| name | skill-construction |
| description | Use this skill when the user wants to build a new Claude skill from scratch, expand a single prompt into a multi-phase skill system, or add a self-improvement feedback loop to an existing skill. Triggers on phrases like "build a skill for", "design a prompt that won't break", "turn this prompt into a full workflow", "make this skill self-improving", "production-ready skill for X", or any request for new skill architecture (not just editing an existing prompt). Handles three construction modes: greenfield design with failure prevention, single-prompt expansion into a phased system, and feedback loop integration. Do NOT use for editing an existing prompt in place (use prompt-refinement) or for evaluation (use prompt-stress-test / prompt-reliability-score). |
Skill Construction
Build new skills or expand existing ones. Bake failure prevention in from day one.
Inputs required
Depends on mode. Ask which mode before proceeding if unclear.
Mode A — Greenfield
- What the skill should do
- Who uses it
- Example inputs it must handle
Mode B — Expansion
- The current single prompt
- The full workflow it sits inside (input → processing → validation → output)
Mode C — Feedback loop
- The current skill
- The recurring quality issues observed
Procedure
Mode A — Greenfield design
- Identify the 3 most likely failure modes for this task type.
- Design constraints, output format, and guardrails to prevent each.
- Draft the skill.
- Stress-test mentally against 5 edge cases.
- Deliver skill + failure prevention map.
Mode B — Single prompt → phased system
- Map the workflow into distinct phases.
- Write a dedicated sub-skill per phase.
- Define explicit handoffs — output of phase N is the exact input of phase N+1.
- Insert checkpoints between phases to catch failures before they propagate.
- Test the full chain against 3 realistic end-to-end cases.
Mode C — Feedback loop integration
- Design structured feedback capture — what to log, when, how.
- Build a reflection step into the skill that treats past failures as active constraints.
- Ensure the skill still works without the feedback layer — it enhances, does not replace.
- Define the improvement cycle trigger (evidence-based, not scheduled).
- Test against baseline.
Rules
- Prevention beats diagnosis — bake resistance in, don't bolt it on.
- Every constraint must have a named reason. No rules without purpose.
- Edge cases are tested before delivery, not after.
- Output format is locked. Ambiguity is the enemy of consistency.
- In Mode B, each sub-skill must work independently — no phase may silently depend on another.
- In Mode C, feedback must be structured. Free-form notes do not improve prompts.
Output format
- Mode — A / B / C
- Skill Draft (A) / Phase Map + Sub-skills (B) / Feedback Layer Design (C)
- Failure Prevention Map — table: failure mode | prevention mechanism
- Edge Case / End-to-End Test Results — 5 cases (A, C) or 3 cases (B)
- Production-Ready Artifact — the final skill(s), ready to deploy
