AI writing anti-patterns · Rules + counter-prompts
AI writing anti-patterns.
Empty prefaces, contrived contrasts and explanations that miss the point. Each rule shows the problem, a possible revision and cases where the text should stay. Its counter-prompt sets out the criteria for that decision.
When AI-style writing obscures the point
A document can contain correct sentences and still miss its task. An introduction explains architecture before the reader can open the app. A heading hedges an established fact. Three paragraphs justify a change instead of describing the current workflow.
The review prompts include context: audience, reader task, supported facts and accepted decisions. An explanation stays when it helps the reader understand a choice, prerequisite or consequence.
Anti-patterns and counter-prompts
Examples make each pattern visible. Open an entry for its reader cost, a possible correction, a counter-prompt and a case where the wording should stay.
Structure and relevance
Taxonomy before the taskOur architecture unifies three review surfaces.
The opening introduces the author’s categories before showing what the reader can do.
Reader cost: Readers must translate the conceptual scheme into their next action.
Pattern example
Our architecture unifies three review surfaces.
Possible revision
Open a folder to review its Markdown files.
The page is a folder-review quickstart and that operation exists.
Can stay as written: This chapter compares three review architectures.
Counter-prompt
Test whether the opening lets this audience locate its task. Suggest a concrete entry only when the stated goal supports it.Review aim: Lead with the reader's task
Recognition, exceptions and sources
The opening explains the author's framework before making the reader's next action or question clear.
Possible cues: long conceptual preface; implementation taxonomy before first action.
Exceptions: A conceptual essay may aim to establish a framework; a reference page may serve lookup rather than onboarding.
- Microsoft Learn style and voice quick start
Organize technical documentation around the customer's task; limit distracting qualifiers and asides. A documentation style policy, not empirical evidence about LLM prevalence or factual truth.
Summary loopsSave your decision. In summary, your choice can be saved. This keeps the decision stored.
The text restates the same conclusion in an introduction, a takeaway and a final summary.
Reader cost: Readers spend attention on repeated information instead of the next useful point.
Pattern example
Save your decision. In summary, your choice can be saved. This keeps the decision stored.
Possible revision
Save your decision.
The repeated sentences add no persistence guarantee or other distinct information.
Can stay as written: Each endpoint lists its request, response and error cases.
Counter-prompt
Compare repeated passages for distinct information. Remove only duplication; preserve parallel structures that make equivalent items easier to compare.Review aim: Make repeated structure earn its place
Recognition, exceptions and sources
Sections repeat the same conclusion, stock transition or fixed template without adding a distinct fact or helping comparison.
Possible cues: repeated wrap-up paragraphs; identical feature-benefit cadence; synonymous summaries.
Exceptions: API references, accessibility cues and parallel comparisons benefit from deliberate consistency. Repetition does not establish machine authorship.
- Use simple words, concise sentences
Prefer precise verbs and consistent terminology; remove words that add no information. Context and meaning still determine whether a modifier or longer explanation is needed.
- Linguistic and Embedding-Based Profiling of Texts Generated by Humans and Large Language Models
Reports aggregate linguistic and semantic variation across eight domains and eleven models. Abstract reviewed. Distributional differences are not per-sentence defects, a detector, or a ranking of current Claude/Codex writing quality.
Formatting on every sentenceQuality: More than a score. Cost: A question of value.
Nearly every sentence becomes a labeled bullet, a heading or an emphasized fragment.
Reader cost: Visual emphasis stops indicating which relationships and decisions matter.
Pattern example
Quality: More than a score. Cost: A question of value.
Possible revision
Review quality and cost
One section compares quality and cost. The question remains useful as a heading for instructions that answer it.
Can stay as written: How do I compare review models?
Counter-prompt
Choose structure for the relationship: sequence, comparison or explanation. Check that informational headings name the section's topic or reader task. Keep genuine reader questions when the section answers them. Distinguish an informational heading from an accepted positioning headline before proposing a change.Review aim: Let structure support reading
Recognition, exceptions and sources
Structure obscures a sequence, comparison or explanation, or an informational heading fails to identify the section's topic or reader task.
Possible cues: one-sentence subsections; nested lists without hierarchy; bold labels on every sentence.
Exceptions: Long references need navigation; procedures need steps; comparison tables can reduce rather than add reading work. A question heading is useful when readers ask that question and the section answers it. An accepted positioning headline can prioritize tone and memorability.
