Use when producing or revising any research artefact to prevent generic, unsupported, repetitive, or mechanically styled output in real time; use ai-slop-audit to grade a concrete completed iteration independently.
Use when producing or revising any research artefact to prevent generic, unsupported, repetitive, or mechanically styled output in real time; use ai-slop-audit to grade a concrete completed iteration independently.
metadata
{"portable":true,"compatible_with":["claude-code","codex"],"priority":"critical","source":"ai-slop-detector research (2026-06-07), verified"}
Anti AI Slop
The guardrail that governs production quality — writing and assembling research outputs so slop never appears. Detection is the companion ai-slop-audit skill. Truthfulness is source-evaluation/references/evidence-discipline.md; this skill is its quality counterpart: a report can be fully sourced and still read as slop (generic, voiceless, template-uniform). Both gates must pass.
Real-time application (a LIVE constraint, not only a final gate)
Apply these rules continuously, as you write — to every sentence, table, finding, and slide at the moment it is drafted, not in one cleanup pass at the end. The moment you reach for a banned word, a generic placeholder, an unverified figure, or a template default, correct it in place. The ship-gate checklist below is the final confirmation, not the first time these rules are consulted.
What "AI slop" is (so you know what you are preventing)
Low-quality content produced in quantity by generative AI and pushed at people who did not ask for it (Merriam-Webster 2025 Word of the Year, verified). Three diagnostic properties (Kommers et al., "Why Slop Matters", arXiv 2601.06060, verified): superficial competence, asymmetric effort, mass producibility. The human tell named in every domain studied: absence of intent. A research output beats slop by re-internalising effort — specificity, verification, and authored judgement — before it reaches the reader.
The seven universal guardrails (apply to EVERY output)
#
Marker to prevent
Avoidance rule you MUST follow
U1
Genericness / averaging
Every section carries >=1 concrete, named, source-specific element (a real figure, named source, dated event, decision) a generic template could not produce.
U2
Superficial competence
Substance floor: include an analytic claim, comparison, or judgement the section could not exist without. If you cannot, it is filler — cut it.
U3
Confident wrongness / hallucination
Verify every statistic, quote, citation, name, and URL before it ships; cite at the point of claim; mark inferences "(inference)" and gaps "(no source found)". Defers to evidence-discipline.
U4
Volume over substance
One sourced, reasoned paragraph beats three hollow ones. Do not pad to length.
U5
Absence of authored voice / judgement
State the analytic point of view: what the evidence means, not just what it says. No relentless hedging, no viewpoint-free summary.
U6
Skipping the hard parts
Address the counter-case, the limitation, the contradicting source, the gap — not just the confirming evidence.
U7
Mechanical uniformity
Vary sentence and section structure. No rule-of-three reflex, no "it's not X, it's Y" formula, no em-dash flood, no uniform paragraph length.
Banned / high-risk vocabulary (the lexical tells)
Over-produced by LLMs (FSU/COLING-2025; PubMed "delve" +400%, verified). Allowed only when genuinely the precise term, never as default register.
Phrases: "in today's fast-paced world", "in the ever-evolving landscape of", "it is important to note that", "let's dive in", "at the end of the day", "in conclusion", "studies show" (without a named study).
Constructions: the "it's not just X, it's Y" antithesis; reflexive rule-of-three lists; em-dash used to manufacture drama; triplet adjectives ("robust, scalable, reliable").
Drop-in guardrail block (inherit in dependent skills)
ANTI-SLOP GUARDRAIL (inherit in every output):
1. SPECIFICITY FLOOR — every section carries >=1 concrete, named, source-specific element.
2. VERIFY-BEFORE-EMIT — no statistic, quote, citation, name, or URL ships unverified;
cite at point of claim; mark inferences and gaps. (Defers to evidence-discipline.)
3. AUTHORED JUDGEMENT — say what the evidence means; no viewpoint-free summary, no sycophancy.
4. COVER THE HARD PARTS — counter-case, limitation, contradicting source, gap.
5. BREAK THE TEMPLATE — vary rhythm and structure; forbid the banned-vocabulary list above.
Domain-specific avoidance (research outputs)
Report / dossier / memo: named sources at the point of every claim; a stated analytic judgement per section; the contradicting evidence shown, not hidden; varied section shapes (not every section a three-bullet list).
