| name | skill-creator |
| description | Create, revise, review, validate, and benchmark Prometheus skills and bundles, including SKILL.md instructions, resources, manifests, tool contracts, recovery paths, verification, and structured prompt routing. Use for explicit requests to make, update, improve, recreate, repair, or test a skill or its triggers. |
Skill Creator
Use this skill only after an explicit user request to create, change, review, or test a skill. The outcome is a focused, discoverable, executable skill whose instructions, metadata, resources, routing, and verification agree.
Authority and evidence
- A direct user request authorizes the requested skill edit. Inspect overlap, make the smallest coherent change, and validate it.
- A reusable lesson noticed during ordinary work is not authorization. Identify the owning skill and ask through
ask_prometheus_questions; submit skill_candidate_submit only when the user asks to capture the lesson. Use skill_curator for review or governed application.
- Treat user messages as evidence of preferences and approval. Do not infer approval from assistant summaries, successful tool calls, or repeated wording.
- Decline skills that would enable unauthorized access, deception, malware, exfiltration, or other surprising behavior.
Creation workflow
- Capture the contract. Extract the job, inputs, observable outcome, output/artifact, dependencies, platform limits, examples, success criteria, and likely positive and negative trigger prompts. Ask only for missing decisions.
- Search overlap. Call
skill_list with a task-shaped query, then skill_read for the strongest candidate and skill_inspect when metadata, health, ownership, or resources matter. Prefer narrowing an existing skill, adding a resource, or merging work over creating a broad duplicate.
- Design the playbook. State the outcome, preflight gates, exact tool/category contract, numbered primary path, evidence and success condition for every step, known recovery branches, final verification, delivery, and cleanup. Use imperative instructions and explain why a fragile step matters.
- Choose the package. Use
skill_create for a genuinely self-contained one-file skill. Use skill_create_bundle when schemas, examples, scripts, templates, or references are needed; write them with skill_resource_write. Give SKILL.md as much space as the workflow needs; use one-level-deep linked resources when selective loading, reuse, or maintenance benefits from separation. Keep frontmatter to name and description, and do not add README, changelog, or setup-guide clutter.
- Implement with the native contract. For an existing skill use
skill_update_from_source for an upstream refresh, skill_update_metadata for a metadata overlay, or skill_manifest_write for a reviewed manifest patch. Use skill_resource_list/skill_resource_read before changing resources and skill_resource_delete only when removal is authorized. Preserve upstream source and record local overlays.
- Author routing as policy. Keep legacy
triggers to high-confidence multiword phrases (maximum 12). In skill.json, use promptSignals with phrases, alternative allOf groups, anyOf, vetoing noneOf, and minScore. A phrase scores 4, each term in a matched allOf group scores 2, each anyOf term scores 1, and any noneOf match excludes; matching requires score >= minScore. Use implicitInvocation: false for broad, role, style, or manual skills. Test positive, negative, and near-miss prompts before widening routing.
- Validate and iterate. Call
skill_read, skill_inspect, and skill_audit_all; verify frontmatter, manifest parity, resource paths, limits, ownership, health, and the final artifact. For metadata or routing changes, provide triggerPositivePrompts and triggerNegativePrompts (or the prompt-signal aliases) to the mutation tool. For skill_repair_metadata, preview first and apply only with reviewed repairs and confirm:true. Use the evaluation loop for baseline/with-skill comparisons, objective assertions, qualitative review, timing, tokens, and variance; never claim an untested dependency-backed path is ready.
Tool contract
The normal authoring surface is skill_list → skill_read/skill_inspect → skill_create or skill_create_bundle → skill_resource_write → skill_manifest_write/skill_update_metadata → skill_audit_all. Use skill_import_bundle, skill_export_bundle, or skill_update_from_source only when the source boundary is explicit. Keep skill_candidate_submit, skill_curator, and skill_repair_metadata inside their evidence and confirmation gates.
Completion
Report the skill ID and exact changed resources, the routing tests and validation gates run, any known limitations or unverified paths, and whether the change is a direct user-authorized edit or a Curator candidate. Leave no disposable eval artifacts in the skill directory.
