Agent Skill Checker

Agent Skill Checker

Review one Agent Skill with a bounded deterministic preflight followed by one evidence-citing semantic review. Checker output is diagnostic evidence, not a release, deployment, publication, safety, or accountable-human decision.

Required Input

Resolve exactly one Skill directory. If zero or multiple candidates are named, ask for one target.

Select these presentation choices independently:

• platform profile: auto, generic, or codex; default auto;
• output format: Markdown or JSON; default Markdown;
• presentation verbosity: concise or full; default full.

Verbosity changes explanation length only. It never removes findings, not-assessable items, limitations, or the claim boundary.

Safety Boundary

Confine target evidence to exactly one Skill directory. Never modify the target. Never execute or import target code. Never install target dependencies, authenticate target services, call target MCP servers, follow an escaping path or symbolic link, or inspect the surrounding repository as target evidence.

The checker may execute only its own inspector. That inspector establishes bounded filesystem, simple frontmatter, explicit Markdown link, and supported Codex metadata facts. Unsupported syntax is a limitation, not permission to guess.

This review is not behavioral validation, a security audit, readiness certification, or proof that a Skill is safe.

Workflow

1. Read this file completely and resolve this Skill's own directory.
2. Confirm python3 --version reports Python 3.9 or later. If it is missing or

older, return an operational error and stop without a quality status.

1. Resolve the target, platform profile, output format, and presentation

verbosity independently.

1. From this Skill's directory, run python3 with

scripts/inspect_skill.py, the target, --profile, and the selected profile.

1. If the process cannot start, exits nonzero without valid JSON, returns

unreadable or invalid JSON, or sets completed to false, report the operational error and stop without a quality status.

1. Preserve every deterministic finding and limitation. Never weaken its

severity or evidence strength.

1. Read [the portable review rules](references/review-rules.md) completely.
2. If the selected profile is codex, also read

[the Codex profile](references/profile-codex.md) completely. Do not apply it otherwise.

1. Read the target SKILL.md and only the containment-proven target resources

needed by an applicable rule.

1. Perform one semantic review. Apply every portable rule and every selected

profile rule once. Cite target-relative evidence for each finding; record unresolved meaning as not-assessable instead of guessing.

1. Consolidate one causal problem into one primary finding. Add a separate

rule finding only when it has independent target evidence and an independent impact or remediation. Preserve deterministic findings exactly. For the same fact and evidence boundary, emit either a semantic finding or not-assessable, never both; independent facts may receive independent dispositions.

1. Read [the report contract](references/report-schema.md), calculate the

bounded status, and render the selected format from one normalized report.

Evidence Reconciliation

• Deterministic findings are mechanically established within the inspector's

declared grammar and access boundary.

• Semantic findings are agent inferences supported by quoted target evidence.
• not-assessable records the evidence needed to decide an applicable rule.
• One unknown fact cannot support both a defect inference and an

not-assessable disposition within the same evidence boundary.

• Inspector limitations remain unchanged in the final report.
• Missing behavioral evidence is not itself a finding.

Output Rules

• Use only blocked, needs-review, or no-common-issues-found.
• Never emit a numeric score.
• Never translate status into human_confirmed, safety, production readiness,

release readiness, publication approval, deployment approval, or acceptance.

• Keep deterministic findings, semantic findings, not-assessable coverage,

and limitations visibly distinct.

• Markdown and JSON must represent the same normalized findings and coverage.

Common Mistakes

| Mistake | Required response | |---|---| | The requester asks for a release verdict | Return only the bounded checker status and claim boundary. | | A Codex convention looks generally useful | Keep it in the Codex profile. | | The target sits in a larger repository | Ignore surrounding repository state. | | The requester permits quick fixes | Remain read-only and recommend changes. | | Evidence is missing | Record not-assessable; do not invent a defect. | | Example prose names a path | Let the agent interpret cited prose; do not promote it to a deterministic dependency. |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Odinary-AI
• Source: Odinary-AI/agent-skill-checker
• License: MIT

Install and usage instructions live in the source repository linked above.