Standards
August 18, 2026 · View on GitHub
This page states which specifications this repository follows, where it deliberately goes beyond them, and where the machine-readable source of truth lives. It links the canonical documents instead of restating them — when this page and a linked document disagree, the linked document wins.
The floor: open specs we follow
Two upstream specifications define baseline validity for every skill in this repo:
- agentskills.io/specification — the open
Agent Skills standard that Claude Code follows. Requires only
nameanddescription; documentslicense,compatibility,metadata, andallowed-toolsas optional. - code.claude.com/docs/en/skills — Anthropic's Claude Code skills reference, the primary source for frontmatter field semantics. (API surface: platform.claude.com Agent Skills overview.)
Both are intentionally permissive. Everything valid under this repo's rules is also valid under both upstream specs — the overlay below is additive, never subtractive.
The overlay: where we deliberately diverge
The Intent Solutions marketplace tier is intentionally stricter than the upstream floor. The divergences, all documented in the schema changelog's NON-NEGOTIABLES and the master spec:
- 8 required frontmatter fields, not 2:
name,description,allowed-tools,version,author,license,compatibility,tags(theALWAYS_REQUIREDset). Upstream treats most of these as optional; the marketplace requires them for trust signals, downstream pinning, and legal clarity. - Missing required fields are ERRORS at marketplace tier, not warnings. The gate
is
scripts/validate-skills-schema.py, run in CI byvalidate-plugins.yml. capabilitiesas a frontmatter field is banned (a## Capabilitiesheading is fine and is credited as an## Overviewequivalent).- Heading-equivalence fairness for external contributors — standard-but-different
section names (
## Usagefor## Instructions,## Troubleshootingfor## Error Handling, etc.) are credited, per the mirror-by-default decision record.
Changes to the required-field set, the tier model, or error-vs-warning semantics are architectural and require explicit approval before landing — see the schema changelog's "How we got here" section for why that rule exists.
The direction: kernel single source of truth
The validator's hand-rolled field sets are converging on a machine-readable kernel:
the authoring/v1 schema family in
@intentsolutions/core
(pinned exactly in package.json). The kernel's
schemas/authoring/v1/CHANGELOG.md
is canonical for authoring-contract semantics; this repo's schema changelog cites it
rather than duplicating rationale.
The conformance signal for that cutover is
kernel-shadow-validation.yml: it
runs the kernel's published JSON Schema over the same SKILL.md corpus the prose-spec
validator grades and reports the per-file AGREE/DISAGREE deviation rate. It is
advisory by design — report-only, not in the required-status set, and it never
blocks a merge. Promotion to a blocking gate is a separate, later cutover step.
The catalog: two files, one source of truth
| File | Role | Hand-edit? |
|---|---|---|
.claude-plugin/marketplace.extended.json | Source of truth for every catalog entry | Yes |
.claude-plugin/marketplace.json | CLI-compatible catalog, auto-generated | Never |
pnpm run sync-marketplace regenerates marketplace.json, missing
plugins/**/package.json files, and the README AUTO-TOC block. The pre-commit hook
runs it automatically when the extended catalog is staged, and CI fails if any
derived file drifts. Authoring workflow: CONTRIBUTING.md.
External plugins: mirror by default
Externally-synced plugins (registered in sources.yaml, a small minority of the
catalog) follow mirror by default · upstream improvements · never clobber, per
the decision record:
- The contributor's own repo is the source of truth; the weekly sync mirrors it and opens a PR a human reviews. We do not locally edit a pure mirror.
- Quality improvements are upstreamed to the contributor's repo, not forked into
a divergent local copy. A locally-hardened source is frozen with
curated: trueuntil the improvement lands upstream. - Mirrored paths self-register their lint exclusions: the managed
sync-lint-ignoresblock in.markdownlint-cli2.jsoncis generated fromsources.yaml(viascripts/sync-lint-ignores.mjs), because upstream markdown style is the upstream maintainer's choice, not ours.
Canonical documents
A document in 000-docs/ may declare itself AUTHORITATIVE (or CANONICAL) only if it is
linked from this table. That rule is what the table exists for: it is the single public index
of which document owns which fact class, and it is the reason a document cannot grant itself
authority by editing its own header.
The one-owner-per-fact-class authority map is
blueprint 727 § 11 (with its § 11.1
activation proof). This table is the public index that reaches every owner in ≤ 1 hop; § 11 is
the per-fact-class assignment. A fact class gains an owner only by a row here plus a § 11 map
row — never by a header edit, and never by restating the fact in another document. Governing
files (CLAUDE.md, AGENTS.md, this file) point at owners rather than duplicating their
content; where they do describe a fact operationally, the description names its authority and is
pinned by the doc-governance assertions (validate:doc-fact-assertions).
| Topic | Document |
|---|---|
| Platform master standard (plugins, CI, release, docs governance, canonical contract, adapters, root README landing contract) | 000-docs/727-AT-ARCH-master-modernization-blueprint.md |
| Reference-architecture evidence base (12 pinned primary sources) | 000-docs/728-RA-DATA-reference-architecture-benchmark.md |
| Reference-architecture decision (adopt / modify / reject) | 000-docs/729-AT-ADEC-reference-architecture-synthesis.md |
| Required fields, tier model, history | 000-docs/SCHEMA_CHANGELOG.md |
| Full skills standard (master spec) | 000-docs/6767-b-SPEC-DR-STND-claude-skills-standard.md |
| External-sync policy | 000-docs/694-AT-DECR-external-sync-mirror-by-default-model.md |
| Skill submission intake standard | 000-docs/700-DR-GUID-skill-submission-standard.md |
| External-PR review standard | 000-docs/709-DR-GUID-reviewing-external-prs.md |
| Document filing standard (numbering, category codes, 000-docs layout) | 000-docs/000-DR-STND-document-filing-system.md |
| Cross-repo source-of-truth map (which system owns which fact class) | 000-docs/718-AT-ARCH-source-of-truth-map.md |
| Contribution requirements | CONTRIBUTING.md |
| Security policy | SECURITY.md |
The blueprint (727) is the platform master standard and supersedes
000-docs/6767-h-SPEC-DR-STND-claude-code-extensions-master.md, whose section-level disposition
it records. It does not supersede the master spec (6767-b) above, which remains the owner of
the skill-frontmatter rubric. The blueprint's own authority-graph proof — one owner per fact
class, with this table as the public index — is its § 11.1.