agentleFS
peterbamuhigirepeterbamuhigire/chwezi-dev-engine/skills/sdlc-meta/ai-slop-audit/SKILL.md

ai-slop-audit

Skill31 starsChanged 18 days agoMITView on GitHub

Use when auditing or scoring an artefact for AI slop after a major iteration or before release. Produces evidence-backed findings, severity, fixes, genericness score, and an A/B/C/F verdict.

3 other projects use a version of this, 2 of them edited.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

8 rules. 3 are common, 4 are distinctive, and 1 is very specific to this project.

  • 3

    Common

    Most of the 3 edited versions we read keep these.

  • 4

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    After EACH major iteration of work on the project at hand — a drafted section, finished feature/module, completed deck, significant …

    line 21
  2. Distinctive:

    Also run as the final gate: before publishing or delivering any engine output.

    line 40
  3. Common:

    Run every applicable check.

    line 60
  4. Common:

    Every finding cites concrete evidence from the artefact (a quote, a line number, a colour value, a screenshot region, a URL).

    line 131
  5. Common:

    Do not invent a flaw to pad the report.

    line 132
  6. Very specific:

    When the source context or runtime evidence is unavailable, mark the affected check NOT_ASSESSED.

    line 162
  7. Distinctive:

    Default to read-only. Use search, execution, visual inspection, or network verification only when available and authorised.

    line 175
  8. Distinctive:

    Do not invent findings, collapse unassessed checks into passes, or report a grade that contradicts a blocking marker.

    line 196

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

AI Slop Audit

Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178.

The detector. Given any artefact, it decides how much it reads as AI slop, names exactly why, and says how to fix each finding. Production-side prevention is the companion anti-ai-slop skill.

<!-- dual-compat-start -->

Use When

  • After EACH major iteration of work on the project at hand — a drafted section, finished feature/module, completed deck, significant revision, or milestone. Log the verdict; block progression on grade F.
  • The user asks to analyse, review, evaluate, audit, critique, score, or de-slop any website, web/mobile app, business plan, SRS/spec, proposal/EoI, blog/article, social post, marketing copy, document, image/video, or codebase.
  • The user asks "is this AI slop / does this look AI-generated / why does this feel off?".
  • As the final gate before publishing or delivering any output.

Do Not Use When

  • There is no concrete artefact to inspect yet — generate it first under anti-ai-slop, then audit.

Required Inputs

  • The artefact itself (text, files, screenshots, repo path, or URL) and its claimed purpose/audience.
  • For code: access to the dependency manifest so packages can be resolved against their registry.

When this runs

Cadence — run after EACH major iteration of work on the project at hand. This is the default mode: whenever a meaningful unit of work is completed — a drafted section, a finished feature or module, a completed slide deck, a significant revision, a milestone — run this audit on what was just produced before moving on. Log the verdict. If the verdict is F (Blocked), do not progress to the next iteration until the blocking findings are fixed. Treat it like a test suite that runs at every checkpoint, not a one-time final review.

Also auto-run on request: when the user asks to analyse, review, evaluate, audit, critique, score, or de-slop any of: a website, web/mobile app, business plan, SRS or technical spec, proposal/EoI, blog or article, social-media post or campaign, marketing copy, document, image/video, or codebase — or asks "is this AI slop / does this look AI-generated / why does this feel off?".

Also run as the final gate before publishing or delivering any engine output.

The companion anti-ai-slop skill runs continuously during generation; this audit runs at each checkpoint to catch what slipped through.

What slop is (the yardstick)

The audit also checks the production method: was the artefact built in a named slice, was existing context inspected, were hard cases exercised, and was at least one concrete refinement made from observed evidence? If the record contains only a bulk-generation claim, headings, or a final screenshot, mark craft evidence NOT ASSESSED and do not award readiness from structure alone.

Low-quality content produced in quantity by AI and pushed at people who did not ask for it (Merriam-Webster 2025 WOTY, verified). Three diagnostic properties (Kommers et al., arXiv 2601.06060): superficial competence, asymmetric effort, mass producibility. The human tell: absence of intent. You are measuring how strongly an artefact exhibits these.

Audit method — layered, cheapest first

Step 1 — Identify artefact type and load the right checklist

Map the artefact to one or more domains: written content (EN/FR), UI/UX, app/product, image/video, code. A "project" (e.g. a website or app) usually spans several — audit each layer.

