| name | skill-creator |
| description | Author or refine workspace-local skills (SKILL.md) so future turns load concise, agent-oriented procedures. |
Skill creator (isanagent workspace skills)
Use this when you need a durable procedure (checklist, tool order, project conventions) that should apply across sessions without pasting the same instructions every time.
Where to write
- Path pattern:
workspace/skills/<skill-name>/SKILL.md relative to the onboarded workspace root (same tree the load_skill_instructions tool indexes).
- One skill = one directory = one
SKILL.md. Avoid scattering extra markdown unless it is truly shared assets.
Frontmatter (required)
Every SKILL.md must start with YAML:
name: short identifier, usually matches the directory.
description: one line, optimized for skill search—describe when another model should load this using concrete triggers (tools, file types, error symptoms), not marketing language.
Optional keys (requires, always, …) follow the same conventions as built-in templates—omit unknown keys.
Body: write for another model
- Use imperative steps (“Call … then …”), not documentation addressed to operators.
- Name isanagent tools exactly as they appear in your tool list (
execution_run, glob_files, …).
- Keep skills short and skimmable; link or cite repo docs for long policy (sandbox rules, harness config).
- Put execution-relevant guardrails here (ordering, caps, anti-patterns). Do not duplicate entire
AGENTS.md—reference it for global behavior.
- Prefer decision tables or bullet checklists over prose essays.
Editing flow
- Read the existing
SKILL.md if present; minimally extend rather than rewriting unrelated sections.
- Save; no compile step—skills are picked up on next load/search.
- If the skill is niche, set
always: false (default) so it only loads when search matches.