skill-portability

Audits whether an agent skill loads and behaves correctly across the major providers — Claude Code, Cursor, OpenAI Codex CLI, and the Agents-Skills Baseline (covering Gemini CLI, OpenCode, Pi). See scripts/providers.mjs for the canonical list. The audit covers three layers:

1. Format-level — does the file/dir structure and frontmatter match what

the provider expects?

1. Body-level — does the body reference provider-specific conventions

(XML tags, headings) that other providers ignore?

1. Tool-surface-level — does the body name tools (e.g.,

AskUserQuestion) that only exist in one provider?

The canonical provider list, required/forbidden frontmatter fields, and bundled doc snapshots live in scripts/providers.mjs and references/providers/. Provider docs are refreshed by hand on cadence (quarterly+) and committed — the audit never fetches at runtime, so per-provider verdicts are deterministic and offline-capable. See CONTRIBUTING.md for the refresh workflow.

Inputs

$ARGUMENTS — one of:

• A path to a skill directory (skills/code-reviewer/ or

skills/skill-creator/)

• A path to a single SKILL.md / .mdc / AGENTS.md file
• A glob (skills/*) — bulk audit
• A literal frontmatter+body block pasted into the prompt
• Empty — discover via ls skills/ and ask which to audit

Workflow

1. Resolve the input

Determine what the user is auditing.

| Input | Action | | ------------------ | -------------------------------------------------------- | | Directory path | Read the SKILL.md (or .mdc / AGENTS.md) inside | | Single file path | Read the file directly | | Glob | Expand and audit each match | | Pasted content | Treat the prompt body as the skill content | | Empty $ARGUMENTS | List skills/*/SKILL.md and ask the user which to audit |

Parse the frontmatter (YAML between leading --- markers) and remember the body content. Both feed into the per-provider checks.

2. Load the canonical provider list

Run the bundled ESM script — no shell required, works on macOS, Linux, and Windows:

bash
node skills/skill-portability/scripts/providers.mjs --pretty

Capture the JSON. Each entry has id, name, fileFormat, fileLocation, docUrls, localDocPath, requiredFrontmatter, optionalFrontmatter, ignoredFrontmatter, forbiddenFrontmatter, toolSurface, and notes.

This script is the single source of truth. Don't hardcode provider details into the audit body — re-run the script each time.

3. Fan out: one agent per provider, in parallel

Dispatch one Agent call per provider returned by step 2 (concurrent — one tool message with all tool uses). Don't hardcode the count — providers.mjs is the source of truth and may grow. Each agent gets:

• The provider entry from step 2 (id, format requirements, toolSurface, notes)
• The bundled provider doc snapshot read from localDocPath (e.g.

references/providers/cursor.md). Read the file with the Read tool before dispatching the subagent and pass its contents inline in the subagent prompt — do NOT instruct the subagent to fetch.

• The skill content from step 1 (frontmatter + body)
• These instructions to the subagent:

> Use the provider doc snapshot included in this prompt — it was bundled > with the skill at authoring time and is the authoritative reference > for this audit. Do NOT WebFetch any URL; if the snapshot is missing > details, surface that as a NOTES finding rather than fetching. > > Evaluate the skill against three layers: > > 1. Format: does it match fileFormat? Are all requiredFrontmatter > fields present? Are any forbiddenFrontmatter fields present? > 2. Body: does the body lean on conventions this provider doesn't > parse (e.g., XML tags Cursor strips, headings the baseline ignores)? > 3. Tool surface: does the body name tools not in this provider's > toolSurface? List each unmatched tool name. > > Return a structured verdict: > > > VERDICT: compatible | partial | incompatible > FORMAT: <one-line summary> > BODY: <one-line summary> > TOOLS: <comma-separated unmatched tool names, or "none"> > NOTES: <2–3 specific findings with file/line references where possible> >

Use subagent_type: "general-purpose". Run all dispatches in one message so they execute concurrently.

4. Aggregate the verdicts into a matrix

Build a markdown table with one row per provider and one column per layer:

| Provider                | Verdict      | Format              | Body          | Tools          |
| ----------------------- | ------------ | ------------------- | ------------- | -------------- |
| Claude Code             | compatible   | SKILL.md ✓          | clean         | -              |
| Cursor                  | partial      | needs .mdc          | XML stripped  | -              |
| OpenAI Codex CLI        | partial      | rename to AGENTS.md | clean         | Bash → shell   |
| Agents-Skills Baseline  | compatible   | SKILL.md ✓          | clean         | tool names vary |

Add a NOTES paragraph below summarizing the most actionable change to make the skill more portable.

