AI writing quirks rarely come down to one forbidden word or punctuation mark. They appear in clusters: generic openings, inflated claims, repeated rhetorical formulas, automatic praise, polished paragraphs that never commit to a point, and specificity the source never supplied.
For a go-to-market (GTM) team, these habits do more than make copy sound synthetic. They can blur product claims, overstate evidence, flatten a company's voice, and turn customer insight into generic content. This guide gives editors a practical way to find and repair 48 common patterns across blog posts, reports, emails, product copy, and social writing.
The most useful question is: What about this writing makes it feel generic, overproduced, evasive, or unnatural? Identify that problem, then fix it without changing the facts or erasing the author's voice.
An em dash, an unusual word, or a tidy list does not establish AI authorship. Humans use these devices too. This guide is an editorial taxonomy, not a validated AI detector or a ranking of which phrases models use most often. A 2023 study found that the detectors it tested frequently misclassified writing by non-native English speakers; that finding is specific to its sample and tested systems, but it illustrates the risk of treating a style judgment as proof. Liang et al., Patterns
All examples and numbers below are invented for illustration. They are not claims about a real company. A rewrite may use only details supplied in its example; where a better sentence needs missing facts, the guide explicitly says so.
How to use this guide
- For a fast cleanup, use the two-minute check and the final checklist.
- For a stubborn draft, identify the relevant patterns in the catalogue and follow the editing workflow.
- For repeatable AI editing, use the reusable prompt near the end.
- For a team or application, use the review rubric and automation guidance.
The two-minute check
Read the opening, a paragraph from the middle, and the ending. Ask:
- Does this say anything specific? Can you identify the claim, action, evidence, or observation?
- Could this describe almost any organisation? Swap the company name. If nothing breaks, check whether the passage needs more substance.
- Is it explaining the point or announcing its importance? “This is a powerful shift” needs a description of what changed.
- Does it keep using the same construction? Watch for repeated contrasts, trios, rhetorical questions, and identical paragraph endings.
- Does the tone belong to this writer and situation? A bug report probably does not need motivational language.
- Would deleting the first or last paragraph remove information? If neither contributes, cut them.
- Did polish conceal uncertainty? Check numbers, causation, promises, and assertions of authority before polishing further.
Treat a cluster of these problems as a reason to edit. Do not convert the count into a probability of AI authorship.
The pattern catalogue
1. Stock vocabulary and inflated claims
The problem is usually the work a phrase avoids doing. Replacing one grand word with another leaves the underlying weakness intact.
| Pattern | What to notice | Better editing move |
|---|---|---|
| 1. Fashionable verbs | “Delve,” “unlock,” “elevate,” “harness,” “foster,” “leverage” appear where a plain action would be clearer. | Name the action: examine, open, improve, use, support. Check the intended meaning before substituting. |
| 2. Praise without a criterion | “Powerful,” “robust,” “seamless,” “intuitive,” “comprehensive.” | State the capability or the evidence supporting the adjective. If unknown, remove the claim or request evidence. |
| 3. Borrowed importance | “Transformative,” “revolutionary,” “groundbreaking,” “a paradigm shift.” | Describe the change and its consequence. Retain the stronger label only when justified. |
| 4. Corporate abstractions | “Drive meaningful outcomes,” “deliver value,” “enable excellence.” | Ask which outcome, for whom, and through what action. |
| 5. Scenic metaphors | “Navigate the landscape,” “a rich tapestry,” “embark on a journey,” “an ecosystem of possibilities.” | Name the market, issue, history, process, or group being discussed. |
| 6. Unnecessary intensifiers | “Truly,” “deeply,” “incredibly,” “genuinely,” “absolutely.” | Delete experimentally. Keep only if the distinction matters. |
| 7. Inflated verb phrases | “Conduct an evaluation of,” “facilitate the implementation of.” | Use “evaluate” or “help implement,” if that preserves meaning. |
| 8. Decorative precision | “Multifaceted,” “nuanced,” “holistic,” “dynamic.” | Identify the relevant dimensions, competing considerations, scope, or change. |
Before: “Our platform leverages powerful automation to streamline workflows and unlock meaningful productivity gains.”
