agentleFS
peterbamuhigirepeterbamuhigire/business-plan-skills/skills/meta-utility/ai-slop-audit/SKILL.md

ai-slop-audit

Skill4 starsChanged 19 days agoMITView on GitHub

Use when auditing a completed business-plan artefact for AI slop after a major iteration or before release; unlike `anti-ai-slop`, this read-only gate grades concrete evidence and does not produce the original content.

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

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. 4 are common, 9 are distinctive, and 2 are very specific to this project.

  • 4

    Common

    Most of the 3 edited versions we read keep these.

  • 9

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use when the user asks to analyse, review, evaluate, audit, critique, score, or de-slop a business plan, plan section, pitch deck, …

    line 25
  2. Distinctive:

    Use when the user asks "is this AI slop / does this look AI-generated / why does this feel off?".

    line 26
  3. Distinctive:

    Use as the final gate before publishing or submitting any suite output.

    line 27
  4. Common:

    Do not use to invent a flaw to pad the report — "this artefact is clean" is a wanted verdict.

    line 31
  5. Distinctive:

    Do not use as a substitute for the commercial-logic gate.

    line 32
  6. Very specific:

    Do not use to rewrite — this skill grades and prescribes.

    line 33
  7. Distinctive:

    Run the automated gates — any hit on a blocking (✗) marker fails the artefact (Step 2).

    line 44
  8. Distinctive:

    Run the human-judgement review, including the business-plan domain checks (Step 4).

    line 46
  9. Distinctive:

    Every finding cites concrete evidence from the artefact (a quote, a line, a figure, a slide).

    line 52
  10. Common:

    Run every applicable check.

    line 82
  11. Common:

    Every finding cites concrete evidence from the artefact (a quote, a line number, a figure, a slide reference, a URL).

    line 157
  12. Common:

    Do not invent a flaw to pad the report.

    line 158
  13. Distinctive:

    Do not grade a plan as AI-written from punctuation, bullet counts, headings, names, or formulaic phrases.

    line 163
  14. Very specific:

    When a file, renderer, network source, dependency registry, or visual surface is unavailable, mark that layer not assessed and return the …

    line 201
  15. Distinctive:

    Run applicable automated and human-judgement checks.

    line 228

“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

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

Overview

This is the detector. Given any artefact, it decides how strongly it reads as AI slop, names exactly why with evidence, scores it, and says how to fix each finding. Production-side prevention is the companion anti-ai-slop skill. It runs as the final gate before any plan, deck, or narrative leaves the suite, and on demand whenever a user asks for a slop review.

Cadence — run after each major iteration

This is the default mode: run this audit after each major iteration of work on the plan — each drafted plan section, each completed deck, each financial-narrative module, each significant revision, each milestone — before moving to the next. Log the verdict each time. If the verdict is F (Blocked), do not progress to the next section or iteration until the blocking findings are fixed. Treat it like a test suite that runs at every checkpoint, not a one-time final review. This is in addition to running on request and as the final pre-submission gate; the companion anti-ai-slop skill runs continuously during drafting, and this audit runs at each checkpoint to catch what slipped through.

Use When

  • Use when the user asks to analyse, review, evaluate, audit, critique, score, or de-slop a business plan, plan section, pitch deck, investment case, financial model or narrative, GTM/pricing narrative, proposal/EoI, blog post, document, or codebase.
  • Use when the user asks "is this AI slop / does this look AI-generated / why does this feel off?".
  • Use as the final gate before publishing or submitting any suite output.

Do Not Use When

  • Do not use to invent a flaw to pad the report — "this artefact is clean" is a wanted verdict.
  • Do not use as a substitute for the commercial-logic gate; a slop-clean plan can still be a bad plan.
  • Do not use to rewrite — this skill grades and prescribes; anti-ai-slop and the section skills do the fixing.

Required Inputs

  • The artefact to audit (text, plan, deck, model, document, or code)
  • The artefact type and intended audience
  • Any source list the artefact's claims are supposed to rest on

Workflow

  1. Identify the artefact type(s) and load the right checklist (Step 1 below).
  2. Run the automated gates — any hit on a blocking (✗) marker fails the artefact (Step 2).
  3. Compute the 0–100 genericness score and name its drivers (Step 3).
  4. Run the human-judgement review, including the business-plan domain checks (Step 4).
  5. Grade A/B/C/F and write the report in the fixed output format.
  6. Hand blocking and high-severity findings to anti-ai-slop and the relevant section skill for fixing.