5. Write COMPAT.md and print the matrix inline

Two outputs:

• Inline: print the matrix + notes in chat so the user sees it immediately.
• Persisted: write <skill-dir>/COMPAT.md (or ./COMPAT.md if no skill

dir), containing: the matrix, the per-provider verdict bodies, the timestamp, and the providers.mjs docUrls actually fetched. The user can diff this over time to spot regressions when providers move docs.

For bulk audits (glob input), produce one combined matrix where rows are skills, columns are providers, and write COMPAT.md at the repo root.

6. Handle missing or stale snapshots

If any provider's localDocPath doesn't exist or its snapshot is empty / truncated to a placeholder, surface this in the inline output:

⚠ Provider snapshot missing or sparse: `cursor` — references/providers/cursor.md
  is 503 bytes (likely SPA-rendered upstream). Verdict relies on the
  `notes` field. Refresh the snapshot by hand (see CONTRIBUTING.md).

Don't silently fall back to WebFetch — that's exactly the runtime fetch behavior the bundled-snapshot design avoids. If a snapshot is genuinely missing, treat it as unknown in the verdict rather than guessing.

To verify URLs still resolve (HEAD requests, no body fetch):

bash
node skills/skill-portability/scripts/providers.mjs --check

Examples

<example> <input>"is my code-reviewer skill cross-provider safe? it's at skills/code-reviewer/"</input> <output>

1. Resolve: read skills/code-reviewer/SKILL.md → frontmatter has name,

description, argument-hint, user-invocable. Body uses <example> blocks and references AskUserQuestion, gh pr diff, and a sibling scripts/detect-clis.mjs.

1. Load providers: node skills/skill-portability/scripts/providers.mjs --pretty

→ entries returned (currently 4).

1. Fan out one parallel Agent call per provider — Claude Code, Cursor, Codex CLI, Agents-Skills Baseline.
2. Aggregate verdicts:

| Provider | Verdict | Notes | | ---------------------- | ---------- | --------------------------------------------------------------- | | Claude Code | compatible | native format | | Cursor | partial | needs .cursor/rules/code-reviewer.mdc; references/ won't load | | OpenAI Codex CLI | partial | rename to AGENTS.md; AskUserQuestion not in tool surface | | Agents-Skills Baseline | compatible | loads from .agents/skills/; tool names vary per consumer |

1. Write skills/code-reviewer/COMPAT.md with full per-provider bodies.

</output> </example>

<example> <input>Pasted: a .mdc file with globs: and alwaysApply: in frontmatter — "will this load in claude code?"</input> <output>

1. Resolve: parse pasted content. Frontmatter: description, globs,

alwaysApply. No name field.

1. Load providers list.
2. Fan out 4 parallel agents.
3. Verdicts:

• Claude Code: incompatible — missing required name field; skills

CLI rejects without it. globs/alwaysApply are silently ignored.

• Cursor: compatible — this is native .mdc format.
• Codex CLI: partial — Codex reads AGENTS.md as plain markdown;

frontmatter is silently ignored, content would still apply.

• Continue.dev: partial — move to .continue/rules/<name>.md;

globs is supported, alwaysApply is not.

1. Recommend: add name: <kebab-case> to make portable; consider dropping

alwaysApply since only Cursor honors it. </output> </example>

<example> <good> Ran node scripts/providers.mjs --pretty once at step 2. For each provider, Read its localDocPath snapshot and inlined the contents into the subagent prompt. Subagents evaluated against bundled docs — no runtime WebFetch. </good>

<bad> Instructed subagents to WebFetch the provider docUrls at audit time. Result: slow audits, stale-doc false negatives, runtime dependency on every provider's docs site staying up. </bad>

The bad version reintroduces the runtime URL fetch the snapshot pattern exists to avoid. Snapshots are committed; refresh by hand on cadence (see CONTRIBUTING.md). </example>

References

• scripts/providers.mjs — canonical provider list,

format requirements, localDocPath for each provider's bundled snapshot. Run --check to verify upstream URLs still resolve.

• references/providers/ — per-provider doc

snapshots (committed; refresh by hand on cadence — see CONTRIBUTING.md).

• references/provider-formats.md —

per-provider deep dive: frontmatter shape, file location conventions, tool surface, common porting gotchas.

• references/audit-prompt.md — the verbatim

prompt template handed to each subagent in step 3.