Minimal edit: “Our platform automates parts of the workflow.”
The minimal edit removes unsupported promotion, but it is still vague. The next step is to obtain the actual workflow and capability. If the supplied facts say that the product assigns each incoming request to its account owner, write that. Do not invent a feature or time saving to make the sentence sound concrete.
When to keep the original wording: “Robust” can be a technical property, “ecosystem” can describe a biological system, and “leverage” can refer to borrowing. An accurate term does not become wrong because AI also uses it.
2. Formulaic rhetoric
| Pattern | Example | Repair |
|---|---|---|
| 9. Forced contrast | “This isn't about better reports. It's about better decisions.” | State the actual relationship: “Better reports help managers decide which accounts need attention.” Use only when that purpose is established. |
| 10. Serial three-part slogans | “Capture. Connect. Create.” | Expand the useful action, or retain the trio only if it accurately describes the process. |
| 11. Predictable question and answer | “The result? A smarter team.” | State what happened. If the result is unknown, describe the intended benefit as an intention. |
| 12. Stacked symmetrical sentences | “Less friction. More focus. Less noise. More clarity.” | Explain the specific improvement in a connected sentence. |
| 13. Fake revelation | “Here's the thing,” “The real secret,” “What nobody tells you.” | Begin with the claim and substantiate it. |
| 14. Universal scene-setting | “In today's fast-paced digital world…” | Begin with the reader's actual problem, finding, or task. |
| 15. Manufactured suspense | “But something remarkable happened next.” | Name the event unless suspense is an intentional feature of the piece. |
| 16. Stock concession | “While X has benefits, it also presents challenges.” | Identify the relevant benefit and cost, then explain how they affect the decision. |
Before: “This isn't just a dashboard. It's a decision engine. The result? Clarity, confidence, and control.”
Repair: Describe what the dashboard displays and which decision it supports. There is not enough factual content in the original to supply that answer honestly.
When to keep the device: Do not flatten a founder's purposeful rhythm into generic explanatory prose. A punchy contrast, a running metaphor, or a repeated phrase can carry the argument; repair the unsupported claim, not the personality. A real distinction can prevent a mistake: “The export contains account IDs, not customer names.” Three steps can genuinely be the right number. A rhetorical question can help introduce a question the reader already has. The issue is repetition or empty emphasis.
3. Paragraphs and structures that feel assembled
| Pattern | What it looks like | Repair |
|---|---|---|
| 17. The repeated template | Every section has a definition, three benefits, an example, and a mini-summary. | Give each section the structure its argument needs. |
| 18. Excessive signposting | “First, let's explore…” followed by “Next, we'll examine…” | Move directly to the substance when the headings already establish order. |
| 19. Repeated conclusions | Opening summary, section summaries, key takeaways, final thoughts all repeat the same claim. | Keep repetition where it serves navigation; remove repeated paragraphs that add no use. |
| 20. Bullets replacing explanation | Related causes and consequences are broken into disconnected fragments. | Use prose for the relationship, bullets for independent items. |
| 21. Heading overload | A heading introduces every two sentences. | Combine related material unless the document is deliberately a lookup reference. |
| 22. Artificial balance | Every option gets exactly the same space, even when one has much stronger evidence. | Allocate space by relevance, complexity, and evidence. |
| 23. Synonym drift | “Client,” “customer,” “user,” and “buyer” rotate as if interchangeable. | Use stable names for stable concepts. Define real differences. |
| 24. Empty closure | “By embracing these principles, organisations can thrive in an ever-changing world.” | End with the decision, implication, unresolved question, or final useful fact. Stop if the piece is already complete. |
Before: “The team should improve onboarding. This will enhance the onboarding experience. A better onboarding experience can support stronger onboarding outcomes.”
After: “The team should improve onboarding.”
The original contains one recommendation stated three times. The next editorial question is what should change and why. More varied synonyms cannot supply the missing thinking.
