Critique Skill · 5 维度专家评审
Produce a single-file HTML "design review report" that scores any artifact across 5 dimensions and proposes actionable fixes. Inspired by the huashu-design expert-critique flow.
When to use
- After the agent (or user) generates an artifact (deck / prototype / landing page) and the user asks "what's wrong with this?" or "review this"
- As a self-check loop the agent can run on its own output before emitting it
- For comparing two variants of the same design
What you produce
A single self-contained <artifact type="text/html"> review report
including:
- Header — what artifact was reviewed, date, reviewer ("OD · Critique skill"), 1-line verdict
- Radar chart (inline SVG, no library) showing the 5 scores
- Five dimension cards, each with:
- Score 0–10 (with band: 0–4 Broken · 5–6 Functional · 7–8 Strong · 9–10 Exceptional)
- 1-paragraph evidence (cite specific elements / files / lines)
- One Keep / Fix / Quick-win bullet
- Combined action lists at the bottom:
- Keep — what's working, don't touch
- Fix — P0 / P1 issues that are visually expensive
- Quick wins — 5–15 minute tweaks with disproportionate impact
The 5 dimensions
Each dimension is independent — a deck can be 9/10 on Innovation but 4/10 on Hierarchy and the report should say so plainly. Don't average away interesting failures.
1. Philosophy consistency · 哲学一致性
Does the artifact pick a clear direction and stick to it through every micro-decision (chrome / kicker / spacing / accent)?
Evidence to look for:
- Is there one declared design direction (e.g. Monocle / WIRED / Kinfolk) or is it three styles in a trench coat?
- Does the chrome / kicker vocabulary stay in one register, or does page 3 say "Vol.04 · Spring" and page 7 say "BUT WAIT 🔥"?
- Are accent / serif / mono used by the same rule throughout?
0–4 Three styles fighting each other. 5–6 One direction but half the elements drift. 7–8 Coherent, occasional drift on edge pages. 9–10 Every element argues for the same thesis.
2. Visual hierarchy · 视觉层级
Can a stranger figure out what to read first, second, third — without being told?
Evidence to look for:
- Is the largest type clearly the most important thing on each page?
- Do mono / serif / sans roles match the information's role (meta / body / display)?
- Lots of "loud" elements competing? Or a clear primary + secondary + tertiary tier?
0–4 Everything shouts. 5–6 Hierarchy works on hero pages but breaks on body. 7–8 Clear tiers, occasional collision. 9–10 Eye moves with zero friction.
3. Detail execution · 细节执行
The 90/10 stuff — alignment, leading, kerning at large sizes, image framing, foot/chrome polish, edge-case spacing.
Evidence to look for:
- Big-stat pages: does the number sit on a baseline, or float?
- Left/right column tops aligned in
grid-2-7-5? frame-img+ caption proportions consistent across pages?- Mono labels: same letter-spacing? same uppercase rule?
- Any orphaned
<br>causing 1-character lines?
0–4 Visible tape and string. 5–6 Most pages clean, 1–2 ragged. 7–8 Polished, expert eye finds 2–3 misses. 9–10 Magazine-grade — the kind of detail that makes printed-by-hand typographers nod.
4. Functionality · 功能性
Does the artifact work for its intended use? Click targets, nav, readability at presentation distance, copy-paste-ability for code blocks, mobile fallback if relevant.
Evidence to look for:
- Deck: keyboard / wheel / touch nav all working? Iframe scroll fallback?
- Landing: CTA above the fold? Phone number tappable on mobile?
- Runbook: code blocks copyable, mono font, no smart quotes?
- Critical info readable from 4m away (large screen presentation)?
0–4 Visually fine but doesn't accomplish its job. 5–6 Core flow works, edge cases broken. 7–8 Robust through normal use. 9–10 Defensively engineered — handles iframe / fullscreen / paste / print without flinching.
5. Innovation · 创新性
Does this push past the median? Is there one element that makes people lean in?
Evidence to look for:
- One unexpected layout / motion / typographic move that wasn't required?
- Or 100% safe — could be any deck/landing from any agency?
- Is the innovation earned (matches direction) or grafted on (random WebGL on a Kinfolk slow-living deck)?
0–4 Generic AI-slop median. 5–6 Competent and unmemorable. 7–8 One memorable moment, the rest solid. 9–10 Multiple moves you'd steal — but each one obviously serves the thesis.
Scoring discipline (read before you score)
- Always cite evidence — "scored 4 because hero page mixes Playfair display with Inter sans on the same line" beats "feels inconsistent". Numbers without evidence get rejected.
- Don't average up — if Hierarchy is 5 because page 3 is broken, don't bump to 7 because pages 1 and 2 are fine. The score is the worst sustained band.
- Don't grade-inflate — a 7 means strong, not acceptable. If every score is 7+, you're not reviewing critically.
- Innovation is allowed to be low — 5/10 is fine for production deliverables. Don't punish appropriate conservatism.
Workflow
Step 1 — Acquire the artifact
Three modes:
- Project file — user said "review the index.html I just made": open it from the project folder.
- Pasted HTML — user pasted code in the chat: read it from the message.
- Generated by you in this turn — you just emitted an artifact
above and want to self-critique: re-read your own
<artifact>.
If multiple HTML files exist, ask which one (don't review all).
Step 2 — Read enough to score
Skim the entire <style>, then read 6–8 representative content
blocks. Do not score from frontmatter alone. The score depends on
executed design, not declared intent.
Step 3 — Score with evidence
For each of the 5 dimensions, write the score and a 30–80 word evidence paragraph that names specific elements. Use line numbers, class names, page numbers.
Example:
Dimension: Detail execution Score: 6 / 10 Evidence: Stat-cards on page 3 align cleanly (grid-6, 3×2), but on page 8 the right column foot sits 2vh higher than the left because .callout has 3vh top margin while the figure doesn't. Image captions use mono on page 5 but sans on page 7 — pick one.
Step 4 — Build the action lists
Aggregate the 5 evidence paragraphs into:
- Keep (3–5 bullets) — concrete things working that the user must not break in the next iteration. Cite by class / page / element.
- Fix (3–6 bullets) — must-do, ordered by visual cost saved per minute spent. Each bullet ≤ 1 sentence.
- Quick wins (3–5 bullets) — 5–15 minutes each, high
signal-to-noise (e.g. "swap
display:flexforgridon page 4 to fix the column drift").
Step 5 — Emit the report HTML
Build a single file:
- Header: artifact name + reviewer credit + date
- Big radar chart (SVG)
- 5 dimension cards in a 1-column or 2-column grid
- Three action lists at the bottom with checkbox affordance
Use the active DESIGN.md tokens if one exists; otherwise default to a neutral light theme (off-white background, near-black text, one accent for radar fill).
Output contract
<artifact identifier="critique-<artifact-slug>" type="text/html" title="Critique · <Artifact Title>"> <!doctype html> <html>...</html> </artifact>
One sentence before the artifact ("Reviewed X across 5 dimensions, see
report below.") and stop after </artifact> — do not paraphrase
the report in chat; the user will read the artifact.
Hard rules
- 5 scores, every time — partial reports (e.g. only 3 dimensions) are not allowed.
- Evidence per score — no "feels off" / "needs work". If you can't cite an element, the score is not justified.
- Don't grade-inflate — overall mean above 8 is suspicious; check yourself.
- Don't review your own artifact in the same turn — the user needs to see it first. Self-critique only on explicit request ("now critique what you just made").
- Single-file HTML only — no external CSS/JS. Inline everything.
- Radar chart is mandatory — gives the report a recognizable silhouette and lets the user spot weak axes at a glance.
