AI will happily tell you every screen needs “more whitespace” and “clearer CTAs.” That is not critique. Critique for freelance UI/UX work is evidence-tied, severity-ranked, and scoped to what you can fix in this engagement.
This guide shows how to prompt for useful critique, what to paste in, how to triage output for clients, and a set of copy-paste prompts you can keep in Notion.
What good critique output looks like
For each issue you want:
- Observation — what is on the screen / in the flow
- Why it hurts — task success, comprehension, trust, a11y, or consistency
- Severity — High / Med / Low relative to the user goal
- Fix direction — not a full redesign; a lever (hierarchy, copy, grouping, state, pattern)
- Out of scope flag — if fixing it requires brand strategy, eng architecture, or new research
If the model skips severity or invents user research you did not provide, the prompt failed.
Golden rule: paste constraints or get fiction
Always include:
- Primary user and job-to-be-done
- Platform (web/mobile) and breakpoints you are shipping
- What is frozen (brand colors, nav IA, tech limits)
- What phase you are in (flows vs UI polish vs handoff)
- How many days/hours you have for this pass
Never ask “fully redesign this app.” Ask for a punch list.
Prompt pattern (reuse forever)
You are assisting a freelance product designer on a scoped critique pass.
User goal: [goal]
Constraints: [brand / platform / time]
In scope: [screens or flow]
Out of scope: [list]
Critique the UI/flow below. For each issue:
- severity (High/Med/Low)
- evidence
- fix direction (not a full redesign)
Do not invent analytics or quotes.
Call out missing states (empty/error/loading/success) if relevant.
Material:
[paste structure, descriptions, or copy]
Screenshots help multimodal models; for text-only models, describe layout top-to-bottom and list interactive elements.
Five critique prompts worth keeping
1. Heuristic punch list
Ask for modern product-UI heuristics (clarity, feedback, error prevention, consistency, recognition). Cap at 10 issues. Require severity.
2. Flow drop-off audit
Provide ordered steps. Ask where users hesitate, what is missing between steps, and which edge cases are must-fix before launch.
3. Hierarchy & density
Describe type sizes, grouping, and competing CTAs. Ask what to demote, promote, or group — without proposing a new brand.
4. Accessibility smoke critique
Feed color notes, form patterns, and touch targets. Ask for a prioritized a11y punch list suitable for a freelance fix pass (not a full audit report for lawyers).
5. Stakeholder feedback triage
Paste raw Figma/Slack comments. Ask the model to cluster: must-fix / nice / out of scope / unclear — and draft a decision-order reply for the decision owner.
Turning model output into client-facing critique
AI drafts; you edit.
- Delete anything you cannot defend from the actual UI.
- Merge duplicates.
- Re-rank severity using their success metric from the SOW.
- Rewrite in your voice — short, calm, specific.
- Attach issues to frames in Figma; do not send a wall of chatbot text.
For tense clients, use AI to draft the email, then strip apologies and buzzwords. Your reputation is tone plus judgment, not prompt volume.
Critique rituals that keep freelancers sane
Async round: You leave numbered comments. Decision owner replies in 72 hours with Accept / Reject / Defer. You implement Accepts once.
Live round (45 min): You present top 5 High issues only. Park Low. End with decisions written in Meeting Notes.
Red team pass (solo): Prompt the model to act as a rushed user and a confused first-time user. Keep the failures that match real constraints.
Log each critique in Notion: Type = Critique, Decisions filled, Actions assigned. Without a decision log, AI critique becomes content theater.
Failure modes (and fixes)
| Failure | Fix |
|---|---|
| Generic “improve contrast” | Paste actual hex/roles; ask for which pairs fail |
| Model redesigns whole IA | State IA is frozen; ask for local fixes |
| 40 low-value nits | Cap issues at 8–12; require severity |
| Fake user quotes | Ban invented research in the prompt |
| Scope creep via “while we’re at it” | Out-of-scope section mandatory |
Example: from slop to useful
Slop: “The dashboard feels cluttered and could use better visual hierarchy.”
Useful: “High — Primary action ‘Create project’ competes with 4 equally weighted metrics cards above the fold. Fix direction: one primary button, metrics as secondary row, default empty state with a single next step when data is zero.”
Teach the model that bar by giving one example in the prompt (“Match this specificity”).
Ethics and honesty
Do not paste confidential client screens into public model endpoints if your agreement forbids it. Use redacted descriptions, local models, or enterprise privacy settings as appropriate. Never fabricate usability findings for a pitch.
AI critique is a force multiplier for structured looking, not a substitute for sitting with the user goal.
Where prompts live in your OS
Keep a Notion page AI Prompts with numbered entries. Tag by moment: Discovery, Critique, Handoff, Difficult reply, Proposal. After a win, note “used on [Project].”
A fuller set (50 prompts across those moments) ships in kits like Clientrail, alongside proposal/SOW templates and a Notion client OS. Whether you collect 10 or 50, the standard is the same: constraints in, severity out, you remain the designer of record.
Quick start this week
- Pick one live flow.
- Run the flow drop-off prompt.
- Keep only High/Med you agree with.
- File a Critique meeting note with decisions.
- Implement in one pass; do not reopen Low nits until Handoff QA.
That loop beats generating another “10 AI tools for designers” bookmark you will never open.
Want the full Client OS?
Clientrail packs Notion structure, 50 prompts, and deal templates for freelance UI/UX — no fake testimonials, no income claims.