Tables / data: every figure carries a source and date; no round-number "estimates" presented as measured; flag computed/derived values.
.docx output: apply the professional-word-output design standard (named styles, two typefaces, real hierarchy) so the document looks human-made, not template-extruded.
Academic / proposal: visible logic (evidence -> warrant -> implication); no inflated superlatives; British English and house tone per the project.
Ship gate (run before delivering ANY output)
Every section has >=1 concrete, named, source-specific element (U1/U2).
Every stat, quote, citation, name, URL verified and cited at point of claim (U3 / evidence-discipline).
No banned vocabulary used as register; output scanned against the list.
Output states an analytic judgement; no viewpoint-free summary, no sycophancy (U5).
Counter-case, limitation, or gap addressed (U6).
Sentence and section structure varied; no rule-of-three reflex, no antithesis formula, no em-dash flood (U7).
For .docx, professional-word-output standard applied.
When in doubt, run ai-slop-audit on the draft.
If any box is unticked, the output is not ready to ship.
See also
Use When
Use throughout production and revision of every research artefact.
Do Not Use When
Do not substitute it for the independent ai-slop-audit release verdict.
Required Inputs
Input
Source/provider
If absent
Draft artefact, purpose, audience
Producing workflow
Stop generic generation and request context
Verifiable facts and dependencies
Evidence register
Remove or qualify unsupported specificity
Workflow
Identify the artefact type and intended action.
Apply specificity, verification, voice, hard-case, and variation controls while drafting.
Stop on fabricated or unverifiable content and recover by removing it or marking a gap.
Run the domain ship gate, then route the iteration to ai-slop-audit.
Outputs
Artefact
Consumer
Acceptance condition
Guarded artefact
Requesting workflow
Every section earns its place and all claims are verified or qualified
Production Evidence Guidance
The completed ship gate and verification-gap log support release review.
Production Capability Notes
Apply only capabilities authorised for the parent task. Mutation, publication, destructive action, spending, or certification requires explicit authority.
Degraded Mode
Fallback when a fact, dependency, citation, or visual check is unavailable: name the gap and return the narrowest useful qualified artefact; never turn an unassessed check into a pass.
Decision Rules
Finding
Action
Failure/risk avoided
Unsupported specificity
Verify, qualify, or remove
Hallucination
Generic non-blocking prose
Rewrite or cut
Superficial content
Intentional verified material
Preserve
Destructive over-editing
Quality Standards
Every claim is verified or qualified, every section carries intent, and errors, empty states, risks, and countercases are addressed where applicable.
Anti-Patterns
Inventing a number for specificity. Fix: verify it or remove it.
Keeping polished filler. Fix: add a sourced judgement or cut it.
Waiting for the final audit. Fix: apply controls live.
extracted-books/human-english-craft-synthesis-2026.md — human-English reader, genre, rhythm, grammar, and proof overlay; routed from the engine-level AGENTS.md.
Evidence Produced
Evidence
Consumer
Acceptance condition
Completed ship gate and verification-gap log
Producing workflow and release reviewer
Every section has intentional substance and every claim is verified, qualified, or removed
Capability Contract
Minimum capability is read-only access to the draft, facts, and dependencies. Any editing, mutation, publication, destructive action, spending, or certification requires explicit authorisation from the parent task.
Production Pitfalls
Inventing a number for specificity. Fix: verify it or remove it.
Keeping polished filler. Fix: add a decision or evidence, otherwise cut it.
Waiting for the final audit. Fix: apply controls during production.
Flattening an intentional voice. Fix: preserve authored choices that serve the audience.
Hiding an unavailable check. Fix: name the gap and its release consequence.
See also
ai-slop-audit — the per-iteration detection/audit companion.
source-evaluation / references/evidence-discipline.md — the truth gate this pairs with.
professional-word-output — document-design standard for .docx outputs.