- OpenAI model guidance: Personality and writing style
Explicit control of prose, terminology and repeated rhetorical patterns. Rolling model documentation; not a comparative study or a rule for every OpenAI/Codex output.
- Claude prompting best practices
Positive directions and examples can specify the desired output style. Provider advice, not proof that this catalogue improves a particular model. The document describes different defaults across models.
- Measuring AI ‘Slop’ in Text
Reader judgments depend on relevance, information and structure in the evaluated domains. English news and QA with a small expert panel. Context for review criteria, not proof that these prompts improve models.
- Problem Patterns in AI: Beyond Hallucinations
Practitioner examples distinguish requested information from added interpretation. Editorial observations and an informal poll; no measured failure rates or established training causes.
Optional-detail detoursBefore your first review, read about theme choice, display density and protocol envelopes.
A required workflow is interrupted by settings, internals or background that do not affect the next action.
Reader cost: Readers must decide which details are prerequisites and which can wait.
Pattern example
Before your first review, read about theme choice, display density and protocol envelopes.
Possible revision
Open a file and select a passage. See the reference for display options.
Display options are optional; the bridge is a genuine prerequisite for that live-page workflow.
Can stay as written: Enable the development bridge before reviewing the running page.
Counter-prompt
Keep prerequisites, decisions and consequences beside the relevant action. Move independent optional explanations to a discoverable reference.Review aim: Give detail where it changes the next action
Recognition, exceptions and sources
Optional controls, deep implementation details or tangential background interrupt the shortest complete route to the stated task.
Possible cues: layout settings before first review; authentication protocol before opening the app; unrelated architecture aside.
Exceptions: A prerequisite or material failure mode may be essential before the action; reference pages intentionally contain detail.
- Microsoft Learn style and voice quick start
Organize technical documentation around the customer's task; limit distracting qualifiers and asides. A documentation style policy, not empirical evidence about LLM prevalence or factual truth.
Unprioritized laundry listBefore opening a file, consider login, themes, caching, locales, analytics, rollout and keyboard shortcuts.
The response lists every adjacent concern without separating prerequisites, relevant choices and unrelated possibilities.
Reader cost: The reader inherits the work of deciding what matters and where to start.
Pattern example
Before opening a file, consider login, themes, caching, locales, analytics, rollout and keyboard shortcuts.
Possible revision
Open the registered folder and choose a file. Configure a model route if you need generated alternatives.
The folder is already registered; the listed display and operations topics are not prerequisites for opening its file.
Can stay as written: This reference lists every supported configuration key.
Counter-prompt
Identify the smallest complete set of required actions and relevant choices. Keep optional reference material discoverable without presenting every possibility as a prerequisite.Review aim: Prioritize the decisions required for this task
Recognition, exceptions and sources
An answer to a bounded task becomes an exhaustive list of possible concerns, with no ordering by relevance or dependency.
Possible cues: consider A, B, C and many unrelated topics; flat list of optional and required work.
Exceptions: A requested exhaustive inventory, compliance checklist or API reference needs complete coverage. Preserve actual dependencies and material risks.
- Microsoft Learn style and voice quick start
Organize technical documentation around the customer's task; limit distracting qualifiers and asides. A documentation style policy, not empirical evidence about LLM prevalence or factual truth.
Forced triads and canned Q&AWhy source? Because clarity. Why review? Because confidence. Why Git? Because trust.
Distinct ideas are squeezed into the same three-part slogan or question-and-answer rhythm even when that form adds no meaning.
Reader cost: The repeated performance obscures which facts and relationships actually differ.
Pattern example
Why source? Because clarity. Why review? Because confidence. Why Git? Because trust.
Possible revision
Review decisions retain the source version. Git records changes when you commit the files.
The product behavior is established; the recurring questions are ornamental rather than actual user questions. The separate FAQ answers a real compatibility question.
Can stay as written: Q: Which version is supported? A: Protocol version 1.
Counter-prompt
Check whether each rhetorical unit answers a real question or names a distinct fact. Let the number and form follow the content, while preserving useful parallel comparison.Review aim: Let the content determine its rhetorical structure
Recognition, exceptions and sources
A fixed rhetorical template determines the number and shape of points instead of the actual ideas, or filler is added to complete a triad.
Possible cues: why X? because Y repeated without an actual reader question; a third point added only to complete a triad.
Exceptions: A real FAQ, a mnemonic with three necessary parts, deliberate approved positioning and parallel API entries can justify a recurring form.
