designing-workflow-skills

Design reliable multi-step workflow skills by enforcing structured SKILL.md construction with numbered phases, explicit entry and exit criteria, and a final verification step. This ensures deterministic execution and verifiable outputs for sequential pipelines.

Skills fail to activate when the frontmatter description lacks exact trigger conditions. Activation-safe design ensures the description matches specific request intents, allowing the skill to load correctly for the target workflow tasks.

Keep SKILL.md compact and push detailed patterns, anti-patterns, and workflow steps into references/ and workflows/ directories using one-hop links. This separation maintains clarity and prevents the main file from becoming unmanageable.

Prevent Cartesian-product tool-call explosions by enforcing least-privilege allowed-tools and bounding tool-call patterns. Avoid unbounded subagent spawning to maintain tooling and scalability discipline within the workflow design.

Yes, you can implement safety gates for destructive actions by routing between independent paths and applying task-driven execution with progress tracking. This ensures destructive operations are gated and verified before proceeding.

Phased execution with entry and exit criteria is needed when designing skills with sequential pipelines, routing between independent paths, or task-driven execution. It ensures each phase is verified before the next begins, producing dependable outputs.