paoding-skill

庖丁 · Skill 解剖术

Purpose

Turn a skill package into useful creator notes. Explain how the skill changes an agent's behavior, what bad output it prevents, which parts are worth copying, and which parts are too heavy for the job.

Default to the smallest useful output. A web handbook is an output mode, not the identity of this skill.

Start Here

1. Map the source package first.
2. Choose the output mode from the user's request:
   • Review / keep-cut advice: write a direct Markdown review.
   • Pattern extraction: write human notes and optional JSON/YAML.
   • Web handbook / 解剖手册: read references/handbook-spec.md, then references/web-production-flow.md.
3. Load only the companion reference needed for the current section. Examples are calibration only; references and source files are the contract.

Workflow

1. Map the Package

Treat the input as a single SKILL.md, a skill directory, or a folder of skills.

• Find every SKILL.md.
• Note references/, scripts/, assets/, examples/, tests/, and metadata.
• Read the entry SKILL.md first.
• Read referenced files only when they explain the skill's design.
• Inspect scripts to see which fragile jobs they take away from the agent — and run every script that has a valid in-package input (mandatory; see 可执行工件必跑 in references/evidence-collection.md).

For a batch, analyze representative skills first, then scan the rest for repeated structures.

2. Find the Real Task Pain

Read references/pain-dimensions.md, then run its item-driven sweep:

1. Imagine doing the task bare (a human novice, then a default agent) and list where each would fail.
2. Inventory every rule, script, intermediate artifact, and checkpoint in the package. List the intermediate artifacts completely.
3. Run each item through the three-question pain test (counterfactual, baseline, evidence grade) and tag it with one of the seven dimension labels.
4. Reconcile: pains the skill catches that you missed, pains you expected that it ignores (blind spots), and residue that fits nowhere.