Quality Bar

  • Every finding cites concrete evidence from the artefact (a quote, a line, a figure, a slide).
  • The verdict, the score, and the score drivers are all stated.
  • Blocking findings (hallucinated stat, fabricated citation, missing assumption behind a projection) are separated from minor ones.
  • "What's good" is listed so the fix does not strip substance.

Anti-Patterns

  • A finding with no evidence from the artefact.
  • Inventing flaws to make the report look thorough.
  • Presenting an inference as a measured fact.
  • Grading on tone alone while ignoring fabricated market data.

Outputs

  • A graded slop audit report (A/B/C/F) with a genericness score, per-marker findings, evidence, fixes, and a recommended next step.

What slop is (the yardstick)

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), business plan / pitch deck / financial narrative, proposal/EoI, code. A full plan usually spans several (prose + market data + financial model) — 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 / plan prose

  • 🤖 Focal-word density — delve/tapestry/realm/navigate/underscore/pivotal/intricate/leverage/robust/seamless 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", "in today's competitive landscape", "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.

Business plan / pitch deck / financial narrative (the suite's primary domain)

  • ✗ 🤖 Fabricated market stats: any TAM/SAM/SOM figure, growth rate, or benchmark with no named source — resolve each against its cited source.
  • ✗ 🤖 "Studies show" / "research indicates" / "industry experts agree" with no named study, report, or institution and year.
  • 🤖 Generic TAM/SAM filler that would fit any sector unchanged; round-number TAM with no stated method (top-down/bottom-up).
  • 🤖 Undifferentiated value proposition — a moat or positioning statement that would describe any competitor in the sector.
  • 🤖 Hockey-stick projections with no stated driver or assumption behind the curve (no customers × price × conversion, no source).
  • 🤖 No authored strategy — generic ambition ("become the leading provider") in place of a named choice of segment, position, and basis of competition.
  • 🤖 Funding ask that does not reconcile with the use of funds or the implementation timeline.
  • 🤖 Deck slides that could appear in any pitch unchanged; more than one assertion per slide with no supporting evidence.

Code / financial-model formulas

  • ✗ 🤖 Hallucinated/uninstallable imports & packages (slopsquatting) — resolve every dependency against its registry.
  • ✗ 🤖 Hardcoded secrets; SQL built by string interpolation; insecure defaults.
  • 🤖 Placeholder stubs/TODO/NotImplementedError/... in shipped work; dead code; duplication; tautological tests (assert true); model cells with no assumption behind them.

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 or decide that required real work (a primary customer interview, a sourced market figure, a costed assumption)? If nothing — slop.
  • 👁 Intent / authored voice: is there a strategy and a point of view, or is it relentlessly positive and viewpoint-free?
  • 👁 Specificity: real named customers, competitors, institutions, and UGX figures, or generic placeholders?
  • 👁 Hard parts: are risks, the do-nothing case, competitor response, and the downside scenario handled?
  • 👁 Domain-specific (per artefact):
    • Business plan / investment case: fabricated market stats, generic TAM/SAM filler, undifferentiated value proposition, hockey-stick projections with no basis, no authored strategy, "studies show" without a named study, funding ask that does not reconcile with the plan.
    • Pitch deck: slides that fit any company, more than one claim per slide with no evidence, no clear ask.
    • GTM / pricing narrative: cost-plus or competitor-match pricing with no stated value logic; channel list with no economics behind it.
    • Proposal / EoI: inflated superlatives, hollow analogies, unverifiable claims, no visible logic (evidence → warrant → implication).
    • Blog / social: engagement-bait, no lived experience, clichés.

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 (fabricated stat/citation, hallucinated package, secret) 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/figure/slide 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 figure, a slide reference, 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.

Responsibility audit

Do not grade a plan as AI-written from punctuation, bullet counts, headings, names, or formulaic phrases. Grade the publishing failure: unsupported claims, unlabeled examples, missing periods or definitions, absent assumptions, generic strategy, and paragraphs with no decision value. Require exact evidence for style findings and keep them separate from factual blockers. Missing source, model, or render evidence is NOT_ASSESSED.

