writing-skills

Create a directory containing a SKILL.md file with YAML frontmatter defining name and description, followed by markdown instructions. Test it first by running pressure scenarios without the skill to document baseline agent failures, then write content addressing those specific failures.

Run realistic pressure scenarios with subagents, combining three or more pressures like time constraints, sunk cost, and authority. Compare agent behavior with and without the skill loaded, and capture exact rationalizations verbatim to identify loopholes.

The description should start with "Use when..." and describe only triggering conditions, symptoms, and situations in third person. Never summarize the skill's workflow, because agents may follow the description shortcut instead of reading the full skill body.

Skip skill creation for one-off solutions, standard practices already well-documented elsewhere, project-specific conventions that belong in CLAUDE.md, and mechanical constraints that can be enforced with regex or validation scripts instead of documentation.

Agents rationalize rule violations when facing time pressure, sunk costs, or authority figures. Skills must explicitly close each loophole with negations, rationalization tables, and red flag lists, then be re-tested until agents comply under maximum pressure.