craft-critique — Claude Code skill: what it does, how to install
The craft-critique Claude Code skill reviews a rendered screen and can apply focused design fixes. It activates for critique or polish requests, before a screen is shown, or before a design goes to code.
MIT ★ 393 updated 2026-11-15
/plugin marketplace add diego3g/craft
/plugin install craft@craftWhat craft-critique does
The skill renders the screen, then checks hierarchy, spacing, stability, interaction, contrast, text, consistency, and motion. It reads DESIGN.md if present and checks the project’s design system, which takes precedence over default checklist rules. For rendering, it selects an adapter from references/adapters/: paper.md, figma.md, or html.md. If none fits the tool, it performs the same actions manually; with no design tool, it builds the screen in HTML.
In review mode, it ranks findings by impact and reports them without editing files. In refine and self-check modes, it fixes only the named issues, renders the screen again, and checks the changes. Fixes to shared components go to the craft-system process because they affect every screen that uses the component.
What it helps with
- spot hierarchy and contrast issues in a screen
- find why a screen feels visually crowded
- rank design findings by their impact
- apply identified fixes and check the screen again
Example prompts for the agent
- Review this screen and rank the problems by their impact.
- Polish this screen, fix only the issues you find, then check it again.
- Check this finished screen before I show it and list the changes it needs.
How to install craft-critique
1Plugin marketplace
/plugin marketplace add diego3g/craft/plugin install craft@craftThis installs the whole craft plugin with all its skills
2Skills CLI (Vercel)
npx skills add diego3g/craft --skill craft-critique -a claude-code3Manually: put the files in place
~/.claude/skills/craft-critique/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/craft-critique/
1Skills CLI (Vercel)
npx skills add diego3g/craft --skill craft-critique -a codex2Manually: put the files in place
.agents/skills/craft-critique/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add diego3g/craft --skill craft-critique -a cursor2Manually: put the files in place
.cursor/skills/craft-critique/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install diego3g/craft skills/craft-critique2Manually: put the files in place
.github/skills/craft-critique/SKILL.md1Skills CLI (Vercel)
npx skills add diego3g/craft --skill craft-critique -a gemini-cli2Platform CLI
gemini skills install https://github.com/diego3g/craft --path skills/craft-critiqueManually: put the files in place
.devin/skills/craft-critique/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about craft-critique
How do I install craft-critique in Claude Code?
Add the marketplace with /plugin marketplace add diego3g/craft, then install the skill with /plugin install craft@craft.
What does craft-critique do?
It renders a screen, checks it against design rules, and reports ranked findings. For improvement requests, it applies only the identified fixes and checks the result again.
How do review and refine modes differ?
Review mode reports findings and makes no edits. Refine mode applies the named fixes, then renders and checks the screen again.
Source text
SKILL.md · MIT
---
name: craft-critique
description: Use when asked to review, critique, refine or polish a design or screen, when something "feels off", "looks busy" or "isn't good yet", after finishing a screen and before showing it, or before handing a design to code.
---
# Design critique
## Overview
Critique the screen people will **see**, not the source you can read: render it first. Run the [rules](references/rules.md) as a checklist, rank what you find by how much it hurts the person using the screen, and **report before you change anything**, unless you were asked to refine. When you refine, fix only what the findings name, then look again.
## When to use
- "Review this", "what's wrong with this screen?", "refine", "polish", "it feels off".
- You just finished a screen and are about to show it.
- Before handing a design to code.
## When NOT to use
- A specific edit ("fix the typo", "make it blue"): make it. You may mention a serious problem you noticed in one line, without fixing it.
## Before you start
0. Read `DESIGN.md` at the project root if it exists: it is the project's design brief.
1. Read the project's design system. **Where it differs from a default in the rules, the project wins.** Don't flag it as a defect; note the difference once.
2. Pick your tool's adapter in `references/adapters/` (`paper.md`, `figma.md`, `html.md`). If no adapter matches your tool, do the same verbs in your tool and say which you mapped by hand. With no design tool at all, build in HTML (`html.md`). If an image is all you have, say that values read from it are estimates.
3. Decide the mode:
- **review:** the person asks about a design ("review", "critique", "what's wrong"). Report only.
- **refine:** the person asks to improve one ("refine", "polish", "fix it", "make it better").
- **self-check:** your own work that the person hasn't seen yet, including when another design skill calls this one. Fix what you find within the task's scope, render again, then show it with a short list of what the check changed.
## Steps
1. **Render it and look.** Take a screenshot at the size people use, and at a phone width if it's responsive. Look at it as the person using it would, and answer:
- What do I see first?
- Is that the most important thing?
- What is the one next action?
If you cannot render, say so at the top of your answer, and mark the findings that need a render to confirm.
2. **Run every rule group** in [the rules](references/rules.md): hierarchy, space, stability, interaction feel, contrast and colour, text, consistency, motion. Check each state the screen shows, not only the default.
3. **Rank the findings by impact** on the person:
- blocks or misleads;
- slows down;
- looks careless.
For each one: the problem, where it is, why it matters, and the fix.
4. **Review mode:** deliver the ranked report and **stop. Don't write a fixed version or edit any file** until the person picks the fixes.
**Refine and self-check modes:**
- apply the fixes, **limited to what the findings name**: no new layout, no new sections, no new files, no restructured files;
- **render again** with the same command and sizes as step 1 (if it worked then, it works now: try it before saying you can't);
- re-run the checks on what changed;
- report each fix and anything that is still open.
5. A fix that changes a **shared component** (a button used on twenty screens) is not a local fix. Name it, and hand it to the craft-system process, so every screen that uses the piece changes together.
## Rules (summary; detail in references/rules.md)
| Group | The core check |
|---|---|
| Hierarchy | One thing wins the first glance; at most one primary action; when in doubt, remove |
| Space | No dead or oversized areas, no cramped groups; repeated rows align in fixed-width slots |
| Stability | Nothing changes size when loading, swapping labels, selecting or toggling |
| Interaction feel | Hover is quieter than selected; focus ring only from the keyboard; pointer cursor only on links (on the web, follow the project's convention) |
| Contrast and colour | Small text readable on tinted and grey grounds; one strong colour moment |
| Text | No one-word last lines; titles capped in lines; plain words; says what happens next |
| Consistency | The same piece looks the same everywhere; banners only for whole-app problems |
| Motion | Only where it helps; opening slower than closing; respects reduced motion |
## Done when
- **Review:** a rendered look (or an explicit "couldn't render") and every rule group checked; a ranked report; nothing changed.
- **Refine / self-check:** the named fixes applied, re-rendered and re-checked, and the report says what changed and what's still open.
## Common mistakes
| Thought | Reality |
|---|---|
| "I can read the HTML/CSS, that's enough" | People see pixels. Wrapped titles, overlaps and shifts only show when rendered. Render, or say you couldn't. |
| "I'll include a fixed version so they see the result" | In review mode that's choosing for them. Report; fix what they pick. |
| "The loading button looks broken" | Name the class of problem: it **changes width**. Layout shift is the finding, and the fix is a fixed width with the label hidden under a spinner. |
| "The headline is long" | Check *how* it breaks: a one-word last line is a finding on its own. Balance the lines. |
| "While refining, I'll also add a max width, a phone layout and split the CSS" | Refine fixes the findings. Anything else is a new proposal: list it, don't do it. |
| "I couldn't render after the fixes" (when the first render worked) | Run the same command again. The re-check is the point of refining. |
| "Their hover equals their selected, that's a bug" | Check the project's guide first. If it says so on purpose, it's their rule. Note it, don't flag it. |
© Diego Fernandes · MIT · diego3g/craft