References

  • anti-ai-slop supplies the production-time controls; this skill remains the independent detector.

Required Inputs

Input artefactSource/providerRequiredBehaviour when missing
Concrete artefact and intended audienceRequester or authorised workspaceYesStop grading and request the artefact.
Source, dependency, or render evidenceArtefact owner and available toolsConditionalMark the affected layer not assessed and qualify the verdict.

Outputs

ArtefactConsumerAcceptance condition
Graded AI-slop auditAuthor and release ownerEvery finding has evidence, severity, a correction, and a verdict consistent with blocking rules.

Evidence Produced

EvidenceFormatAcceptance condition
Graded audit reportA/B/C/F verdict, 0-100 genericness score, cited findings, and fixesEvery finding points to a phrase, line, value, screenshot region, dependency, or named unassessed check.
Release decisionPass, revise, or blocked noteAny fabricated claim, broken citation, security defect, or absent substance produces grade F and blocks release.

Capability Contract

Default to read-only. Read and search the supplied artefact; execute text, link, dependency, or rendering checks only within the authorised scope. Do not edit the artefact unless remediation is separately requested. Network access verifies current claims when available; publishing, deletion, spending, and certification remain outside this audit.

Degraded Mode

When a file, renderer, network source, dependency registry, or visual surface is unavailable, mark that layer not assessed and return the narrowest qualified grade supported by inspected evidence. Never lower the risk, invent a finding, or call an inaccessible check passed.

Decision Rules

ConditionActionFailure or risk avoided
A blocking factual, citation, security, accessibility, or dependency defect is evidencedGrade F, stop release, and name the acceptance evidence for the fixHarmful or fabricated material reaching a decision-maker.
Several non-blocking markers or weak authored intent are evidencedGrade C and require a substantive reworkCosmetic edits masking generic or empty content.
Isolated language or formatting markers remainGrade B and list targeted correctionsOver-rewriting useful authored material.
No blocker is found and the artefact has specific evidence and intentGrade A and releaseInventing audit debt merely to fill a report.

Anti-Patterns

  • Calling an artefact "AI-like" without a location. Correction: cite the exact phrase, line, value, colour, import, or missing state.
  • Editing during the default audit. Correction: preserve read-only scope and request separate remediation authority.
  • Treating an unavailable renderer or source as passed. Correction: mark the layer not assessed and qualify the verdict.
  • Padding a clean audit with invented weaknesses. Correction: allow grade A when evidence supports it.
  • Removing an unusual but purposeful phrase. Correction: record authored material worth preserving before recommending changes.
  • Giving grade B despite a fabricated statistic or broken citation. Correction: apply the blocking rule consistently and grade F.

Worked Example

A lender plan states a market growth rate without a source. Cite the exact line, record the missing source as a blocking factual defect, grade the artefact F, and require a dated authoritative source or removal of the claim. Do not rewrite the plan during the read-only audit.

Workflow

  1. Classify the artefact and inspect the supplied evidence in read-only mode.
  2. Run applicable automated and human-judgement checks; stop on a blocking marker.
  3. Score genericness, grade the artefact, and cite every finding.
  4. Recover a failed artefact by handing concrete corrections to the authorised author, then audit the revised version afresh.
<!-- dual-compat-end -->

Machine-error audit extension

Report ME1-ME7 with the affected paragraph, slide, table, or model row and the missing information delta. A repeated number is not automatically a finding when it reconciles the model. A semantic finding requires human comparison of meaning; otherwise mark it NOT_ASSESSED.

Impeccable-derived overlay audit

For dashboards, websites, decks, and rendered plans, audit AS1-AS7 with cli, browser, llm_only, or human_review evidence. Purple gradients, glassmorphism, neon glow, AI-beige defaults, decorative editorial scaffolding, and decorative motion are blocking visual findings unless a task, state, accessibility, data, or approved design-system reason is recorded. Mark visual checks not_applicable for text-only plans and unavailable render checks NOT_ASSESSED.

  • ../../meta-utility/anti-ai-slop/SKILL.md — prevention companion (write/model/design so slop never appears).
  • ../../meta-strategy/meta-critical-thinking-business-logic/SKILL.md — commercial-logic gate; a slop-clean plan can still be a bad plan.
  • ../../language/writing-quality/SKILL.md — the suite's language-quality layer; 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.