PDCA Scaffold
A tool for generating domain-specific PDCA skills. You describe a complex repeatable task; this skill guides you through structured discovery, then generates a valid installable Claude skill tailored to your domain — with a built-in learning loop that sharpens the skill after each cycle.
How This Works
Phase 1: Discovery — Socratic questioning across 5 layers to understand your task, how human and AI responsibilities divide, what quality looks like, where intervention is needed, and what you want to learn from each cycle.
Phase 2: Generation — A domain-specific PDCA skill is generated in your project directory. The skill includes phase prompts, working agreements, quality gates, and intervention triggers derived directly from your discovery answers.
Phase 3: Refinement (per cycle) — After each ACT retrospective in the generated skill, the refinement protocol proposes specific diffs back to the skill's reference files. You approve each change. The skill gets sharper without growing longer.
Start Here
Read references/discovery-guide.md and begin Layer 1 questions. Do not skip layers or combine them — each layer builds on the previous. The confirmation gate after Layer 5 is mandatory before generating anything.
After Generation
The generated skill goes in [domain]-pdca/ in the user's current project directory. It is committed to the project repo and version-controlled alongside the work it governs.
Before using the skill in production, validate it with /skill-creator:
- Verifies SKILL.md format and that the description triggers correctly
- Runs with-skill vs. without-skill eval comparisons to confirm behavioral effectiveness
- Surfaces improvements before the first real cycle
A generated skill that hasn't been through eval is an untested assumption. /skill-creator is the quality gate between generation and production use.
Reference Files
references/discovery-guide.md— The 5-layer Socratic question framework with confirmation gatereferences/generation-templates.md— Templates for SKILL.md and all references files in the generated skillreferences/refinement-protocol.md— The active learning loop: how to propose, approve, and commit skill refinements
Anti-Patterns to Avoid
Skipping the confirmation gate — The summary after Layer 5 is the primary HITL checkpoint. Errors in the summary propagate into every future cycle. Always present and confirm before generating.
Generating generic rules — "The human owns the important parts" is not a working agreement. Push for specificity: which parts, which decisions, which moments.
Rewriting instead of refining — Refinement is a diff, not a rewrite. If the generated skill needs substantial changes, use /skill-creator for a structured revision session.
Growing the skill longer each cycle — The anti-drift rule is ≤ +10 net lines per refinement. For every addition, identify something to narrow or remove.