Back to Lobehub

UX Audit

.agents/skills/ux-audit/SKILL.md

2.2.1019.1 KB
Original Source

UX Audit

A repeatable, standards-based UX review of one surface at a time. The benchmark is two things together:

  1. Jenifer Tidwell, Designing Interfaces — the pattern language for what a good interface is made of. See references/pattern-catalog.md.
  2. The ux skill — LobeHub's execution checklists for how a flow should behave.

The audit answers: which patterns does the surface use (and how well), and where is the experience weak (each gap tied to a checklist item). Recurring gaps feed back as new ux checklist items; the audit itself becomes a worked-example reference.

Do one surface per run — a full-app sweep is too much for a single pass. Re-run per page as the product grows; that's the "continuous" part.

Three layers — pick by what you need to catch

An audit is not one activity. A finding is only trustworthy from a layer that can actually see it. Each layer has its own procedure file; run the ones the surface needs.

LayerFileWhat it doesCatchesCost
L1 Staticlayer-1-static.mdRead the codeMissing states/branches (empty/error/retry), no draft persist, absent patterns, structural issuescheap, offline, every audit
L2 Visuallayer-2-visual.mdScreenshots of the rendered surfaceReal visual hierarchy & dominant control, spacing/contrast/alignment, truncation/overflow, how empty/loading/error actually look, responsive breakpoints, dark/lightmedium; needs a render
L3 Dynamiclayer-3-dynamic.mdDrive the real user journey via agent-testing + instrumentIn-progress/locked states, forced error/empty states, does step N lead to N+1, focus/keyboard, quantified CLS / LCP / INP / long-taskshigh; needs a running env + auth

Coverage matrix — which layer can conclude what

The core rule: a verdict must come from a layer that can see it. Don't tick a visual or runtime verdict off the code.

Finding typeL1L2L3
Missing empty/error branch, no retry, draft not persisted, absent pattern
Real visual hierarchy / is the dominant control the primary action❌ misleads
Spacing / alignment / contrast / truncation / overflow / dark mode
Off-screen selection; what empty/loading/error actually render as
Responsive breakpoints (narrow / mobile)
In-progress / locked states; forced error / empty; capability-gated
Journey stitching (forward momentum across steps)weakweak
Focus order / keyboard reachability
CLS / LCP / INP / long-task numbersqualitative only
Which of two variants is better (A/B winner)❌ misleads❌ misleads✅ (+analytics)

⚠️ The recurring trap this prevents: ticking "one primary button" or "empty is a real page" from a variant prop in the code. Those are L2 verdicts — confirm them on the render, never from L1 alone.

Tiering — don't run all three every time

  • L1 always — fast, complete-coverage baseline for every surface.
  • Add L2 when the findings are about layout, hierarchy, rendered states, or responsive.
  • Add L3 when you need to walk a journey, force states L1/L2 can't reach, or measure performance (CLS etc.).

--l1 / --l2 / --l3 scopes a run to one layer; default is L1 (+ L2 if screenshots are supplied).

Ground rule: evidence, not vibes

Every finding cites its evidence — file:line (L1), a screenshot you verified with the Read tool (L2), or a captured value / snapshot (L3). Before asserting a load-bearing claim, confirm it in the layer that owns it; a wrong "it's missing" is worse than no finding.

Ground rule: benchmark the surface class, not just our own artifact

Reading our code can only surface flaws in what we built — it is structurally blind to a capability we never built at all, because an entirely-absent affordance leaves no file:line, no dead branch, no half-wired button to grep for. The checklists guard the quality of the states that exist; they do not tell you which states a surface of this class is expected to have.

So before (or alongside) reading code, name the surface's class and its domain conventions: how do the mature, comparable products build this exact screen, and what do they offer that a first version forgets? An OAuth consent screen's class norms, for example (GitHub / Google / Okta): show which identity you're authorizing as and let the user switch account / re-authenticate, name the requesting app, list the scopes, allow deny, and point to later revocation. A file picker, a checkout, a share dialog each carry their own class norms. Write this expected-capability list first, then audit gaps against it — otherwise the audit only ever polishes the paths that already exist and silently blesses a missing one.

❌ The first pass of the OAuth audit measured consent against our internal state checklists only and reported button-hierarchy / retry gaps, while missing the biggest one: the consent screen locks the user into the current identity with no switch-account path (OAuthConsent/Login.tsx) — a class norm every comparable OAuth provider ships. A competitor-norms pass catches this on minute one; a code-only pass never can.

Ground rule: comparing two variants — the winner is an outcome verdict, not a craft verdict

When an audit compares two variants of the same surface ("is Agent or Classic onboarding better?"), the trap is judging which is better made (more polished, more patterns, more AI) when the real question is which better gets the user to their goal. For a gateway / interstitial surface — onboarding, consent, paywall, a loading gate — the two diverge hard: the best version is often the least version, because the surface stands between the user and what they came for. Craft is not outcome, and the richer artifact is routinely the worse one.