When to keep the structure: An executive summary and a full report serve different reading paths. Repeated layouts help users compare products and follow procedures. In this guide, tables support lookup; they would be a poor choice for telling a personal story.
4. An unnatural relationship with the reader
| Pattern | Example | Repair |
|---|---|---|
| 25. Automatic praise | “Great question! You're absolutely right to focus on this.” | Answer the question. Acknowledge a specific useful observation if relevant. |
| 26. Unfounded agreement | “You're spot on” before checking the claim. | Assess the claim first. Explain agreement or disagreement. |
| 27. Canned empathy | “I understand how incredibly frustrating this must be.” | Acknowledge the stated problem proportionately and offer a useful next step. |
| 28. Simulated familiarity | “We've all been there,” “You know that feeling.” | Describe the situation without assuming a shared experience. |
| 29. Synthetic personal history | “I've spent years helping teams solve this.” | Retain only if this is a truthful statement by the actual author. |
| 30. Compulsory enthusiasm | Every update is exciting, every feature a breakthrough. | Match emotion to the event and writer. |
| 31. Performative bluntness | “Let's be brutally honest. Most teams are doing this wrong.” | State the supported criticism and identify its scope. |
| 32. Chat residue | “Certainly! Here's your polished draft,” “Would you like me to expand?” | Remove assistant-to-user commentary from the finished deliverable. |
Before: “Absolutely, you're asking all the right questions. Navigating this challenge can feel overwhelming, but don't worry—we've got you covered.”
After: Begin with the answer or the next practical step. The original has no substantive information to preserve.
When to keep warmth: A considerate acknowledgement can matter in support messages or sensitive conversations. Preserve real empathy. Remove excessive reassurance, invented familiarity, and praise that substitutes for a response.
5. Weak thinking disguised as fluent prose
These problems deserve attention before punctuation and vocabulary because editing can otherwise make an unsupported claim more persuasive.
| Pattern | Example | Repair |
|---|---|---|
| 33. Authority without a source | “Research shows,” “Experts agree,” “Industry leaders recognise.” | Identify the supporting source and its scope, or remove the authority claim. |
| 34. Observation promoted into causation | “Retention rose after onboarding changed, proving the new flow works.” | Preserve the observed sequence; qualify the causal interpretation. |
| 35. Certainty without support | “This will eliminate churn.” | State what the intervention is intended or evidenced to affect. |
| 36. Foggy hedging | “This could potentially perhaps have an impact.” | Express the actual uncertainty once: “The effect is unknown.” |
| 37. Hidden actor | “It was decided that implementation would be delayed.” | Name the decision-maker if known and relevant. Do not guess. |
| 38. Pseudo-specific language | “Actionable insights,” “measurable impact,” “strategic alignment.” | Identify the action, measure, or agreement. |
| 39. Unearned significance | “This marks a pivotal moment in the company's evolution.” | Show the specific consequence that makes the change important. |
| 40. Generic advice without a decision rule | “Monitor performance and optimise continuously.” | Identify what to monitor, what result matters, and what action follows; request missing thresholds. |
Before: “Following the rollout, conversion increased from 4% to 5%, proving the new experience drove a 25% improvement.”
After: “Conversion increased from 4% to 5% after the rollout. That is a one-percentage-point increase, or 25% relative growth. The figures alone do not establish that the rollout caused it.”
The edit keeps the numbers, distinguishes percentage points from relative change, and limits the causal claim to what the example supports. In real work, also check the denominator, comparison period, sample size, and definition of conversion.
When to keep uncertainty: “May,” “estimated,” and “in this sample” often carry essential meaning. Removing them for confidence can make a sentence false. Keep necessary conditions even when they make a sentence longer.
