Build a four-lens adversarial critique panel — skeptical user, accessibility, "why will this fail," engineering feasibility — and run a feature concept through it. Every AI finding gets an accept/reject decision with rationale, because the panel below lands real hits, misses seeded flaws, and asserts at least one confident falsehood.
Jae is a product designer at Fernwood, a fictional grocery-delivery app, two days from a concept review for "Pantry Refill" — a feature that predicts what you are running low on and pre-fills a weekly basket. The team likes it too much, which is exactly the danger. Jae's plan is an AI red-team: four adversarial personas that attack the concept before the review does. But an AI critique panel has two failure modes of its own — it produces plausible-sounding objections that are wrong, and it misses flaws a specialist would catch instantly. So the graded skill here is curation: you accept, reject, or escalate every finding, in writing, and you sweep for what the panel never mentioned.
Paste-ready. Read it once as a fan, once as an assassin.
FEATURE CONCEPT — Pantry Refill (Fernwood grocery app)
Every Monday 7am, Pantry Refill builds a suggested basket from
your purchase history and predicted run-out dates. The basket
is auto-scheduled for delivery Thursday unless the user edits
or cancels it by Wednesday 6pm. First-time users are enrolled
during checkout via a pre-ticked "Smart refills (recommended)"
checkbox.
Basket screen: predicted items in a grid. Each card shows the
item photo, price, and a confidence ring around the photo —
green ring = high confidence, amber ring = medium, red ring =
low. Ring color is the only confidence indicator. Tapping a
card removes the item; a small "why this?" link opens the
prediction rationale.
Predictions: v1 uses purchase-interval averages per item.
Households sharing an account are treated as one buyer. Items
bought for one-off events (parties, guests) are not
distinguished from staples.
Edits: removing an item this week removes it from all future
baskets (to reduce repeat friction). There is no undo for
this. Cancelling three weeks in a row disables the feature
silently.
Success metric: % of auto-scheduled baskets delivered without
user edits. Target: 60% by Q3.
Do your own two-minute attack first. Before any AI, list the three biggest problems you see. The panel will bury you in findings shortly; this list keeps your judgment ahead of the volume.
Convene the panel, one lens per chat. Run four separate prompts against the pasted brief. Skeptical user: "You are a busy parent who distrusts subscriptions and auto-charges. Attack this concept: list every moment you would feel tricked, surprised, or charged without meaning to be." Accessibility: "You are an accessibility specialist. Audit this concept against WCAG 2.2 AA. Cite the specific success criterion for each issue." Why-will-this-fail: "You are a post-mortem author writing in 2027 about why this feature flopped. Name the failure chain." Engineering feasibility: "You are a staff engineer. What is naive, expensive, or fragile in this spec as written?"
Merge into a findings register. One row per distinct finding: ID | Lens | Finding (quoted) | Severity guess. Deduplicate across lenses — several will spot the same flaw in different words; note the convergence, it usually marks the real problems.
Verify before you judge. Any finding that cites a rule, standard, number, or platform behavior gets checked against a primary source — the actual WCAG success-criterion text, the actual platform guideline. AI critics cite confidently and are wrong often enough that an unchecked citation is worthless. Log each check's result. At least one citation in your panel's accessibility output is likely to be wrong or non-existent — the answer key covers the common fabrications.
Rule on every finding. Add columns: Decision (accept / reject / escalate) | Rationale (one line). Accept means it changes the design. Reject means you can argue why it is wrong or negligible — write the argument, not "disagree." Escalate means it needs data or a specialist. A panel run where everything is accepted means you obeyed the AI; everything rejected means you wasted it.
Sweep for what the panel missed. Recheck the brief yourself against three seeded-flaw families: consent and dark patterns (read the enrollment sentence again), accessibility of the confidence display (read how confidence is communicated), and destructive actions (read what happens when you remove an item). For each miss, add a row marked PANEL MISSED.
Fix the concept. Rewrite the brief's flawed passages — enrollment, confidence display, removal behavior, and anything else you accepted — as v2 text. Keep v1 and v2 side by side.
Write the critique summary. Half a page: findings counted by lens, accepted/rejected/escalated tallies, the panel's best catch, the panel's worst miss, one falsehood you rejected with proof. This is what you would bring to the concept review alongside v2.
Run each lens in a separate chat — a single chat asked for four perspectives converges into one polite reviewer with four hats, and the attacks lose their teeth.
A red-team panel multiplies your critical eyes but not your judgment — the decision log, where every AI attack is accepted with consequences or rejected with proof, is what turns adversarial noise into a sharper concept.