review-doc: the fresh-process document reviewer

The operating contract is blunt: dispatch a separate fresh-process reviewer, never the author; it rubber-stamps. This skill is that reviewer, for documents. It exists because the failure mode of agent-written docs is not absence, it is plausible slop: a spec with every heading present and every section quietly fabricated. A reviewer that only checks shape approves slop. This one checks shape, altitude, and truth.

Note

Never review a document you authored in this same process. If you drafted it (or a skill you ran drafted it), hand this SKILL.md and the document to a fresh process with no conversational context beyond the pointers in Step 0. An author reviewing its own work confirms its own assumptions; that is the rubber stamp this skill exists to prevent.

Step 0: Identify the doc class, then ground yourself

Read the frontmatter key first; it tells you what you are reviewing and which checklist applies:

Frontmatter key	Doc class	Template to check against
job:	Job Spec (durable)	templates/job-spec.md
serves: (on an RFC shape)	Ship-coupled RFC	templates/rfc.md
backs:	Design record backing an invariant	templates/rfc.md
North Star / outcomes / job index	Product Spec	templates/product-spec.md
Situation / Pillars / Deferrals	Strategy	templates/strategy.md
date: + commit:	Decision record	templates/decision-record.md

Then load the grounding set: the template for the class, the anchors (vision, principles, invariants), the product spec (for outcome slugs and the job index), and any document this one points at (serves:, backs:, a cited Job Spec or decision record). You cannot judge an up-anchor you have not read.

Step 1: Run the four checks

Work through all four for every document. Record each finding as it surfaces, with file:line.

1. Template fidelity

Every section of the class template is present, in order, and answered, not occupied. A heading with placeholder prose under it is a finding. Depth must match the stakes (a Quick Win answers in a line; a high-stakes doc expands), but the gates never disappear: a Job Spec with an empty ## Prove it, a missing silent failure in Good / bad, or no Abandon signal fails regardless of stakes. An RFC with fewer than two evidenced struggling moments, or fewer than the four named user-failure modes, fails. A strategy with untagged Situation claims or an empty Deferrals table fails.

2. Altitude

Each layer has a working altitude, and leakage is the most common defect:

• Job Spec: no feature names, flags, files, vendors, or technology nouns; no numbers, dates, account names, or ARR (those live in Job Links and the RFC). If any line would not survive a full rewrite of the feature, it is implementation and it leaks.
• RFC: the Solution Space describes an envelope (MUST / MUST NOT / free to vary), never a blueprint. Prescribed components, layouts, or data models are findings.
• Product Spec: outcome altitude; runtime and architecture belong in docs/.
• Strategy: funds jobs, does not restate them; pillars link to Job Specs rather than re-describing them.

3. Groundedness

The anti-slop check. For every number, quote, evidence link, and named source in the document, ask: where did this come from?

• Evidence links must resolve to something real. A link you cannot verify is flagged unverifiable, not assumed fine.
• Claims must carry their honest label: a Strategy claim is evidenced, hypothesis, or to-validate, and a stated Confidence score must match the evidence actually cited (Confidence 3 with one anecdote is a finding).
• A specific number with no source is presumed fabricated until traced. Fabrication is a BLOCK, not a REQUEST_CHANGES: a wrong document that reads as authoritative is worse than no document, because the whole fleet anchors to it.
• Honest gaps pass. *(coverage gap: no runnable scenario yet)* or a to-validate tag is the system working. The defect is confident invention, never admitted uncertainty.

4. Up-anchor integrity

The hierarchy only works if the pointers hold:

• serves: on a Job Spec names an outcome slug that exists in the Product Spec; serves: on an RFC names a real Job Spec; backs: names a real invariant.
• Named invariants exist in the invariants anchor, verbatim.
• The job index and the Job Specs agree (every indexed job exists; every job ladders to exactly one outcome).
• A change to a Job Spec’s job / outcome / stakes lines, or a material STRATEGY.md change, cites its decision record. Those lines are diff-protected by the operating contract; a silent rewrite is a BLOCK. (Authoring-time review of a new doc skips this; it applies to changes. See amend-durable-doc.)

Step 2: Return the verdict

One of three, using the operating contract’s vocabulary:

• APPROVE — every check passes. Say so plainly; do not invent findings to look thorough. A reviewer that always requests changes is as useless as one that always approves.
• REQUEST_CHANGES — fixable defects. Every finding cites file:line, names which of the four checks it fails, and says what passing looks like.
• BLOCK — fabricated evidence, a crossed invariant, a silent rewrite of diff-protected lines, or a doc pointing at anchors that do not exist. These are not polish items; the document must not be treated as ratified in any form until resolved.

Order findings most-severe first. Iterate with the author to APPROVE; do not let anything downstream (an RFC citing the spec, a build citing the RFC) proceed on an unratified document.

Output

Verdict: APPROVE | REQUEST_CHANGES | BLOCK

Findings (most severe first):
1. [BLOCK] <file>:<line> -- <check failed> -- <what is wrong> -- <what passing looks like>
2. [CHANGE] <file>:<line> -- ...

Checked: template fidelity ✓/✗ · altitude ✓/✗ · groundedness ✓/✗ · up-anchor ✓/✗
Grounding set read: <the template, anchors, and pointed-at docs you actually read>

The Grounding set read line is not decoration: a reviewer that did not read the up-anchors could not have checked them, and saying what you read keeps you honest about what you verified versus assumed.

The self-teaching check (dogfood this skill)

Seed a known-good document with five defects (a fabricated evidence link, a missing silent failure, an implementation noun in a Job Spec, a serves: slug that resolves nowhere, an untagged strategy claim) and hand a fresh agent only this skill and the grounding set. It passes if it finds all five with citations, and separately APPROVES the unseeded original without inventing false blockers. Both halves matter: recall without precision is noise.

Related

• AGENTS.md — the operating contract and reviewer checklist this skill implements for documents
• create-job-spec — authors the Job Specs this skill gates
• create-rfc — authors the RFCs this skill gates
• amend-durable-doc — the edit-time guard for diff-protected lines
• uat-ux-debug — reviews shipped behaviour against the docs; this skill reviews the docs themselves