- OpenAI model guidance: Personality and writing style
Explicit control of prose, terminology and repeated rhetorical patterns. Rolling model documentation; not a comparative study or a rule for every OpenAI/Codex output.
- Linguistic and Embedding-Based Profiling of Texts Generated by Humans and Large Language Models
Reports aggregate linguistic and semantic variation across eight domains and eleven models. Abstract reviewed. Distributional differences are not per-sentence defects, a detector, or a ranking of current Claude/Codex writing quality.
Claims and evidence
Roadmap inside the usage guideThe export will be able to carry Git decisions.
Instructions for the available product drift into features that are only proposed.
Reader cost: Readers cannot reliably tell which instructions work today.
Pattern example
The export will be able to carry Git decisions.
Possible revision
The export contains the selected findings.
Findings export is verified. The Git extension moves to a separate internal plan; no available behavior is removed.
Can stay as written: The job will run after the current upload completes.
Counter-prompt
Check each capability against supplied implementation evidence. Keep the current workflow clear and place proposals in an explicitly identified plan.Review aim: Describe the available behavior
Recognition, exceptions and sources
A usage guide mixes implemented behavior with proposals, temporary rollout history or hypothetical capabilities.
Possible cues: will support; would allow; in a future release.
Exceptions: Roadmaps, migration guides and release notes intentionally discuss time and change; asynchronous behavior may require future tense.
- Present tense
Describe general implemented behavior in the present; preserve future tense for actual future timing. Grammar alone cannot determine feature maturity.
- Future features
Keep unannounced future capabilities out of ordinary product documentation. Google's internal publication-approval requirement is not adopted as a Voice authorization rule. Explicit plans remain valid document types.
Hedge stackingThe file may generally be considered to remain local.
Several vague qualifiers surround an established fact without identifying an actual unknown or choice.
Reader cost: Readers cannot distinguish genuine uncertainty from habitual caution.
Pattern example
The file may generally be considered to remain local.
Possible revision
The file stays in the registered folder.
Local storage is verified; the separate stale-result condition remains necessary.
Can stay as written: The result may be stale if the source changed outside Studio.
Counter-prompt
Distinguish unknown facts from optional choices. Retain each necessary condition and uncertainty once, attached to the affected claim; do not turn an estimate into certainty.Review aim: Keep uncertainty specific and necessary
Recognition, exceptions and sources
Stacked hedges weaken a verified fact, or a suggested simplification would remove an actual condition, unknown or risk.
Possible cues: generally may perhaps; preferred without a stated alternative; repeated generic disclaimers.
Exceptions: Scientific uncertainty, security conditions, configuration dependence, estimates and genuine alternatives require qualification.
- Microsoft Learn style and voice quick start
Organize technical documentation around the customer's task; limit distracting qualifiers and asides. A documentation style policy, not empirical evidence about LLM prevalence or factual truth.
Invented cause and benefitGit storage therefore makes every review easier to understand.
A feature is treated as proof of an outcome even though the link is missing from the evidence.
Reader cost: Readers may make a decision based on a benefit that was never established.
Pattern example
Git storage therefore makes every review easier to understand.
Possible revision
Git storage records changes to the review files.
File history is supported; universal comprehension improvement has not been established.
Can stay as written: A stale version is rejected because its hash differs from the current source.
Counter-prompt
Trace the stated cause to the claimed outcome. If the link is missing, describe the observed capability or identify the missing evidence without inventing a benefit.Review aim: Show the link between cause and benefit
Recognition, exceptions and sources
A feature is followed by a benefit, conclusion or recommendation that the provided evidence does not establish.
Possible cues: therefore reduces cognitive load; this ensures success; because it is Git-based.
Exceptions: A documented mechanism, bounded measurement or explicitly stated hypothesis can justify a causal explanation.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
Claim without evidenceAll integrations are verified.
A guarantee, measurement, readiness claim or evaluative modifier asserts more than the evidence or stated criteria support.
Reader cost: The apparent certainty hides what was and was not checked.
Pattern example
All integrations are verified.
Possible revision
The TypeScript and JavaScript examples passed the consumer checks.
Those checks ran and support exactly that subset; no broader integration claim follows.
Can stay as written: Both examples passed the two recorded checks.
