---
name: design-heuristic-review
description: "Use when handed an interface — a screen, a flow, a screenshot, or a written description of one — and asked to review it, audit its usability, critique the UX, or find what's wrong for users. Runs a structured Nielsen heuristic evaluation instead of free-form reaction: every finding is tagged to one of the ten named heuristics, rated 0-4 for severity, and paired with a concrete fix. Do NOT use for writing the interface text itself (see ui-microcopy-conventions), choosing colors/typography/visual style, implementing a mockup in code, or designing the user research study behind the interface."
metadata:
  decimalai-skill-id: "a5f58a3e-418c-4441-8bc2-20210d3c5e93"
  decimalai-version-id: "262e6f75-da74-4a9d-8db8-3f2557cfef15"
  decimalai-version: "1"
  decimalai-body-sha256: "ac830dd694ee2dde2e347e7c9be4f3bffd32449606da97ae9a281a9c6e85c02a"
  decimalai-name: "design-heuristic-review"
  decimalai-format: "1"
  decimalai-invocation: "model"
  decimalai-category: "design-media"
  decimalai-source: "https://app.decimal.ai/s/design-heuristic-review@1/SKILL.md"
---
# Design Heuristic Review

Handed a screen and asked "what's wrong with this UX?", the base model free-associates:
"the layout feels cluttered, the button could be clearer, maybe add some spacing." Real
observations, but delivered as loose reactions — untagged, unranked, and unactionable. A
team can't triage "feels cluttered." They can triage a **major visibility-of-system-status
gap with a one-line fix**.

A heuristic evaluation replaces reaction with a discipline. Every finding does three things,
or it doesn't ship as a finding:

1. **Tags to one named heuristic** — the specific one it violates, from the ten below.
2. **Carries a 0-4 severity** — on the scale below, so the team knows what to fix first.
3. **Pairs with a concrete fix** — the specific change, not "improve this."

If an observation can't be pinned to a heuristic, either you haven't found the real problem
yet (keep looking) or it's taste, not usability (drop it). This is the line the skill holds:
grade the interface against the heuristics, not against your preferences.

## The ten heuristics

Tag every finding to exactly one. Use the number and the name.

1. **Visibility of system status** — the system keeps the user informed about what is going
   on, through timely feedback. (Missing spinners, no saved/sent confirmation, no progress.)
2. **Match between the system and the real world** — speaks the user's language, real-world
   conventions, natural order. (System jargon, developer labels, unnatural sequencing.)
3. **User control and freedom** — clearly marked exits, undo and redo, no forced paths.
   (No undo, no back, no cancel, no way out of a state the user entered by mistake.)
4. **Consistency and standards** — same words, situations, actions mean the same thing;
   follows platform conventions. (Same action labeled three ways; non-standard controls.)
5. **Error prevention** — designs that stop problems before they happen. (Destructive
   actions with no confirm; inputs that permit invalid data instead of constraining it.)
6. **Recognition rather than recall** — make objects, actions, options visible; don't force
   the user to remember information across screens. (Codes to memorize, hidden options.)
7. **Flexibility and efficiency of use** — accelerators for experts, unseen by novices.
   (No shortcuts, no defaults, no way to speed up a repeated task.)
8. **Aesthetic and minimalist design** — no irrelevant or rarely needed information competing
   with the relevant. (Cluttered screens, competing calls to action, noise around the task.)
9. **Help users recognize, diagnose, and recover from errors** — plain-language error
   messages that state the problem and a way forward. (Codes like "Error 500", blame with no
   next step.)
10. **Help and documentation** — help that's easy to search, task-focused, concrete, not too
    large. (Complex fields with no guidance, no tooltips, no examples where they're needed.)

## The severity scale

Rate each finding 0-4. Severity is a function of **frequency** (how often it hits), **impact**
(how hard it is to get past), and **persistence** (does it keep biting, or is it a one-time
learn). Rate the finding, not the fix cost.

- **0 — not a usability problem.** Don't file it as a finding.
- **1 — cosmetic.** Fix only if spare time. (Minor misalignment, slightly-off wording.)
- **2 — minor.** Low priority. Users notice but work around it easily.
- **3 — major.** High priority, important to fix. Blocks or badly slows many users.
- **4 — catastrophe.** Imperative to fix before release. Causes data loss, task failure, or
  strands the user with no way forward.

Calibration anchors: a one-time cosmetic wobble is a 1, never a 3. An unconfirmed destructive
action that can lose the user's work is a 4, never a 2. Don't inflate everything to 3; a report
where every finding is "major" is uncalibrated and useless for triage.

## The finding format

One block per finding. Order the report by severity, highest first.

```
[H4 · Consistency and standards] · Severity 3 (major)
Issue: The primary action is labeled "Submit" on step 1, "Continue" on step 2, and "Done"
       on step 3, so the same forward action reads as three different things.
Fix:   Use one verb for the forward action across all steps (e.g. "Continue", with "Finish"
       only on the final step).
```

Keep the issue to what a user experiences and where; keep the fix to the specific change.

## Rules that hold the discipline

- **No untagged findings.** If it isn't tied to one of the ten, it's not in the report.
- **No unrated findings.** Every finding gets a 0-4. "This is bad" without a number can't
  be triaged.
- **No fix-less findings.** Name the concrete change. "Consider improving" is not a fix.
- **One heuristic per finding.** If an issue truly spans two, file the sharper one; don't
  smear one observation across three heuristics to pad the count.
- **Don't over-flag a clean screen.** If the interface is sound, say so and file the few
  low-severity items honestly. A short calibrated report beats a padded one.
- **Positive findings are allowed but not required.** If something is handled well, one line
  noting it is fine; it doesn't need a severity.