- Home
- Skills
- UX & UI Design
- Design QA: Designer Judge
Design QA: Designer Judge
A senior-designer rubric for your agent, with 8 anchored criteria, 3 isolated runs, median scores and a stability check.
$9
Works with the AI Agents you already use
Design QA: Designer Judge
Example session with this skill installed
Judge my mobile pricing screen (cand@390x844.png) against the reference (ref@390x844.png). Give me scores and the top fixes.
- Read your context and instructions
- Compiled the design qa
Mode
absolute. Ran 3 isolated judges (separate sub-agents) on the same two PNGs, with no metrics or builder notes shown. Saved judge.r1.json, judge.r2.json, judge.r3.json, then:
python3 S/scripts/judge_merge.py runs/20261114-1520-1/judge.r1.json runs/20261114-1520-1/judge.r2.json runs/20261114-1520-1/judge.r3.json --md
Judge - pricing @ 390x844 (absolute, 3 runs)
Overall 3.71 / 5, worst spread 1
| criterion | score | spread | evidence |
|---|---|---|---|
| hierarchy | 4 | 0 | hero: price is the clear focus; secondary link competes slightly |
| typography | 3 | 1 | plan cards: 5 different sizes, title wraps to 2 lines |
| color | 4 | 0 | accent used only on primary CTA, matches reference |
| grid_rhythm | 3 | 1 | card gaps 24px vs 16px section rhythm |
| states_interactive | n/a | - | no state screenshots |
| a11y_visual | 4 | 1 | grey price note looks low-contrast |
| reference_fidelity | 4 | 0 | same structure and order as the reference |
| generic_ai_look | 4 | 0 | no gradient blobs or emoji bullets |
Top fixes
- pricing cards: reduce to 3 type sizes (20/16/14) so plan titles fit on one line
- card list: use a 16px gap to match the section rhythm
- price note: darken the grey text (check exact ratio with axe)
No criterion has spread > 1, so the medians are stable. Files: judge.json (merged, pass it to verdict.py --judge) and judge.md in .design-qa/pricing/runs/20261114-1520-1/.
Connects securely to your tools. The creator never sees your data.
About this skill
Asking a model "is this design good?" gives vague and unstable answers. This skill gives the agent an anchored rubric with 8 criteria: hierarchy, typography, colour, grid and rhythm, interactive states, visual accessibility, fidelity to the reference, and a "generic AI look" detector. Every score must name the screen region it is based on.
Three modes
- Absolute: 1-5 scores per criterion
- Pairwise: A-vs-B preference (more reliable than comparing scores)
- Defects: a defect list with severities
Bias-resistant protocol
The judge runs three times in isolation, with the order swapped in one pairwise run to avoid position bias. Metrics, verdicts and the builder's reasoning are hidden from the judge to avoid anchoring. A bundled script merges the runs into median scores with a per-criterion spread, so you can see when the judge is unsure.
Output is JSON plus a short Markdown summary with the top fixes. An optional brand profile file makes the judge score your typography and colours against your brand instead of penalising them. The output plugs straight into the Design QA verdict (gate G7).
What you get
SKILL.md: an 8-step judging workflow, including what to hide from the judgereferences/rubric.md: 8 criteria with anchors for scores 2 and 5, null rules and generic-AI flagsreferences/judge-prompt.md: ready prompt, brand-mode line, JSON schemas for all 3 modes, merge rulesscripts/judge_merge.py: median/spread, majority voting with label mapping, defect union (standard library only)templates/brand-profile.example.md: brand profile template (type scale, colour tokens, rules)
Permissions / network
- No network calls. Reads and writes local JSON and Markdown only
- Writes judge files into the folder you choose (normally
.design-qa/<screen>/runs/<run>/)
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for UX & UI Design.
LLM judges vary from run to run; the 3-run median and spread make this visible but do not remove it. Contrast is judged visually, use axe (Fidelity Check skill) for exact ratios. 3 model calls per screen, viewport and mode.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
- Software
Python 3.8+Multimodal AI model with image input
Permissions
File Scopes
Terminal only runs the bundled merge script (Python standard library). Reads the screenshots you pass and writes judge JSON/Markdown into the folder you choose. No network calls, no env vars. Note: 3 model calls per screen, viewport and mode.