| name | skill-creator |
| description | Load when creating, updating, adapting, or evaluating a standard agent skill; do not load for ordinary documentation edits or one-off prompt advice. |
| license | Anthropic primary terms in LICENSE.txt; Matt Pocock additions in LICENSE.mattpocock.txt |
| metadata | {"source":"anthropics/skills skills/skill-creator","upstream_commit":"690f15cac7f7b4c055c5ab109c79ed9259934081","additional_source":"mattpocock/skills skills/productivity/writing-great-skills","additional_upstream_commit":"e9fcdf95b402d360f90f1db8d776d5dd450f9234","additional_license":"MIT; complete terms in LICENSE.mattpocock.txt","adaptation":"Platform-neutral Anthropic base with selected Matt Pocock completion, context-load, leading-word, and pruning guidance; host-specific eval scripts and UI metadata are excluded."} |
Skill Creator
Use this skill to create or improve a standard skill folder with SKILL.md plus optional references/, scripts/, and assets/.
Creation Workflow
- Define the trigger contract: real user phrases, positive cases, negative cases, and forbidden loads.
- Decide whether a skill is warranted. Prefer project docs or root instructions when the behavior is broad, obvious, or already covered.
- Budget context load before choosing invocation. A model-routed description is present in the Host's initial Skill inventory, so each word and each new Skill must earn its recurring cost; side-effect-heavy or rarely used Skills stay explicit-only.
- Write a concise
description that front-loads a useful leading word, starts from user intent, includes the nearest dangerous overlap, and avoids workflow internals or duplicate synonyms.
- Give every ordered step a checkable completion criterion. State the evidence that distinguishes complete from merely attempted, especially where premature completion would be costly.
- Keep
SKILL.md short. Move branch-specific examples, schemas, API tables, templates, or scripts behind precise context pointers in spokes.
- Record source, license, and upstream commit when adapting third-party material.
- Register the Skill in the install graph only after path, routing, overlap, and risk are clear.
- Validate with the repository's nearest Skill validator, routing tests, install checks, and the declared completion criteria.
Quality Gates
- The skill changes agent behavior in a recurring, bounded way.
- The name is lowercase hyphen-case and matches the folder.
- The description is narrow enough to avoid trigger noise.
- The body contains decisions and gotchas, not generic advice.
- Heavy or volatile content uses progressive disclosure.
- Side-effect-heavy workflows are explicit-only until safety boundaries are documented.
- Each meaning has one canonical owner; references point to it instead of restating it.
Pruning Gate
Run the no-op test on the Skill and then on each sentence: if removing it would not change recurring agent behavior, delete it instead of polishing it. Reject or remove the Skill when project instructions, an existing Skill, or a native Host capability already produces the same behavior.
Split only when a branch has an independently useful trigger or later steps repeatedly cause premature completion. Do not split merely to shorten a file; progressive disclosure is cheaper than another model-routed description.
Evaluation Story
Before making a skill default-route material, capture at least:
- positive routing cases;
- negative routing cases;
- forbidden-load cases;
- one progressive-loading case for each spoke file that matters;
- one completion check showing the workflow reaches the intended output.
Use the active project's skill-quality rules when they exist.