6. Punctuation, formatting, and invented labels
| Pattern | What to notice | Repair |
|---|---|---|
| 41. Em dashes doing every job | Dashes introduce explanations, parenthetical remarks, conclusions, and emphasis in nearly every paragraph. | Choose punctuation by function; split overloaded sentences. Keep useful dashes. |
| 42. Colon-heavy micro-headlines | “The challenge: complexity. The answer: integration. The benefit: clarity.” | Use a normal sentence where the labels contribute no navigation. |
| 43. Constant bold emphasis | Every paragraph contains several bold phrases. | Emphasise the few points a scanning reader must find. |
| 44. Emoji as a default outline | Every heading begins with a rocket, target, bulb, or check mark. | Use icons only when appropriate to the channel and helpful to navigation. |
| 45. Excessive fragments | “A better process. A clearer picture. A stronger team.” | Reconnect claims so the relationship becomes explicit. |
| 46. Invented compounds | “Insight-led revenue-readiness flywheel” or “context-aware outcome enablement.” | Describe the mechanism with familiar nouns and verbs. |
| 47. Formatting debris | Unresolved placeholders, broken citations, raw Markdown, “insert example here.” | Check the final rendering and resolve or deliberately label each remaining placeholder. |
| 48. Unnecessary frameworks | Ordinary steps are given a capitalised name and a strained acronym. | Keep a framework only if its categories help someone make or remember a decision. |
Before: “Build a unified, insight-powered, customer-centric growth-enablement layer.”
Repair: Ask what is being built, which information it contains, and who uses it. If those answers are unavailable, the sentence is a proposal label, not an explanation.
When to keep technical language: “Knowledge graph,” “retrieval-augmented generation,” and “role-based access control” can identify specific concepts. Explain them for unfamiliar readers. Replacing a precise term with “smart system” makes the writing less useful.
The phrase watchlist
Use this as a search list. A match means “inspect in context,” not “delete automatically.”
| Search for | Ask | Usual action |
|---|---|---|
| “It's worth noting,” “Importantly,” “It is important to remember” | Does the sentence need an announcement of importance? | Start with the fact. |
| “In order to” | Does “to” mean the same thing here? | Shorten if it does. |
| “Due to the fact that” | Does “because” preserve the relation? | Usually replace. |
| “In today's,” “In an era of,” “As the world becomes” | Does the broad opening establish necessary context? | Replace with the specific situation or cut. |
| “Unlock the potential,” “Take it to the next level” | Which capability or change? | Specify or remove. |
| “Seamlessly integrates” | What connects, how, and with what limits? | State the integration and substantiated constraints. |
| “Best-in-class,” “Industry-leading” | Compared with whom, on which measure? | Substantiate or remove. |
| “A testament to,” “Underscores the importance of” | What does the evidence actually show? | Explain its implication. |
| “Not just X, but Y,” “It's not about X” | Is the contrast real and useful? | Preserve valid distinctions; simplify ornamental ones. |
| “The result?”, “The kicker?”, “Here's why” | Is this transition earning its space? | Usually connect the sentences directly. |
| “Ultimately,” “In conclusion,” “Final thoughts” | Is the ending adding a useful synthesis? | Keep the synthesis; remove repetitive packaging. |
| “A wide range of,” “Various,” “Multiple” | Does the reader need the actual items? | Name the relevant ones when known. |
Do not replace “leverage” with “utilise” and call the job finished. Nor should you rotate synonyms throughout a document solely for variation. A reader should not have to infer whether a “workspace,” “portal,” and “hub” are the same thing.
Five complete editing examples
Marketing copy
Draft
“In today's fast-paced business landscape, teams need more than just another tool. They need a powerful partner that unlocks productivity. Our platform brings project updates and customer notes together, empowering teams to navigate complexity with confidence.”
Edited
“Our platform brings project updates and customer notes together.”
What changed: Removed generic context, a forced contrast, an unsupported productivity claim, and an abstract benefit. The remaining sentence preserves the supplied capability. To strengthen it further, obtain a concrete use case or supported benefit from the author.
Executive update
Draft
“We're excited to share meaningful momentum across the business. New MRR reached $12,000, while churned MRR was $9,000. This underscores the strength of our strategy and positions us for sustainable growth.”
Edited
“New MRR was $12,000 and churned MRR was $9,000. New MRR exceeded churned MRR by $3,000. These figures alone do not establish sustainable growth.”
