specific extension fields are marked with compatibility annotations.

Generator-assessor separation prevents biased self-evaluation of one's own artifact. The package enforces this with separate Grader, Comparator, and Analyzer subagent prompts that run in isolated contexts and score against the four-discipline audit rubric.

Skip skill authoring when a plain prompt handles the task in one shot, when the logic is fully deterministic and belongs in a script, when you need an autonomous agent, or when an MCP tool exposing an external API suffices. The skill's pre-flight check maps each situation to the cheaper primitive.

Remove explicit chain-of-thought, few-shot examples, self-consistency, least-to-most decomposition, and skeleton-of-thought scaffolding, which degrade frontier reasoning models. Replace them with a Goal plus Constraints plus Context structure written as outcome-based declarative instructions.