Step 2 — Automated gates (🤖, machine-checkable) — any hit is hard evidence

Run every applicable check; a hit on a blocking marker (✗) fails the artefact outright.

Written content

  • 🤖 Focal-word density — delve/tapestry/realm/navigate/underscore/pivotal/intricate/leverage etc. >2 per 500 words.
  • 🤖 Em-dash density >1 per paragraph; reflexive rule-of-three; "it's not X, it's Y" repetition; uniform 15-25-word sentences (low burstiness).
  • 🤖 Transition clichés ("in today's fast-paced world", "let's dive in", "in conclusion").
  • 🤖 Mechanical formatting: Title-Case headers, excess bold, decorative emoji, leftover tool markup ("oaicite", "contentReference").
  • ✗ 🤖 Broken/fake citations: dead URLs, invalid DOI/ISBN, fabricated stats, utm_source params copied in.
  • French: "plongeons dans", "il est important de noter que", "force est de constater", filler connectors.

UI/UX

  • 🤖 Indigo/purple-gradient default (HSL 250-280°, sat 70%+); Inter/Roboto/Poppins-only; uniform border-radius; glassmorphism; gradient text; shadcn coloured card-border.
  • ✗ 🤖 Dark-mode/body contrast <4.5:1 (WCAG fail).
  • ✗ 🤖 Missing states (error/empty/loading/focus/disabled) — state-coverage audit.

Code

  • ✗ 🤖 Hallucinated/uninstallable imports & packages (slopsquatting) — resolve every dependency against its registry. Package-hallucination is measured, not hypothetical: Spracklen et al. (USENIX Security 2025) found 19.7% of LLM-recommended dependencies did not exist.
  • ✗ 🤖 Hardcoded secrets; SQL built by string interpolation; innerHTML = userInput (XSS); insecure defaults. AI-generated code fails security checks at scale: Veracode found 45% of AI code samples introduced a known vulnerability, rising to 86% for XSS and 88% for log injection.
  • 🤖 Placeholder stubs/TODO/NotImplementedError/... in shipped code; dead code; duplication; bare-except; cross-language constructs; inconsistent style in one file; tautological tests (assert true). Copy-paste/duplication is rising in AI-assisted repos: GitClear measured cloned code blocks up from 8.3% (2020) to 12.3% (2024).

Image/video

  • 🤖 Missing/contradictory C2PA provenance; SynthID absence (Google-only — absence != authentic); ELA/JPEG-forensics anomalies.

Step 3 — Structural score (🤖) → 0-100 "genericness"

Combine burstiness, focal-word density, duplication, and template-similarity into a single genericness score. Higher = more slop-like. Report the score and its drivers.

Step 4 — Human-judgement review (👁) — the checklist no tool replaces

  • 👁 Substance: what does this assert/decide that required real work? If nothing — slop.
  • 👁 Intent / authored voice: is there a point of view, or is it relentlessly positive and viewpoint-free?
  • 👁 Specificity: real named examples/people/numbers, or generic placeholders?
  • 👁 Hard parts: are errors, edge cases, risks, counter-arguments handled?
  • 👁 Visuals: anatomy (hands/eyes/teeth), "AI sheen", garbled text-in-image, impossible geometry, video "boiling"/lip-sync.
  • 👁 Product/app: wrapper with nothing added? fleeceware/dark-pattern billing? deceptive AI claims? fake reviews?
  • 👁 Domain-specific (per artefact):
    • Business plan: fabricated market stats, generic TAM/SAM filler, no authored strategy, "studies show" without a named study.
    • SRS/spec: vague requirements, placeholder "Challenges and Future Prospects" sections, missing edge/error specs, hallucinated APIs.
    • Proposal/EoI: inflated superlatives, hollow analogies, unverifiable claims, no visible logic (evidence→warrant→implication).
    • Blog/social: engagement-bait, no lived experience, clichés, AI-sheen imagery.

