writing-skills

Create a directory containing a SKILL.md file with YAML frontmatter defining name, description, and git_hash, followed by the instruction body. Run baseline pressure scenarios with subagents first, then write the minimal skill addressing observed failures.

Test skills by running pressure scenarios with subagents, first without the skill to capture baseline failures, then with the skill to verify compliance. Combine three or more pressures like time, sunk cost, and authority, and iterate until no new rationalizations appear.

The description should only state triggering conditions, starting with "Use when..." and listing specific symptoms and situations. Never summarize the skill's workflow, because agents may follow the description instead of reading the full skill body.

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

Agents rationalize violations when facing time pressure, sunk costs, or authority conflicts. Counter this with explicit negations of each loophole, rationalization tables, red flag lists, and authority-based language like "No exceptions."