Iteration & skill-creator
Use skill-creator to bootstrap, then iterate based on under/over-triggering signals and execution failures.
Bootstrap With skill-creator
skill-creator is a skill that creates skills. It generates properly formatted SKILL.md with frontmatter, suggests trigger phrases, flags common issues (vague descriptions, missing triggers, structural problems), and helps iterate by absorbing your edge cases back into improvements. If you have an MCP server and know your top two or three workflows, skill-creator can produce a functional skill in 15–30 minutes.
# Kick off
Use the skill-creator skill to help me build a skill for [your use case]
# Iterate after finding an edge case
Use the issue & solution identified in this chat to improve how the skill handles [specific edge case]skill-creator helps you design and refine skills. It does not execute automated test suites or produce quantitative evaluation results — that still requires manual testing in Claude.ai, scripted tests in Claude Code, or the Skills API.
Iterate Based on Feedback Signals
Skills are living documents. Three signal clusters tell you exactly what to change, and they require different fixes — confusing them is the fastest way to make the skill worse.
| Signal Cluster | What You See | Fix |
|---|---|---|
| Undertriggering | Skill doesn't load when it should; users manually enable it; support questions about when to use it | Add more detail and nuance to the description, including domain-specific keywords |
| Overtriggering | Skill loads for irrelevant queries; users disable it; confusion about purpose | Add negative triggers, narrow scope, mention what the skill does NOT do |
| Execution issues | Inconsistent results; API call failures; user corrections needed | Improve instructions, add error handling, add explicit validation steps |
Debug Trigger Problems
When a trigger-related issue surfaces, ask Claude directly: "When would you use the [skill name] skill?" Claude will quote the description back. Compare what it echoes against the scenarios you expected — the mismatch tells you exactly which words or phrases are missing or pulling the skill into unrelated conversations.
# Before — overtriggers on any CSV
description: Advanced data analysis for CSV files.
# After — negative trigger plus narrowing
description: Advanced data analysis for CSV files. Use for statistical modeling, regression, clustering. Do NOT use for simple data exploration (use data-viz skill instead).When Claude Loads the Skill But Ignores It
Once the skill is loading correctly, the next failure mode is Claude loading it but not following the instructions. Three common causes:
- ▸Instructions too verbose — keep SKILL.md concise, move detail to references/.
- ▸Instructions buried — put critical steps at the top under ## Critical or ## Important headers.
- ▸Ambiguous language — replace "Make sure to validate" with explicit checks ("CRITICAL: Before calling create_project, verify name is non-empty, at least one team member assigned, and start date is not in the past").
For critical validations, bundle a script that performs the checks programmatically instead of relying on natural-language instructions. Code is deterministic; language interpretation is not. The Office skills are a canonical example of this pattern.