Anti Slop Review

Anti-Slop Review

Overview

Review whether an interface is specific, truthful, coherent, usable, implemented faithfully, and verified in context. “AI slop” is not a visual genre or list of banned words; it is observable loss of product meaning through unsupported copy, borrowed identity, arbitrary composition, fake data, missing behavior, or decorative polish that outruns function.

Do not pass a surface because it looks polished in screenshots. Do not invent evidence, lower severity to meet a deadline, or call skipped checks passed. Claims, authorization, persistence, formulas, accessibility, responsive behavior, performance, SEO delivery, and motion semantics require the evidence appropriate to each claim.

When to use

• Review an interface before design-system approval, implementation completion, merge, release, or publication.
• Audit generic copy, landing-page sameness, card-wall composition, trend collage, design drift, fake metrics, missing states, motion excess, accessibility, responsive, or browser behavior.
• Compare implemented frontend against approved product, reference, design-system, and project UI artifacts.
• Turn subjective “this feels AI-generated” feedback into specific evidence and fixes.

Do not use for choosing the initial direction, writing all copy from scratch, penetration testing, or proving backend/security behavior from UI evidence.

Workflow

1. Establish review truth and scope

Inspect the requested surface, Git diff, repository instructions, product routes/journeys, PRODUCT-STORY.md, PAGE-COPY.md, CLAIMS.md, REFERENCE-LEDGER.md, DESIGN.md, tokens.json, COMPONENTS.md, compiled project UI skill, MOTION.md where applicable, implementation, tests, and current browser output.

Name baseline failures separately from changes under review. Identify which claims require source, code, test, browser, accessibility-tree, network, data, or human evidence.

Load [the review criteria](references/review-criteria.md) for severity and dimension-specific checks.

Complete when: scope, approved sources of truth, affected journeys/viewports/states, and evidence methods are explicit.

2. Audit truth and copy

Map visible claims, metrics, testimonials, logos, trust marks, urgency, comparisons, and availability to CLAIMS.md and primary evidence. Check that copy:

• names the actual user, mechanism, constraint, or task;
• does not survive unchanged under a competitor's brand;
• uses proof at the supported scope;
• preserves human voice rather than formulaic cadence;
• gives forms, loading, empty, error, success, permission, and destructive states actionable language;
• avoids invented specificity and decorative jargon.

Flag a common word only when its usage is vague or unsupported. Do not run a word blacklist as a quality metric.

Complete when: every visible factual claim is supported or reported, and each copy finding quotes the exact text and explains the product impact.

3. Audit direction and composition

Compare the surface with the selected visual thesis and no-go list. Check:

• attention follows user decisions;
• layout area reflects content importance;
• sections/components have distinct jobs;
• cards, borders, gradients, glows, blur, particles, and illustrations carry meaning;
• references were adapted into principles rather than recognizable fragments;
• typography, spacing, density, and imagery/data language remain coherent;
• product-specific signature is present without overwhelming the task.

Uniform grids are not automatically wrong; they fail when unequal information is forced into equal containers without product reason.

Complete when: each visual finding points to an approved principle or observable task/hierarchy failure rather than personal taste alone.

4. Audit design-system fidelity and states

Trace rendered values and components to semantic tokens, primitives, and COMPONENTS.md. Find:

• arbitrary colors, spacing, radius, shadows, typography, durations, or z-index values;
• duplicated/forked primitives;
• undocumented variants;
• broken theme/mode mapping;
• missing default, hover, focus-visible, active/selected, disabled, loading, success, error, empty, and skeleton states where applicable;
• responsive rules that merely stack without preserving identity, priority, or action context.

Inspect source and computed/rendered output. A clean token file does not prove product code uses it.

Complete when: every divergence is documented, justified and approved, or reported with exact source/rendered evidence.

5. Audit product and data behavior

Exercise the actual journey with safe representative data. For websites, inspect conversion, form, metadata, canonical/indexability, social, semantic heading, trust/legal, and analytics event behavior. For applications, inspect routes, permissions, state ownership, mutations, retries, and reconciliation. For dashboards, inspect source, grain, formula, freshness, filter scope, card/chart/table agreement, realistic fixtures, permissions, actions, partial failure, and post-action readback.

Synthetic data must be labeled and plausible enough to expose layout/state behavior without masquerading as production evidence.

Complete when: the interface's success, failure, data, and action semantics are tested through the appropriate implementation and source boundaries.

6. Audit motion

For every material animation, identify its intent: causality, feedback, orientation, continuity, progression, or hierarchy. Verify trigger, timing token, easing, exit, interruption, repeated input, focus, data relationship, and reduced-motion equivalent.