What changed: Replaced celebration with the result and removed the unsupported strategic conclusion. Do not label the $3,000 as total net MRR growth without checking expansion, contraction, and any other movements included in the company's definition.
Customer email
Draft
“I hope this message finds you well. I wanted to take a moment to reach out regarding the implementation timeline. It would be greatly appreciated if you could kindly confirm whether Friday remains feasible for your team.”
Edited
“Could you confirm whether Friday still works for implementation?”
What changed: Removed duplicate politeness and narration of the act of sending an email. A greeting or a brief personal line can stay if it fits the relationship. No response deadline has been invented.
Technical explanation
Draft
“By leveraging a robust caching mechanism, the system seamlessly optimises performance. Cached responses expire after five minutes. This ensures users always benefit from fresh, reliable information.”
Edited
“The system caches responses for five minutes. Users may receive information that is up to five minutes old relative to the cached response.”
What changed: Removed unmeasured performance language and corrected the unsupported promise of freshness. The example says nothing about when the underlying source was last updated, so the edit does not imply the source data itself is always less than five minutes old.
Professional social post
Draft
“Hot take: more dashboards won't save your business. Better decisions will. Last month, our team spent six hours reconciling two reports that used different definitions of an active customer. The lesson? Clarity. Alignment. Execution.”
Edited
“Last month, our team spent six hours reconciling two reports because they used different definitions of an active customer. We need to agree on that definition before comparing the reports again.”
What changed: Kept the specific experience supplied by the draft and replaced the slogans with a relevant next action. If the anecdote is not true of the actual author, neither version is publishable as a personal account.
A reliable editing workflow
Step 1 — Establish the job of the piece
Complete this sentence: “This is for [reader], who needs to [understand, decide, or do something].”
Then identify the format, desired tone, and necessary detail. An internal engineering note, a condolence message, and a product announcement should not all sound alike. If the author's voice matters, use two or three representative samples they actually wrote.
Step 2 — Protect the meaning
Before rewriting, identify the claims that must survive: names, dates, figures, comparisons, negations, conditions, definitions, quotations, attribution, and degree of certainty. Mark unsupported statements separately.
Watch especially for changes from “may” to “will,” “some” to “all,” “associated with” to “caused,” and “we recommend” to “we have decided.” These are factual or commitment changes, even when the new sentence sounds better.
Step 3 — Find the actual substance
Label each paragraph by its purpose: claim, evidence, explanation, example, instruction, decision, or necessary context. Paragraphs with no identifiable purpose are candidates for deletion or revision.
Ask the author for missing substance only when it matters. Keep a restrained general statement when the available information is genuinely general. Specificity is useful only when true.
Step 4 — Repair the structure
Move the answer or main finding where the reader can find it. Combine repeated points. Reorder evidence so it supports the claim near it. Remove template sections that do not serve the document. Preserve a summary when readers will need a separate fast reading path.
Step 5 — Repair sentences
Name the actor where useful, bring the verb closer to its subject, and replace abstract actions with concrete ones. Break overloaded sentences, but retain longer sentences when they express a useful relationship clearly. Leave accurate technical terms intact.
Step 6 — Review tone and patterns
Look across the whole draft for repeated constructions, forced enthusiasm, borrowed authority, chat residue, and decorative formatting. Judge the concentration of a device rather than treating each occurrence as an error.
Read aloud. Listen for passages that sound like a keynote speech when they should sound like a colleague explaining a result. Do not force a particular sentence-length distribution or add slang to manufacture variety.
Step 7 — Compare against the source
Check every number, quotation, proper noun, qualification, and commitment. Ensure the edit did not add a capability, experience, customer outcome, deadline, or causal claim. Check that citations still support the edited wording. Where information remains missing, flag it separately from the publishable draft.
Step 8 — Stop when the piece works
The stopping rule is practical: the reader can understand the point, the evidence supports the claims, the voice fits, and no distracting repetition remains. Do not keep rewriting merely to make the prose less recognisable as AI-assisted.
Preserve personality while editing
Good editing should preserve an author's distinctive observations, humour, precise terminology, justified emotion, and useful rhythm. It can also preserve an occasional fragment, unusual turn of phrase, em dash, or repeated word.