Scoring & verdict

Aggregate into a grade:

GradeMeaningTrigger
A — CleanNo blocking hits; genericness low; substance & intent presentship
B — Minor slopA few automated hits, no blockers; some genericnessfix listed items
C — SlopyMultiple automated hits or weak substance/intentrework before ship
F — BlockedAny ✗ blocker (hallucinated fact/citation/package, secret, WCAG fail, missing states) OR no substance at alldo not ship

Output format (the audit report)

# AI Slop Audit — <artefact name> — <date>
Verdict: <A/B/C/F>   Genericness score: <0-100>
Artefact type(s): <...>

## Blocking findings (✗) — must fix
- [marker] <what was found> · evidence: <quote/line/URL/screenshot ref> · fix: <concrete action>

## Slop findings (by severity)
- [marker] <finding> · evidence: <...> · fix: <...>

## What's good (so it isn't stripped in the fix)
- <substantive, specific, authored elements worth keeping>

## Recommended next step
- <rework / targeted fixes / ship>

Discipline (anti-hallucination — applies to the audit itself)

  • Every finding cites concrete evidence from the artefact (a quote, a line number, a colour value, a screenshot region, a URL). No finding without evidence.
  • Do not invent a flaw to pad the report. "This artefact is clean" is a valid, wanted verdict.
  • Mark inferences "(inference)"; never present a guess as a measured fact.

See also

Inputs

ArtefactRequired?Purpose
Concrete artefact and intended audienceyesGround every finding
Source or dependency evidenceconditionalVerify citations, packages, and claims

Outputs

ArtefactConsumerAcceptance condition
Graded audit reportAuthor or release ownerEvery finding cites concrete evidence and a fix

Evidence Produced

CategoryArtifactFormatExample
CorrectnessAI slop audit reportA/B/C/F verdict, genericness score, evidence, fixes, and unassessed checksdocs/audits/ai-slop-<artefact>.md
<!-- dual-compat-end -->

Machine-error audit extension

Report ME1-ME7 with exact file, line, component, or section evidence. Separate semantic duplication from intentionally repeated identifiers, schemas, error strings, security warnings, and trace links. When the source context or runtime evidence is unavailable, mark the affected check NOT_ASSESSED.

Impeccable-derived overlay audit

For UI, web, dashboard, presentation, or rendered-document artefacts, report AS1-AS7 with exact file/line or screenshot evidence and an evidence mode (cli, browser, llm_only, or human_review). Purple gradients, glassmorphism, neon glow, AI-beige defaults, decorative editorial scaffolding, and decorative motion are blocking visual findings unless a functional state, accessibility need, data encoding, or approved design-system reason is recorded. Mark unavailable visual checks NOT_ASSESSED; do not flag functional repetition.

Capability contract

Default to read-only. Use search, execution, visual inspection, or network verification only when available and authorised. Never edit the audited artefact unless remediation is separately requested.

Degraded mode

Mark inaccessible checks not assessed; do not lower their risk silently or invent findings to complete the template.

Decision rules

EvidenceGrade effectAction
Any blocking markerFBlock release and name the fix
Multiple non-blocking markers or weak substanceCRework before release
Minor isolated markersBApply targeted fixes
No blockers, low genericness, clear intentAShip

Workflow

Classify the artefact, run applicable automated checks, assess substance and intent, grade it, and cite evidence for every finding.

Quality Standards

Do not invent findings, collapse unassessed checks into passes, or report a grade that contradicts a blocking marker.

Anti-Patterns

  • Reporting a vague “AI feel”. Fix: cite the exact phrase, line, colour, import, or state gap.
  • Padding a clean audit with invented flaws. Fix: allow an A verdict.
  • Editing during a read-only audit. Fix: separate diagnosis from remediation authority.
  • Treating an inaccessible check as passed. Fix: mark it not assessed.
  • Removing useful authored material during cleanup. Fix: record what must be preserved.

Responsibility audit

The 25 writing signs are editorial warnings, not proof that a person used AI. Report a style pattern only with exact artefact evidence and a reader-value or clarity consequence. Audit every load-bearing claim for scope, date, source fit, transformation, limitation, and decision effect; audit examples for hypothetical labelling. Keep authorship questions separate from unsupported claims, fabricated evidence, provenance gaps, security defects, accessibility failures, and genericness. Never turn a missing source or render into a pass: record NOT_ASSESSED.

  • AI-slop responsible publishing standard

References

  • anti-ai-slop supplies the production-time prevention controls.

See also

  • anti-ai-slop — prevention companion (write/design/code so slop never appears).
  • Host engine house-style skill — apply domain tone on top.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.