Report motion that delays action, implies success early, copies a reference, competes with content, or has no information role. Also report abrupt changes where motion is necessary to preserve spatial understanding.

Complete when: retained motion earns its cost and normal/reduced behavior is exercised in context.

7. Audit browser, accessibility, responsive, and performance evidence

Use representative narrow and wide viewports and relevant user settings. Inspect:

• landmarks, headings, names, roles, values, descriptions, and announcements;
• keyboard order, focus-visible, focus movement/return, skip paths, and traps;
• contrast, non-color cues, target size, zoom/reflow, overflow, and content wrapping;
• loading, empty, partial, error, permission, offline, and retry states;
• console errors, failed/duplicated requests, hydration/runtime warnings;
• image/font loading, layout shift, interaction delay, and bundle/network cost proportionately;
• metadata and social output in the actual document/response for public pages.

Separate unavailable, skipped, failed, and not applicable checks.

Complete when: browser claims come from fresh post-mutation execution, not source inspection or old screenshots alone.

8. Write findings, fix blockers, and rerun

Copy [the quality report template](templates/quality-report.md) to .design-flow/QUALITY-REPORT.md. Each finding records:

• stable ID and severity;
• dimension and affected locator;
• violated source rule or observable user failure;
• evidence;
• impact;
• smallest durable correction;
• verification result after correction.

Severities:

• BLOCKER — false/misleading claim, inaccessible primary task, broken core journey, authorization/data deception, gate bypass, or material unsupported behavior;
• MAJOR — significant hierarchy, state, responsive, system, motion, conversion, or recovery failure;
• MINOR — bounded inconsistency with low task impact;
• NOTE — evidence-backed opportunity, not required for pass.

Do not hide findings in a single aggregate score. Fix blockers and relevant majors, rerun affected evidence, then report residual risk honestly.

Complete when: ## Evidence and ## Findings are substantive, every blocker has a disposition, final checks are fresh, and skipped coverage is named.

Output contract

Anti-slop review: PASS | PASS_WITH_NOTES | FAIL | BLOCKED
Surface/journey: 
Approved baseline: 

Findings
- Blocker: 
- Major: 
- Minor: 
- Notes: 

Dimensions
- Truth/copy: PASS | FAIL | SKIPPED
- Direction/composition: PASS | FAIL | SKIPPED
- Design system/states: PASS | FAIL | SKIPPED
- Product/data behavior: PASS | FAIL | SKIPPED
- Motion: PASS | FAIL | NOT_APPLICABLE | SKIPPED
- Accessibility/responsive: PASS | FAIL | SKIPPED
- Marketing/runtime/performance: PASS | FAIL | NOT_APPLICABLE | SKIPPED

Evidence
- Source/claims: 
- Static/tests/build: 
- Browser/viewports/states: 
- Accessibility tree/keyboard/contrast: 
- Network/data/action readback: 

Residual gaps
- 

Common pitfalls

• Using an unexplained numeric slop score as the verdict.
• Treating personal taste as a violation without a principle or task impact.
• Banning common words, gradients, cards, or animation categorically.
• Reviewing only the happy path or desktop screenshot.
• Calling synthetic dashboard data production evidence.
• Checking token files without tracing actual rendered values.
• Checking hidden buttons instead of server authorization.
• Calling metadata SEO readiness without runtime/indexability evidence.
• Treating automated accessibility scans as keyboard or screen-reader proof.
• Fixing visible symptoms while leaving the system rule or duplicated component unchanged.
• Passing skipped browser, role, or reduced-motion coverage silently.

Verification checklist

• [ ] Current diff, source artifacts, implementation, tests, and rendered output were inspected.
• [ ] Baseline failures are separated from regressions or scope findings.
• [ ] Every visible factual claim maps to evidence and exact allowed scope.
• [ ] Copy findings quote exact text and identify product/user impact.
• [ ] Composition findings trace to approved direction or observable hierarchy/task failure.
• [ ] Rendered values, components, variants, modes, and states trace to the approved design system.
• [ ] Product routes, conversion, permissions, state, failure, data, and action behavior were exercised as applicable.
• [ ] Dashboard metrics and fixtures are source-defined, reconciled, realistic, and honestly labeled.
• [ ] Every material animation has intent, interruption, exit, and reduced-motion evidence.
• [ ] Narrow/wide, keyboard/focus/tree/contrast, overflow/wrapping, console/network, and relevant performance checks ran after final mutation.
• [ ] Every finding includes severity, locator, evidence, impact, fix, and recheck status.
• [ ] Blockers and relevant majors were fixed or explicitly remain blocking.
• [ ] Skipped, unavailable, failed, and not-applicable evidence are reported distinctly.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: BRYANN2K
• Source: BRYANN2K/skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.