| name | skill-creator |
| description | Design or revise a focused Prometheus skill when the user explicitly asks to create, update, improve, review, or troubleshoot a skill. Route inferred reusable lessons through candidate submission and Brain Curator review; do not mutate skills after ordinary workflows. |
Skill Creator
Design small, precise skills without recreating the catalog feedback loop.
Authority boundary
There are two distinct paths:
- Direct user request: the user explicitly asks to create or edit a skill. Inspect overlap, propose the focused change, then perform the authorized mutation and validate it.
- Learned candidate: a workflow, Thought, Dream, cleanup, or normal chat suggests a reusable lesson. Submit a structured candidate with
skill_candidate_submit. Do not create or update skill files.
Brain Curator is the sole automatic writer. Candidate approval for a new skill authorizes design/proposal work only; it does not itself create the skill. Instruction, trigger, resource, and new-skill changes remain pending review unless the user directly requested the edit. Exact non-behavioral metadata repairs may use the Curator's safe path.
Evidence gate
Treat only user messages as evidence of user preference or approval. Assistant summaries, praise, completion claims, tool count, and repeated assistant wording are not evidence.
Submit a learned candidate only when supported by either:
- explicit user language such as “remember this” or “use this next time”; or
- repeated evidence across distinct sessions.
Record tool failures and successful validation as separate signals. Attach source session IDs and concise evidence; never infer confidence from the number of tool calls.
Design workflow
- Define one focused job and its observable output.
- Search the catalog for overlap. Prefer narrowing, merging into, or adding a resource to an existing skill over creating another broad entrypoint.
- Write a narrow description with clear positive and negative scope. Avoid generic single-word triggers.
- Keep
SKILL.md around 500–750 words where practical. Move schemas, background, detailed examples, and provider variants into stable canonical resources such as references/recovery.md or topic-specific files.
- Set broad role, style, and manually invoked skills to
implicitInvocation: false.
- Test positive and negative prompts before activating triggers. At most one high-confidence match should be mandatory; other matches remain suggestions.
- Run any bundled script or dependency against a disposable fixture. If it fails, record the failure and exact correction needed; do not claim the capability works.
- Validate frontmatter, manifest parity, resource paths, routing, and catalog health.
Candidate shape
For inferred improvements, call skill_candidate_submit with:
- candidate type;
- existing skill ID when applicable;
- concrete observed problem;
- proposed focused change;
- evidence/session references;
- confidence and trigger context.
Use create_new_skill_candidate only after overlap analysis shows no existing skill can absorb the job. Curator clusters equivalent candidates and suppresses rejected or duplicate suggestions.
Do not
- Create or update a skill after every successful workflow.
- Use assistant text as user approval or preference evidence.
- add dated resource filenames for recurring lessons;
- activate broad triggers without positive and negative routing tests;
- duplicate instructions across
SKILL.md and references;
- auto-apply behavioral changes merely because they look harmless;
- mark an untested dependency-backed skill ready.