So a variant comparison must:

  1. Name the success metric first, then judge against it. Onboarding = completion rate + time-to-value + drop-off, not pattern richness. A checkout = conversion. Write the metric before scoring, or you'll default to scoring craft.
  2. Gate the winner verdict on L3 / analytics. "Which variant is better" is a behavioral outcome — it lives in the coverage matrix's L3 row alongside CLS/INP. From L1/L2 you may compare mechanics ("A's error recovery is more complete", "B is fewer steps"); you may not declare a winner. No funnel data → say "insufficient evidence, here's what I'd need", and stop. A confident winner call with a buried "needs L3" caveat is the failure — the caveat does not license the verdict.
  3. Cost to the user is a first-class axis, inverse-weighted for gateway surfaces. Time, steps, tokens, latency. Past a threshold, richness is a liability on any surface the user wants to get through, not into.
  4. Anchor 意义感 on the user's real goal, not the feature's richness. A flow that detains the user in itself when their goal is elsewhere is less Meaningful even if more engaging — read correctly, 意义感 and 自然 favor the fast path there. Don't mistake "more conversational /more AI /more crafted"for"more meaningful".
  5. Read the org's revealed preference as evidence. Feature flags, which variant is the fallback, which the most-constrained platform is forced onto (desktop), recent reverts. When the universal fallback is variant B, B is the trusted baseline and the burden of proof is on A — don't explain these signals away as "ceiling vs floor".
  6. Pick the right reference class, and weight by the real intent distribution. An AI tool's first-run benchmarks against ChatGPT / Claude / Cursor (near-zero onboarding, straight to the box), not SaaS setup wizards (Notion / Linear) that reward thorough onboarding. Score the modal user (who wants to skip), not the ideal engaged one.

❌ This skill's own miss: an L1 read judged Agent onboarding "better" than Classic because it was more polished /conversational (richer completion panel, name suggestions, view transitions), citing 意义感 ≳ 自然 > 确定性. In production Agent's effective-guidance completion was not high, users found it too slow (they wanted the tool, not a chat), and the org rolled back to Classic. Every error above was present: craft mistaken for outcome, 意义感 scored backwards, cost footnoted, a winner declared from L1 on what is an L3/analytics metric, and the flag-gated /degrades-to-Classic/desktop-excluded signals explained away.

Ground rule: a "redundant" control is a composition question before a subtraction one

When L1 spots two controls that seem to do the same thing, the reflex is subtraction — delete one, hide one, merge them, differentiate the copy. Resist it. Two controls that share an intent often differ in scope, and the honest fix expresses that scope difference in layout — promote the wider-scope one to a visible, titled sibling — not by removing it. Subtraction is a behavior-layer move; the better answer usually lives in the composition layer, which is precisely the half of the benchmark an audit drops when it walks the ux checklists but never opens references/pattern-catalog.md. The checklists speak states, momentum, and draft-safety; they carry no vocabulary for Titled Sections / Grid of Equals / Center Stage, so a checklist-only read reaches for "dedupe" every time. Walk the pattern-catalog pass (L1 step 2) before writing a remedy for any "redundant / overlapping" finding, and ask: same intent, different scope? → the fix is a sibling, not a delete.

❌ The CC AskUserQuestion audit flagged the per-question "write your own" box and the global "Or type directly" escape as redundant and proposed hiding / merging them — a subtraction. They aren't redundant: the per-question box is question-scoped, the escape is whole-form-scoped. The composition-layer answer is to hide the escape when there's one question and render it as a visible peer to the question block when there are several — a Titled Sections move a checklist-only read never surfaced. Skipping the catalog cost the better answer.

Ground rule: report the good, not only the gaps

An audit that only lists what's broken has drifted into a bug report. The mandate is patterns in use and how well — so the strengths are first-class findings, not table decoration. A well-built state machine, a draft that survives a failed save, an open-redirect guard, a smart default: these are the good cases. Name them, cite their file:line, and mark the standouts ✅ 亮点, for three reasons:

  • They teach. A good case is the ✅ half of the 回灌 loop — it becomes the positive example a ux checklist item cites, exactly as a gap becomes the ❌ one. An audit that reports no good cases can only ever sharpen the ❌ side of the checklists.
  • They protect. "Don't regress this" is a finding. The next refactor needs to know which behaviors are load-bearing — the Strengths worth preserving block in example/task-detail.md (the loading/not-found state machine, comment-draft-preserved-on-failure, the run-all preview → confirm flow) and the ✅ 亮点 rows in example/home.md (empty-send → day's-hint fallback, drawer-not-navigate surface contract) are exactly that list.
  • They calibrate severity. A gap reads differently against a surface that is otherwise strong than against one weak throughout; the good cases are the baseline the gaps rank on.

