Writing Skills treats creating or editing a skill as test-driven development for process documentation. You run a pressure scenario with subagents to see how the agent fails without the skill, write the skill to address those failures, verify the agent now complies, and refactor to close new loopholes. Guidance covers skill types, directory layout, descriptions, flowcharts and code examples.
What You Get
- A mapping from TDD concepts to skill creation
- Guidance on skill types and file organization
- Rules for descriptions, keywords and naming in a discovery reference
- Flowchart conventions with a Node script that renders the diagrams
- Anthropic’s official skill authoring best practices as a reference
Common Use Cases
- Creating a new skill for a repeated technique
- Editing an existing skill and re-testing it
- Testing a discipline skill under time and sunk-cost pressure
What Makes It Unique
It tests documentation the way code is tested: the baseline failure of an agent without the skill is the failing test. The skill spells out the rationalizations agents use and asks authors to add an explicit counter for each.
Under The Hood
Instructions with reference files and a render-graphs.js script. The script extracts the dot blocks from a skill’s SKILL.md, runs the local graphviz dot program on them and writes SVG files next to the skill, and it exits with install hints when graphviz is missing. The agent writes skill files to your skills directory, such as ~/.claude/skills, and dispatches subagents for pressure tests. One graphviz conventions file ships in a format the label could not read.