Auditing skills

An audit grades a skill against Goodeye's authoring checks: the same checks every public template displays. Run it before you publish to see what a reader will see, or any time after to improve a skill you already shipped.

The checks describe how a skill is written, not the outputs it produces. They are not a guarantee of safety, authorship, accuracy, or fitness for any task. A reader still has to decide whether a template fits their need.

The authoring checks

Every public template shows three named checks: Safety, Runnable, and Well-formed. Together they give a reader an honest, at-a-glance signal of how well a template's skill is authored.

Safety is the safety screen that runs before a template is published. See Templates for the safety states (clean, flagged, and not safety-checked) a reader sees and what they mean.
Runnable means the skill has clear, ordered instructions and a defined input and output contract an agent can follow.
Well-formed means the skill is discoverable, coherent, and held to a stated standard: a clear trigger description, consistent terminology, a definition of what a correct output looks like in checkable terms, and at least one real, specific verifier rather than a holistic "is this good overall?" check.

Check states

The two earned checks (Runnable and Well-formed) each read as one of:

Met: the check was assessed and every part of it passed.
Not met: the check was assessed and at least one part fell short.
Not checked: the check was not assessed.

Safety is a separate badge with its own status (clean, flagged, or not checked); see Templates. Every public template page shows all three checks, including the ones a skill has not earned.

How the checks are computed

The checks are computed once, when you publish a template version, and pinned to that version. Publishing a new version recomputes them; older versions keep the result they were published with.

Authoring checks never block publishing. Only the Safety check can stop a publish; a template that earns no authoring checks still publishes. Publishing reports the full breakdown so you can see exactly what to improve, including non-blocking authoring notes (a missing tag set, an overlong body, a referenced verifier with no calibration) that never change a badge.

Running an audit

An audit is an opinionated health check against these checks. It returns a prompt pack your agent runs locally, inspects the skill body and every directing sibling file, and produces a priority-ranked report: each finding is tagged P0 (blocker), P1 (recommended), or P2 (polish), quotes the text to change, and states one concrete fix. Requires at least view access.

You choose which findings to fix; the audit applies only what you approve and never auto-saves, editing a local copy first when one exists. You can also audit a skill file on disk that is not saved to Goodeye yet, in which case the report recommends saving it as a hosted skill.

CLI: goodeye skills audit <id-or-name> (omit the id to audit a skill file on disk)
MCP tool: audit_skill
REST: POST /v1/skills/{id_or_slug}/audit, or GET /v1/audit/skill-prompt for a skill file on disk

The audit is the place to improve a skill before or after you publish it.

Checking safety on demand

You can run the platform safety checks without publishing, against a hosted skill or a published template version. The call makes no changes and runs the safety checks (billed as verifier runs), returning a status of clean, flagged, blocked, or error. What safety means, the states, and the banner a reader sees are canonical in Templates; publishing computes the authoritative status over the whole bundle and stores it on the version.

Hosted skill: CLI goodeye skills check-safety <id-or-name>, MCP check_skill_safety, REST POST /v1/skills/{id_or_slug}/safety-check.
Published template version (re-checkable by anyone; pass --anonymous to run without an API key, billed against a small per-caller credit budget): CLI goodeye templates check-safety <identifier>, MCP check_template_safety, REST POST /v1/templates/{template_ref}/safety-check.