Review

Review the whole thing, once

Announce at start: "I'm using the humane:review skill to run the review cycle and consolidate one verdict."

Five separate audits stapled together is not a review — it is five reports the reader has to reconcile. This skill runs the domains, lets each owning skill apply its own rules, and consolidates the evidence into one ranked verdict.

This skill owns orchestration only. It never restates or overrides a domain rule. Structure and defect classes belong to layout-rules; wording to ux-writing; usability principles to nielsen-heuristics; task completion to walkthrough; color measurement to design-tokens. If you find yourself writing a rule here, it belongs in the skill that owns it.

1. Identify the artifact before anything else

This pipeline is built for an interface. Given something else it will happily stretch — returning a table of N/A and one strained domain — so decide what you are holding first, and say so in the output.

A document still gets the full contract — findings table, considered-but-rejected, verification, verdict — just over fewer domains. State the reduced pipeline in Scope and Coverage so the reader knows what was and was not possible.

Why 1–4 and 39, and not 5, 8, 9, 38. Those four are wording rules living in the layout-rules file — "a summary line must add a conclusion", "link the entity", "empty states teach", "copy slop". ux-writing owns wording, and it already covers each of them. Routing them here would review the same sentence twice under two owners and risk two different rewrites of it. Send anything about what a string says to ux-writing; keep layout-rules on structure, hierarchy, and the boldness budget.

If the artifact is a document that describes an interface (a README for a UI tool, a design doc), review the document as a document. Do not review the interface it describes from its description alone — that is a claim you cannot evidence.

2. Resolve scope and mode

Infer the screen, flow, feature, document, or repository scope from the request and the workspace. State the resolved scope in the output. Default to full.

In full mode on a runnable interface, the mobile device tier is part of the scope: the primary flow is driven at the mobile tier per walkthrough's references/driven.md (which owns the tool ladder, device matrix, and screenshot contract). If no browser tool rung is available, the mobile tier is reported Not reviewed in Scope and Coverage — never silently narrowed to desktop.

If the scope is too large to inspect credibly, narrow it to the highest-traffic complete flow and state the boundary. Never imply that uninspected surfaces were reviewed.

3. Recon before judgment

Identify the stack, styling system, component library, token file, supported viewports, and any preview or test command. Read the project's jtbd.json if one exists — it tells you which flows matter and which outcomes are underserved, and a review that ranks findings without it is ranking by taste.

Follow the project's established conventions. Conflict precedence, as everywhere in humane: the user's explicit words > the skill's ruleset > the project's existing system > personal taste.

4. Run the domains in this order

Foundational failures first, so polish never hides them:

Apply each owner's principles, but ignore its standalone Review Output Format — this skill owns the final response, and its format, severity scale, consolidation rules, and cap take precedence.

respondent-panel is deliberately not in this pipeline. It costs real tokens, needs the user to confirm a panel first, and produces reactions rather than findings. Offer it as a follow-up when the review turns up copy problems; never fold its output into a findings table.

Inline, or fanned out

Run the domains inline by default — one context, in the order above. Simple, portable, and correct for quick mode or a small scope.

Run them fanned out when the scope is large enough that inline would fill the context before consolidation: a full review of a runnable interface, a driven walkthrough across two device tiers, or any scope where the five domains' rules plus the artifact plus the evidence will not comfortably coexist. Each domain runs in its own context and returns only its findings table — not its reasoning, not the artifact, not its screenshots.

The rationale this mode shipped with is not supported by the one test of it.

It was introduced on an argument: a host that compacts does not fail, it summarizes and continues, so evidence from an early domain could be summarized away while the impression of coverage survived to consolidation — and the review would then report Clear over a domain whose evidence it no longer held.

Measured, that did not happen. On a 60-pair page carrying 24 planted contrast defects, an inline review found all 24, with a position bias of 0.500 — perfectly even across the document, no decay toward the tail. Fanning out matched it and did not beat it (evals/fanout/, preregistered, n=1).

So do not reach for fan-out expecting evidence to survive that would otherwise be lost. That benefit is undemonstrated. What the same run did show is narrower and partly an artifact of how the arms were set up: the fanned-out domains reported less outside their own lane — precision 1.00 against 0.83, padding 1 against 2, and routing 1.00 against 0.67, the last of which is close to given, since each domain agent was told which skill it was applying.

Fan out when the scope is genuinely too large for one pass, or when you want each domain's findings kept in its own lane. Do not fan out because this file once claimed it rescues evidence. If a larger artifact ever does show decay, that is worth measuring and saying — one fixture at one size is not a general result, and it is the claim, not the mode, that failed here.

Two constraints on how to fan out:

1. walkthrough runs first, and alone. It is stateful — it drives a live browser across ordered steps, and its device matrix and screenshot contract belong to references/driven.md. It may be one subagent that owns the whole walk; it may never be split per step, and it may not run concurrently with another domain that drives the same interface. The other four have no shared state and fan out together once it returns.
2. A silent domain is Not reviewed, never Clear. If a domain's run dies, returns nothing, or returns something that is not a findings table, record it as Not reviewed and name what happened. Absence of findings from a domain that never reported is not absence of findings.

