| name | agent-skill-authoring |
| description | Creating, refining, splitting, consolidating, renaming, or auditing an agent skill — drafting a `SKILL.md`, editing frontmatter, tightening a `description`, deciding where a new rule belongs, or running the structure validators. Triggers on "add a skill", "split this skill", "audit skills", "recast this skill as a capability", and any change to a `SKILL.md` or its `references/`. The authoring rules for the agentskills.io format — capability framing, discovery metadata that survives a host's listing truncation, section anatomy, progressive disclosure, cross-references, and three bundled validators, one per kind of edit. |
| user-invocable | false |
Agent Skill Authoring
Use this capability whenever you create, refine, split, consolidate, rename, or audit an agent skill under the host project's skill root. It is what turns a durable convention into a well-formed, discoverable skill and keeps the skill tree coherent as it grows.
Skills authored here follow the agentskills.io format. For the host project's active skill inventory and topic-to-skill routing, defer to each skill's own description discovery metadata and the directory listing under the skill root; where a host also maintains a written index (e.g. AGENTS.md), keep it in sync too.
Guidelines:
- MUST run the bundled validator that owns what a change touched —
scripts/check-skill-frontmatter.mjs after editing frontmatter, scripts/check-skill-body.mjs after editing prose, scripts/check-skill-references.mjs after adding, moving, or renaming a reference file — and all three when unsure; together they are the enforcement path for the frontmatter, naming, discovery length-cap, and reference-linkage rules this skill states nowhere else (see audit-checklist.md).
- SHOULD propose or implement a skill update when any task exposes a reusable convention, outdated guidance, a recurring review issue, or a missing project rule — skill maintenance happens when work reveals durable learning, not after every narrow fix.
- SHOULD skip skill maintenance when the work produced no generalizable learning, and state that it was skipped.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.
Scoping and MECE
See scoping-and-mece.md for:
- choosing a coherent skill boundary, skill name, split, consolidation, or source-of-truth location
- checking overlap with neighboring skills before adding new guidance
- classifying every section as mechanism or judgment when a vendor- or runner-specific skill sits beneath a tool-agnostic owner
- using section length and topic growth as signals for restructuring
Capability Framing
See capability-framing.md for:
- framing a skill as an ability the agent gains rather than a document it reads
- naming the activity a skill enables, and the document-style suffixes to avoid
- the voice of the
description opening clause, the H1, and the opening paragraph
- recasting an existing guideline-style skill in a fixed order, without changing what it requires
Frontmatter and Naming
See frontmatter-and-naming.md for:
- creating or editing discovery-critical
SKILL.md frontmatter
- setting the invocation-control fields (
argument-hint, arguments, user-invocable, disable-model-invocation) by skill archetype — guideline skill vs workflow entry point
- choosing the skill directory name and keeping it aligned with the
name field
- porting or preserving host-project harness fields
Description Writing
See description-writing.md for:
- drafting, trimming, or auditing the
description field against its byte cap
- ordering a
description so the routing decision survives a host's listing truncation
- adding likely user phrasings and symptom-based triggers without over-broadening the skill
Body Content Style
See body-content-style.md for:
- writing or revising substantive skill-body or reference-file sections
- balancing concise topic explanation, examples, and guideline bullets
- citing the upstream documentation URL in a section that pins a version or mirrors a vendor's option surface
- placing normative RFC-2119 requirement bullets in detailed reference content rather than parent routing sections
Progressive Disclosure
A rule is load-bearing when an agent that loads SKILL.md and opens no reference would produce wrong output for want of it — held before the work starts, not looked up once the reader already knows the question exists. That test still sorts a skill's material, but its consequence is a conditional read obligation rather than a relocation: a load-bearing rule's own statement stays in its reference, and SKILL.md carries an RFC-2119 obligation to read that reference before the work its rule governs. A reference nobody is told to read never gets read; a SKILL.md that states every rule directly cannot be tree-shaken by the sessions that will never touch most of them. This test is itself load-bearing for a skill's author, so it is stated here as a rule rather than left behind the pointer below.
See progressive-disclosure.md for:
- deciding when a skill should stay single-file or split into
references/
- the load-bearing test's full sorting table, and how it decides whether a reference earns a conditional read obligation
- the size thresholds that signal a skill or reference file has grown too large
- using the parent routing-section format:
## Topic, See [file.md](./references/file.md) for:, descriptive situation bullets, then a **Guidelines:** block carrying the read obligation
- wording a read obligation's triggering condition narrowly enough to be skippable
- stating the fact a routing bullet points at — the flag, limit, or rule by name — instead of announcing that one exists
- keeping parent routing bullets free of RFC-2119-style requirement keywords so they remain routing cues, not duplicated rules
- the one case where a rule's own statement stays in
SKILL.md instead of moving to its reference
Guidelines:
- MUST treat a rule as load-bearing when an agent that loads
SKILL.md alone would produce wrong output for want of it — a fixed order, a closed set, or a constraint whose violation is not self-evident from the output — and as elaboration otherwise.
- MUST place a load-bearing rule's own statement, its RFC-2119 bullets, and everything that elaborates it in the reference file that governs it, not in
SKILL.md.
- MUST give
SKILL.md a **Guidelines:** block, placed after a reference's routing list, carrying one RFC-2119 bullet per reference that names the reference and states the condition — narrow enough to be skippable — under which it MUST be read.
- MUST NOT restate a load-bearing rule's statement or its RFC-2119 bullets in
SKILL.md once its reference states them, except under the carve-out below.
- MUST keep a rule's own statement in
SKILL.md, never moved to a reference, when the rule's triggering condition is unconditional within its own skill's scope — a pointer that would fire on every turn costs a read and shakes nothing.
Cross-Referencing and Discovery
See cross-referencing.md for:
- adding, renaming, moving, deleting, or linking skills and reference files
- choosing one source of truth instead of copying detailed rules across skills
- using topic-based cross-skill references, verifying intra-skill relative links, and keeping skill discovery current (plus any written index a host maintains)
Project Skill Archetypes
See project-skill-archetypes.md for:
- creating the project-specific skills a scaffolding pass calls for: structure, component, and UI/design
- the three-way ownership triangle and each archetype's skeleton, topics checklist, or table patterns
- growing archetype skeletons with worked examples and mechanical boundary checks
Auditing and Validation
See audit-checklist.md for:
- auditing multiple skills or reporting skill-tree quality
- running the three bundled structure validators —
check-skill-frontmatter.mjs, check-skill-body.mjs, and check-skill-references.mjs — and which kind of edit each one answers for
- running the bundled link-freshness audit (
scripts/link-freshness/check.mjs) on a schedule to catch a cited vendor URL that has gone 404, why only a confirmed-dead link fails it, and why a pull_request trigger on it is a request-forgery primitive
- checking inventory, skill discovery, section anatomy, RFC-2119 bullets, topic-based cross-skill references, and relative links
- identifying overlap, stale assumptions, orphan references, and missing source-of-truth links
