Work to AI Execution Manual

Problem to solve

Teams often finish a project, fix a workflow, or collect hard-won lessons, but that knowledge remains in chat logs, verbal notes, and scattered docs. An agent may still ask 'what next', repeat the same mistake, or fail to verify completion. This skill focuses on turning completed work into a reproducible execution manual instead of writing a vague document first.

How it works

• Choose the path: separate new skill creation from updating an existing skill to avoid overwriting working content.
• Follow the stages: identify fit, extract experience, record teaching, package files, validate reproducibility, and desensitize data.
• Make gotchas executable: write 'if X then Y', not 'watch out for this trap'.
• Keep structure tight: keep SKILL.md as a clear entry point, move detail beyond about 30 lines into references/, and include completion checks.
• Handle updates and privacy: apply incremental updates with SemVer, and desensitize eight sensitive categories when selected by the user.

Boundaries

Best for completed, repeatable work with clear validation. Not suited for one-off tasks, generic tools, or project-only conventions. The skill should come from real failure records and teaching sessions, not theory alone; if it contains internal information, desensitization must be confirmed before delivery.