# Visual Product Slop Audit > Use when auditing imagery, brand assets, UI screens, or AI product features for visual and product slop. Unlike ai-slop-typography-audit, this excludes type; written-copy slop routes to the digital-research engine. - Skill: `peterbamuhigire/visual-product-slop-audit` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add peterbamuhigire/visual-product-slop-audit` - Raw SKILL.md: https://api.skillmd.com/api/skills/peterbamuhigire/visual-product-slop-audit/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Design & Media - Author: peterbamuhigire (https://skillmd.com/u/peterbamuhigire) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/peterbamuhigire/visual-product-slop-audit --- # Visual & Product Slop Audit Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com. ## Use When - Reviewing AI-generated or AI-assisted imagery, key art, thumbnails, or brand/marketing assets. - Auditing a UI screen or a product feature that embeds generative AI. - Gate-checking any visual deliverable before it ships under the Chwezi name. - Checking a specific artifact against the current slop doctrine after the doctrine has already been refreshed. ## Do Not Use When - The concern is the typeface/type system — use `ai-slop-typography-audit`. - The concern is written copy (preambles, buzzword salad, vague attribution) — that is a writing concern owned by the `digital-research-engine` (`anti-ai-slop`); route there. - The question is whether the definition of AI slop, a banned font, or a current visual tell has changed - use `slop-doctrine-refresh-and-research-loop` first, then return here for artifact audit. ## Required Inputs | Input | Supplied by | Required? | Why | |---|---|---|---| | Assets, screens, or live feature | Requester | yes | Defines the audit surface | | Audience, channel, and stakes | Product or brand brief | yes | Calibrates severity | | Current slop taxonomy | Design doctrine | yes | Prevents stale or invented tells | - The asset(s) or screens (images, mockups, the live UI, or a description of the feature). - The context and audience, and whether it is public-facing (raises the bar — corporate slop is the highest-backlash failure, per `doctrine/references/ai-slop-taxonomy.md`). ## Workflow 1. **Run the visual tells checklist** (`references/visual-tells-checklist.md`, which operationalises `ai-slop-taxonomy.md` → Visual & video tells): waxy skin, melted backgrounds, gibberish text, floating objects, extra fingers/impossible anatomy, warped/misspelled logos, uncanny faces, mismatched lighting. **Current 2026 AI-image tells** (the classic six-finger artifact is largely fixed in late-2025/2026 models, so the tells migrated): plastic "default-render" lighting & creamy bokeh, texture-too-perfect / uniform-frequency detail, micro-anatomy & accessory errors visible only on zoom (teeth, earrings, watch faces, fused hair), garbled small text / UI chrome / gauges, impossible reflections & re-tiling patterns, and the "no-author" style-blend aesthetic — perfect, even, styleless competence that looks made by nobody. The 2026 rule of thumb: the tell is now the **absence of authored specificity**, not a hard anomaly. 2. **Run the product/interface tells checklist**: AI feature where nav/search was faster; ungrounded chatbot that can invent commitments; generative output with no verifiability/undo; decorative "AI" badges/gradients with no user benefit. 3. **Classify each finding** — *critical* (public-facing visual anomaly, warped logo, halluc\ inating bot) vs *major/minor*. 4. **Decide remake vs remediate.** Anomalous AI imagery is **remade or replaced** (art-directed, or real photography/illustration), not patched. Product slop is **removed or grounded**. 5. **State the human-craft alternative** — what a skilled designer would ship instead (per the Mission: distinct, authored, not templated). 6. **Report** findings → fixes with before/after. ## Anti-Patterns - "Touching up" a six-fingered hero image instead of remaking it. - Keeping an AI feature because it's impressive when a search bar served users better. - Treating an AI gradient/badge as design. - Declaring any polished generated image slop without naming a concrete tell and evidence. - Spot-fixing anatomy, text, or logo corruption instead of remaking the source asset. ## Machine-error review overlay Apply the shared Digital Research machine-error gate to interface and product structure. Report ME1-ME7 when repeated meaning, decorative symmetry, over-explanation, inflated emphasis, generic examples, recurring visual tics, or insight-shaped product sections are evidenced. Compare adjacent components and the actual user flow. Preserve repeated states, labels, and accessibility cues when their function is documented; otherwise remove or redesign them. Written-copy findings route to the Digital Research gate. ### Impeccable-derived AS1-AS7 review Use `cli`, `browser`, `llm_only`, or `human_review` evidence. Block decorative purple gradients, glassmorphism, neon glow, AI-beige defaults, editorial scaffolding, and decorative motion. Review AS1 default convergence, AS2 unearned hierarchy, AS3 module monoculture, AS4 decorative attention, AS5 placeholder material, AS6 copy tells, and AS7 delivery debt. A missing render, browser, or detector is `NOT_ASSESSED`, not clean. Preserve functional state transitions, accessibility cues, data encodings, and approved design-system decisions only with a recorded reason. ## Quality Standards - Inspect focal details at delivery resolution and evaluate the feature in its actual flow. - Separate anomalies, convergence, provenance gaps, and product-risk findings. - Block public release for corrupted identity, anatomy, text, or ungrounded consequential output. ## Outputs | Output | Consumer | Evidence / acceptance | |---|---|---| | Severity-rated finding register | Designer and owner | Asset location, tell, evidence, and taxonomy mapping | | Remake/remediate disposition | Production team | Named human-craft alternative for every finding | | Gate verdict | Pre-launch QA | PASS, CONDITIONAL, or BLOCKED with unresolved risks | - A findings list (tagged critical/major/minor with the tell it breaks) and a stated, human-craft remediation ready to apply. ## Examples - See `examples/slop-audit-filled.md` for a complete severity and disposition register. ## Decision Rules | Condition | Decision | Wrong-choice failure | |---|---|---| | Identity, anatomy, or embedded text is corrupted | Remake or replace and block ship | Visible anomaly damages trust | | Asset is competent but unauthored and generic | Re-art-direct with a specific point of view | Convergent brand work ships | | AI feature lacks grounding, verification, or undo | Remove or ground it | Generated output is treated as fact | | Evidence is ambiguous | Mark conditional and seek source/provenance | Suspicion is presented as proof | ## Capability Contract Read and high-resolution visual inspection are required. Review is read-only unless remediation is authorised. Network access is optional for provenance checks; editing or generation must preserve brand and usage rights. ## Degraded Mode Without original-resolution assets or live-flow access, audit the supplied evidence, mark hidden details unverified, and withhold a pass. Recover by requesting crops, source files, or a recorded walkthrough. - `examples/slop-audit-filled.md` — a real before/after audit of a Maduuka landing-page hero + embedded AI chat panel: findings (tagged critical/major/minor) → remake/remediate dispositions with the human-craft alternative named for each. ## References - `references/visual-tells-checklist.md` — the concrete, tickable image + product/interface tells (incl. the current 2026 AI-image tells), citing the taxonomy. - `doctrine/references/ai-slop-taxonomy.md`, `doctrine/design-doctrine.md` (Mission). - `doctrine/references/living-slop-refresh-protocol.md` and `slop-doctrine-refresh-and-research-loop` for refreshing changing slop definitions. - Sibling audits: `ai-slop-typography-audit` (type); digital-research `anti-ai-slop` (writing).