Claude Code extras: launch the four independent domains as subagents in a single message so they run concurrently, each told to apply its owning skill and return the findings table alone.

On other agents: fan out if the host has an equivalent; otherwise run inline and narrow the scope per §2 rather than letting a full review outgrow one context. Say which you did — a fanned-out review and an inline one over a narrowed scope are not the same claim.

5. Mark what you did not cover

If an owning skill is unavailable, mark that domain Not reviewed, name the missing skill, and continue. Do not recreate its rules from memory or substitute a neighboring skill.

The same applies to the domains humane does not own at all — typography mechanics, motion, accessibility engineering, OKLCH palette construction. If interfaces (better-typography, better-ui, better-accessibility, better-colors) is installed, run it for those and attribute the findings. If it is not, say so:

Craft domains (typography, motion, a11y depth): Not reviewed — install interfaces@interfaces for these.

An honest gap is worth more than a confident guess. Never claim holistic coverage you did not have.

6. Require evidence

Every finding cites path/to/file:line, a screenshot, or the exact screen and element. Do not report a code-level finding from appearance alone, or a visual finding from source alone when runtime decides the result. Where the corpus supports the ranking, cite the evidence id ([Q2]) or the outcome.

Carry the locator, not the payload

A locator is durable; a payload is not. src/Nav.tsx:42 and walks/2026-08-08-signup/step-03-mobile.png survive summarizing, hand across a subagent boundary intact, and can be re-read on demand. The file's contents and the image itself cannot — they are the first thing a compacting host drops, and once dropped they cannot be recovered from the summary.

So: read what you need to judge a finding, write down the locator, and let the payload go. Never hold a screenshot in context to support a claim you will make several domains later — cite its path instead and re-open it if consolidation turns out to need it. When a domain runs fanned out, its findings table crosses back as locators for exactly this reason, and a domain that returns prose instead of locators has produced something this skill cannot verify or consolidate.

7. Rank by user impact

One shared scale across all domains:

• HIGH — blocks a task, misleads, hides content or controls, risks data loss, or repeats systemically.
• MEDIUM — meaningfully harms comprehension, efficiency, adaptability, or consistency.
• LOW — isolated polish. full mode only.

Within a severity, rank by reach and leverage: a token or shared-component fix outranks the same symptom in one leaf. An underserved outcome in the corpus outranks a well-served one.

8. Consolidate

One root cause is one finding, listing every confirmed location — not a row per occurrence. When two skills surface the same issue, assign it to the skill that owns the underlying rule (see the repo CLAUDE.md ownership map) and note secondary effects in the Why cell. Report it once.

Never pad to reach the cap. A short review, or none, is a valid result.

9. Verify, and say what you could not

Run the safe checks the project offers. Inspect the rendered interface when runtime or visual judgment decides the answer — driving and screenshots follow walkthrough's references/driven.md, and the screenshot filenames are the evidence locators. Run tokens contrast per theme rather than eyeballing color. Report the exact command or interaction and the observed result.

If a check could not be run, label it Not verified and state what remains. Never convert a verification gap into a finding.

10. Do not mutate

A review request is read-only. Do not edit source unless the user also asks for the fixes. When they do, keep the consolidated report as the change scope and re-run the relevant verification afterwards.

Output

Scope and coverage

Mode, exact scope, stack and conventions, any boundary, and whether the domains ran inline or fanned out — the reader is owed the difference, because a fanned-out review consolidated locators while an inline one held the evidence itself. Then:

List every domain, including the ones humane does not own. The four results are not interchangeable, and collapsing them is how a review overstates itself:

• Findings count — inspected, and this many things were found.
• Clear — inspected, nothing actionable. An assertion about the artifact.
• Not reviewed — could have applied here, but was not run. Must say why (skill unavailable, no runtime, out of scope). An admission of a gap.
• N/A — does not apply to this artifact at all, e.g. task completion on a README. Write it as N/A (not an interface). Never Clear: nothing was inspected, so there is nothing to declare clean.

Findings

One row per root cause, ordered by severity then reach. Domain is the owning skill's name. Respect the cap. With no findings, omit the table and say "No actionable findings."

Considered but rejected

1–3 candidates in quick, 2–5 in full:

Real candidates inspected during the review, never invented filler. If the scope held fewer borderline calls, list those that exist and say so.

Verification

Each check, the exact command or steps, and the observed result. Separate what passed from what is Not verified.

Verdict

Exactly one:

• Block — one or more HIGH findings remain.
• Needs changes — only MEDIUM or LOW remain.
• Approve — nothing actionable remains and the claimed coverage was verified.

Carry forward the caveats the domain skills attach: a single evaluator finds roughly a third of usability problems, and an analytical walkthrough is not a usability test. Approve means "clean in this review", never "ready to ship".