Avoid these overcorrections:
- Adding deliberate mistakes. Typos and broken grammar reduce quality; they do not establish authenticity.
- Making everything casual. Contractions, slang, profanity, and chatty asides may not fit the writer or reader.
- Making every sentence short. This can create the same repetitive cadence you were trying to remove.
- Removing every qualification. Directness does not require certainty.
- Inventing opinions or anecdotes. A stronger personal voice must still belong to the actual author.
- Flattening culturally different English. Edit for comprehension and the agreed voice, not conformity to one imagined native speaker.
- Banning whole categories of language. Metaphor, passive voice, technical vocabulary, and parallel structure all have legitimate uses.
- Replacing all warmth with commands. A helpful message can be clear and considerate.
If an unusual phrase is clear, accurate, and characteristic of the author, there may be nothing to fix.
Adjust the review to the format
| Format | Prioritise | Watch particularly for |
|---|---|---|
| Executive update | Result, comparison, implication, decision | Celebration without a benchmark; numbers without definitions; unsupported explanations |
| Marketing page | Audience, problem, capability, substantiated benefit | Interchangeable claims; “seamless”; invented differentiation; claims stronger than the product |
| Article or essay | Argument, evidence, connected explanation | Generic introductions; identical section structures; repeated conclusions |
| Customer email | Purpose, appropriate warmth, request, necessary context | Excessive courtesy; vague asks; made-up deadlines; promises the sender cannot make |
| Technical documentation | Exact behaviour, prerequisites, limits, instructions | Marketing adjectives; shifting terminology; removing important conditions |
| Research summary | Methods, findings, limitations, attribution | Causal overstatement; broad consensus claims; citations that support only part of a sentence |
| Social post | A useful observation or truthful experience | Manufactured controversy; fake confession; one-sentence-per-line theatrics; engagement bait |
| Interface copy | Action, state, consequence, recovery | Cute language hiding errors; unnecessary enthusiasm; explanations of internal implementation |
A review rubric for teams
Use the following as an editorial discussion aid. It is not a validated score or an authorship test.
| Dimension | Needs revision | Acceptable | Strong |
|---|---|---|---|
| Substance | Mostly broad claims or repetition | Main point is present | Each section contributes necessary information |
| Evidence | Unsupported, invented, or overstated | Claims generally match supplied support | Evidence, interpretation, and uncertainty are clearly distinguished |
| Clarity | Reader must decode abstractions | Meaning is understandable | Actions and relationships are easy to follow |
| Voice | Generic or inappropriate | Fits the situation | Fits the author and reader without performance |
| Structure | Formulaic or hard to navigate | Organised and usable | Order and format support the reader's task |
| Fidelity | Facts or commitments changed | Source meaning retained | Important nuances preserved and gaps flagged |
Treat evidence and fidelity problems as release blockers for the affected claims. A pleasing voice cannot compensate for a fabricated fact. Fix reader comprehension before polishing optional stylistic preferences.
Automating parts of the check
A lightweight writing linter can flag exact phrases from the watchlist, unresolved placeholders, repeated openings, suspiciously duplicated passages, and frequent formatting patterns. These are review candidates. Set any frequency thresholds as local preferences, not universal limits or AI probabilities.
Use human or contextual model review for whether a contrast is useful, a metaphor clarifies, a technical term is necessary, or a claim exceeds its evidence. A word search cannot make those judgments reliably.
A useful review record contains:
| Field | Purpose |
|---|---|
| Exact excerpt | Shows what is being criticised |
| Issue | Identifies the specific editorial problem |
| Reader impact | Explains why it matters here |
| Proposed edit | Supplies the smallest adequate repair |
| Meaning risk | Flags possible loss of conditions, scope, or voice |
| Missing evidence | Separates research needs from style changes |
Do not automatically publish model rewrites of claims, quotations, figures, or commitments. Compare those elements with the source. Keep a small test set of representative writing so changes to the editing prompt can be checked for lost nuance and an increasingly uniform voice.