❌ The drift this prevents: an audit whose "Patterns in use" ✅ rows are one-word ticks ("persistent", "strong", "textbook") while every gap gets a paragraph — the good cases demoted to checkboxes. example/task-detail.md is the model to match: each standout pattern carries its file:line and a ✅ 亮点 call-out, so the strengths are as legible as the gaps.

Severity rubric (shared)

  • 🔴 Breaks trust — data / input loss, stuck / permanent states, a misleading "empty" that hides a failure, silent send failure.
  • 🟠 Dead-ends or misleads — no forward path, ambiguous state, missing in-progress feedback, an empty state that isn't a real page.
  • 🟡 Friction / inconsistency / missed delight — predictability, redundant controls, progressive-disclosure gaps, CLS jank. This tier is the easiest to under-report: a pass hunting correctness skews 🔴/🟠 and glides past micro-consistency — sibling elements styled differently (options carry a number chip, the free-text row doesn't), an affordance with no label. When the surface is otherwise solid, deliberately switch into the interface-details lens for one pass, or these never get written down.

Output (shared)

See the worked example, references/example/home.md. Note which layers ran, then:

  1. Patterns in use — table (from L1/L2), grouped by pattern family, with a one-line read; mark each standout ✅ row 亮点 and back it with real evidence (file:line), never a one-word tick.
  2. Strengths / good cases (don't regress) — a dedicated section (## … — Strengths / good cases), not a table footnote: bulleted ✅ 亮点 items, each naming the behavior this surface gets right, its evidence (file:line), and why it's load-bearing — the ✅ half of the 回灌 loop and the "don't regress" list for the next refactor. Flag the ones strong enough to land as ✅ examples in ux (annotate → landed as ux <ref> ✅, wired to the Skill-feedback section). Every audit produces this section — if a surface is genuinely weak throughout, say that explicitly rather than omit it. See example/fleet.md §2 for the shape.
  3. Experience gaps — ranked; each names the finding, the ux checklist item / catalog pattern it violates, the layer + evidence it came from, and a one-line remedy.
  4. Skill feedback — real instances of existing checklist items vs new generalizable gaps worth adding to ux, and good cases worth landing as ✅ examples (see 回灌 below).

Land the findings (shared)

An audit is not finished when the findings are written — it is finished when they are landed. All three steps below are required to close a run:

  • Concrete bugs → fix the top 🔴, or file as Linear sub-issues under the "UX Audit" parent (per-page container issue → one sub-issue per finding).
  • Generalizable gaps → 回灌 ux (mandatory). Every run must close the loop back into the ux skill: for each finding that generalizes beyond this surface, add / strengthen a ux checklist item (rule + ✅/❌ example in the right module, and mirror a line into the ux Quick review), citing the audited surface as the ❌ example. This is what makes the audit continuous — each run leaves the checklists sharper than it found them. If a run genuinely surfaces no generalizable gap, say so explicitly in the report's Skill-feedback section (only validated-existing-rule instances) — silence is not an acceptable close.
  • Exemplary good cases → 回灌 ux, refining the rule (not just decorating it). The 回灌 loop has two halves: a gap sharpens a checklist item's ❌ example, a good case sharpens its ✅ one. But a good case is only worth landing if it teaches the rule something — the ✅ example is the floor; the goal is to sharpen the rule text or extract a latent sub-rule / checklist item the current rule doesn't yet state. So when an audit finds a pattern done notably well, ask: what technique does this do that the rule doesn't already demand? and fold that back — refine the prose, split a one-trigger rule into its real cases, add a checklist line — citing the surface as the ✅ example. A good case that merely re-illustrates an already-complete rule adds little; a good case that reveals a missing distinction is the point. (e.g. Fleet's scroll-into-view extracted "the re-run trigger has two flavors — async arrival vs. imperative-add-then-paint" and "the scroll axis follows the list direction" into Read §1.3; its skeleton extracted "match the text's width proportion, not just height" into Feedback §4.1 — neither was stated before.)
  • The audit → save it as references/example/<page>.md so the next run has a template.

The audit and the ux skill are a closed loop: ux is the benchmark the audit measures against, and the audit is the mechanism that keeps ux honest. Skipping the 回灌 breaks the loop and reduces the audit to a one-off review.

  • ux — the execution checklists this audit measures against, and where generalizable findings get landed.
  • agent-testing — the automation framework L3 drives (agent-browser CDP: snapshot / eval / screenshot / GIF). L3 assumes its Step 0 env + auth are green.
  • deep-review — code-level review; this skill is its design-level sibling (deep-review's ux dimension checks flows inside a code review; this skill audits a whole surface).
  • skills-audit — the same "periodic, evidence-based audit" shape, applied to the skill catalog.