ux-audit-panel
Skill4 starsChanged 3 months agoMITView on GitHub
Run a heuristic UX audit with a panel of independent evaluator subagents and fill out a heuristic audit spreadsheet. Each evaluator scores the same 49 statements from a different designer perspective without seeing the others' work, then their scores are averaged into the template and their disagreements written up as a divergence report. Use when the user asks for a UX audit, a heuristic audit, an audit spreadsheet, a multi-perspective design review, or says "audit this site/app/flow".
In this repositoryOne other file it gives its agents.
- design-critiqueSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 8 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Ask these one at a time, not as a block.
- Distinctive: line 29
Create the run folder, and copy this skill's two reference files into it so the run is self-contained and the evaluator prompts carry no …
- Very specific: line 58
Write <run>/project.json with the project details from step 1, using the keys app_name, flow, goal, link, notes.
- Distinctive: line 87
Never tell an evaluator how many others there are, who they are, or what they found.
- Distinctive: line 88
Never pass one evaluator's output into another's prompt.
- Distinctive: line 89
Never run them sequentially with the earlier results in context.
- Distinctive: line 90
If an evaluator fails, relaunch it with the same prompt.
- Very specific: line 126
Never edit assets/template.xlsx.
- Distinctive: line 127
Never write scores yourself.
- Distinctive: line 128
If fewer than 2 evaluator files exist, stop and say so.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
UX audit panel
Fills the audit template using several evaluators that never see each other's work. The averages go in the spreadsheet. The disagreements are the other half of the deliverable.
Runs entirely on Claude Code subagents under the user's subscription. Never call the Anthropic API, never use an API key, never install or invoke an SDK.
Step 1 — Ask
Ask these one at a time, not as a block. Skip any the user already answered.
- What to audit: one URL, several URLs that form a flow, or a folder of screenshots.
- How many evaluators. Default 3, maximum 5, minimum 2.
- Where to save the output.
- Project details for the "About the project" sheet: app name, flow being audited, goal of the audit, link, any notes.
Then pick perspectives from reference/perspectives.md. Defaults are A1, A2, A3 for
three evaluators. Never assign the same perspective twice. Tell the user which
perspectives are in the room and let them swap any before you start.
Create the run folder, and copy this skill's two reference files into it so the run is self-contained and the evaluator prompts carry no machine-specific paths:
<chosen folder>/<app-slug>-audit-<month>-<year>/
evidence/ raw/ reference/mkdir -p <run>/evidence <run>/raw <run>/reference
cp <skill>/reference/statements.md <skill>/reference/rubric.md <run>/reference/<skill> is this skill's own directory. <run> is the folder you just created. Use absolute
paths for both. Every command below is written relative to <skill>.
Step 2 — Capture
For URLs:
scripts/capture.sh <run>/evidence <url> [<url> ...]This captures desktop fold, desktop full page, mobile fold, mobile full page, the
accessibility snapshot, and console output for each screen, in an isolated browser
session. If the user gave a screenshots folder, copy its contents into
<run>/evidence instead and skip this step.
Write <run>/project.json with the project details from step 1, using the keys
app_name, flow, goal, link, notes.
Step 3 — Evaluate
Launch every evaluator in a single message so they run in parallel, using the
Agent tool with subagent_type: "general-purpose".
Each evaluator gets this prompt, with its own perspective and output path:
You are conducting a heuristic UX audit as a single evaluator.
Your perspective: [full text of the chosen perspective from perspectives.md]
Read
<run>/reference/statements.mdfor the 49 statements and<run>/reference/rubric.mdfor how to score and what to output.The evidence is in
<run>/evidence/. Read every screenshot with the Read tool and every snapshot and console file.urls.tsvmaps screen numbers to URLs.You are the only evaluator on this audit. Do not read, list, or look for any other file in
<run>/raw/. Do not look for other agents' output anywhere. Your judgement must come only from the evidence and your own perspective.Score all 49 statements and write your JSON to
<run>/raw/evaluator-N.json. Write nothing else to disk.
Isolation rules you must hold to:
- Never tell an evaluator how many others there are, who they are, or what they found.
- Never pass one evaluator's output into another's prompt.
- Never run them sequentially with the earlier results in context.
- If an evaluator fails, relaunch it with the same prompt. Do not patch its scores yourself and do not let another evaluator cover for it.
Step 4 — Merge and write
python3 scripts/merge.py <run>/raw <run> <run>/project.json
python3 scripts/write_audit.py <run>/results.json <run>/<app-slug>-audit-<month>-<year>.xlsxmerge.py builds results.json, keeps every evaluator's note in its own voice in the
Notes column, and writes divergence-report.md listing every statement where the
evaluators split by 2 points or more, worst spread first.
Then add a short synthesis at the top of divergence-report.md, under a
## What this means heading, in at most 15 lines: the sharpest disagreements and why
they happened, and the issues that multiple evaluators independently flagged. Agreement
is what to fix. Disagreement is what to take to real users.
Step 5 — Report back
Tell the user the global score, the three weakest sheets, how many of the 49 statements the panel split on, and the paths to both files. Do not paste the whole report.
How the spreadsheet is filled
Evaluators 1 and 2 go in columns C and D. Evaluators 3, 4 and 5 go in J, K and L, to the right of Notes, so nothing marked "don't move" moves. Only the E column formula changes, widening to average across however many evaluators ran. Columns F, G, H and the Summary sheet are never touched.
A statement scored "N/A" by every evaluator has its E, G and H cells cleared so the
row drops out of the sheet grade instead of poisoning it. This is correct and intended.
Constraints
- Never edit
assets/template.xlsx. It is the source template. - Never write scores yourself. Every number in the spreadsheet comes from an evaluator.
- If fewer than 2 evaluator files exist, stop and say so. Do not produce a partial audit.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