Reusable editing prompt
Copy this prompt, fill in the context you know, and append the draft. Missing optional context should lead to conservative editing, not invented facts.
Edit the draft for clarity, substance, and the author's voice.
Context
- Reader: [who will read this]
- Purpose: [what they need to understand, decide, or do]
- Format: [email, report, article, product copy, etc.]
- Voice: [plain description; include authentic samples if available]
- Length: [target or preserve approximately]
- Editing depth: [light copyedit or structural edit]
Process
1. Identify the main point and preserve all supported meaning.
2. Flag factual or reasoning problems separately from style problems.
3. Remove empty introductions, repeated conclusions, generic praise,
chat residue, and claims of importance that add no explanation.
4. Inspect forced contrasts, slogans, rhetorical questions, repeated
three-part constructions, excessive fragments, and decorative formatting.
Keep any instance that serves a clear purpose.
5. Replace vague wording with concrete wording only when the supplied
material supports it. Otherwise retain a restrained statement or flag
the missing detail. Do not invent evidence to improve specificity.
6. Use consistent terminology. Prefer direct verbs and explain useful
relationships. Keep technical terms when they are accurate and needed.
7. Match the tone to the author, reader, and situation. Preserve real
personality and warmth. Do not add slang, deliberate errors, anecdotes,
personal experience, or opinions absent from the source.
8. Preserve numbers, names, dates, definitions, negations, comparisons,
conditions, uncertainty, attribution, and commitments. Do not alter
quoted text. Do not turn association into causation.
9. Compare the edited version with the source for factual and meaning drift.
Do not optimise for an AI-detector score. Do not treat particular words,
em dashes, passive voice, or lists as inherently wrong. Make only changes
that improve this piece for its reader.
Return
A. The clean edited draft, without assistant preamble or closing offer.
B. Up to five material changes and why they help; do not fill a quota.
C. Any unresolved factual questions or missing support, if present.
Draft
[paste here]
For a particularly important passage, first ask for an audit without a rewrite. Require the editor to quote each problem, explain its effect, and propose a minimal correction. This makes it easier to reject changes that flatten a deliberate voice.
Final checklist
- The intended reader and purpose are clear.
- The opening gets to the point at the right pace for the format.
- Each paragraph contributes information, reasoning, useful context, or deliberate effect.
- Broad claims have been specified, supported, qualified, or removed.
- Actions have identifiable actors where relevant and known.
- Repeated contrasts, trios, questions, and sentence patterns have been reviewed.
- Terminology stays consistent.
- Formatting supports reading and navigation.
- Praise, empathy, enthusiasm, and humour fit the writer and situation.
- Assistant commentary and unintended placeholders are gone.
- No facts, experiences, capabilities, deadlines, or commitments were invented.
- Numbers, quotations, citations, conditions, and uncertainty survived the edit accurately.
- The ending adds a useful implication or stops when the work is complete.
- The piece still sounds like its intended author.
Make the company voice available to every draft
Editing a single draft is manageable. The harder problem is keeping every writer and agent grounded in the same product facts, customer language, terminology, evidence, and editorial decisions.
That is one reason company context matters. Generic AI can imitate a style guide, but it does not automatically know which claims your team has approved, how customers describe the problem, or which document is canonical. The GTM Intelligence approach aims to make relevant company knowledge available to AI-assisted work, while Context Factory is a coming-soon project intended to apply reviewed context to content production, not a tool readers can start using today.
For a practical way to assemble that shared context with the tools you already have, read Context drives GTM.
Further reading and scope
The catalogue and examples are practical editorial guidance, not research findings about the statistical frequency of model outputs. Many of these weaknesses long predate generative AI.
The National Archives' plain-language principles offer a concise foundation for reader-focused organisation, direct wording, and useful formatting. National Archives, Top 10 Principles for Plain Language
For the limits and fairness risks of authorship detection, see the original study discussed at the beginning. Its results concern the systems and samples tested in 2023; they should not be presented as a measured error rate for every current detector. Liang et al., GPT detectors are biased against non-native English writers