Counter-prompt
Name the evidence and its actual scope. Flag missing support without declaring the claim false; preserve valid bounded measurements. When an evaluative modifier such as valid, justified or legitimate adds a judgment, check for a stated criterion that helps the reader decide. If it is missing, name the condition or remove the modifier. For example, replace valid exceptions with cases where the text should stay when the surrounding rule supplies those conditions. Keep valid JSON when syntax validity is the actual requirement; distinguish syntax validity from schema validation. Do not ban an adjective based on word-list membership.Review aim: Match the claim to its evidence
Recognition, exceptions and sources
A count, superlative, guarantee, release claim or evaluative modifier exceeds the observed scope, omits its decision criterion, or treats plausible source text as verified product behavior.
Possible cues: all pages verified; production-ready; guaranteed improvement; percentages without conditions; valid exceptions without a stated criterion.
Exceptions: An exact API invariant or properly scoped measurement can support absolute wording. A stated decision criterion can justify a modifier. For example, valid JSON names a syntactic requirement; it does not mean that a schema was checked. Exception criteria may be supplied by the adjacent rule instead of repeated in every heading.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
- Use simple words, concise sentences
Prefer precise verbs and consistent terminology; remove words that add no information. Context and meaning still determine whether a modifier or longer explanation is needed.
Unsupported completion claimThe website is live and fully verified.
A report announces completion or readiness without a matching observation for that exact result.
Reader cost: Readers may treat an unverified step as completed and stop checking too early.
Pattern example
The website is live and fully verified.
Possible revision
The local build passed. The public deployment has not been checked.
Only the local build was observed; publication must have its own evidence.
Can stay as written: The deployment URL returned the recorded build identifier.
Counter-prompt
Tie each result claim to the recorded observation. State failures and missing coverage directly; distinguish a source push, local build and public deployment.Review aim: Report the observed result and its boundary
Recognition, exceptions and sources
A progress or completion claim lacks a matching tool result, or success in one check is presented as broader readiness.
Possible cues: done without result evidence; tests passed used as editorial approval; source push described as deployment.
Exceptions: A brief result can refer to an accessible report rather than repeat every log line. Verified results do not need speculative hedging.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
Unearned praise and automatic agreementBrilliant idea! Your review design is unquestionably the best approach.
The response praises the user or endorses a proposal before examining the claim, evidence or tradeoff.
Reader cost: Approval can replace the useful critique the reader asked for.
Pattern example
Brilliant idea! Your review design is unquestionably the best approach.
Possible revision
The design keeps decisions beside their source. It still needs a rule for stale results.
The reviewed design has source-linked decisions but no stale-result policy; the assessment concerns the design, not the author.
Can stay as written: The new diagram is easier to follow because each arrow names an operation.
Counter-prompt
Assess the proposal against its goal and evidence. Give specific praise where it is earned and state the unresolved weakness without judging the person.Review aim: Evaluate the proposal on its merits
Recognition, exceptions and sources
A request for assessment receives flattering approval without a specific reason, or the answer changes its factual position only to agree with the user.
Possible cues: brilliant idea without assessment; you are absolutely right before checking; reflexive praise before critique.
Exceptions: A supported compliment, ordinary warmth and an explicit celebration are valid. Agreement is appropriate when the evidence supports it.
- Sycophancy in GPT-4o: what happened and what we’re doing about it
Documents a specific GPT-4o update that became excessively agreeable and was rolled back. A historical provider incident, not proof that current OpenAI, Codex or Claude output is generally sycophantic.
- OpenAI Model Spec, 18 December 2025
Describes evidence-proportionate treatment of viewpoints and constructive critique rather than reflexive agreement. A statement of intended behavior, not a measured account of model compliance or a mandate to choose a side when evidence is missing.
Forced neutrality and false balanceApplying the stale diff and refreshing the source are equally reasonable options.
The answer gives alternatives equal weight even though the supplied facts rule one out or support them unequally.
Reader cost: Readers receive an artificial dilemma instead of the supported distinction.
Pattern example
Applying the stale diff and refreshing the source are equally reasonable options.
Possible revision
Refresh the source before preparing the change; the saved diff refers to an earlier version.
The saved source version is known to be stale; the backend does not support applying that stale proposal.
Can stay as written: Both storage adapters meet the requirements. Choose according to the deployment constraints.
Counter-prompt
Represent the alternatives fairly and weight them by the available evidence and stated criteria. Name a real unresolved tradeoff when one remains; do not manufacture certainty.Review aim: Weight alternatives by the evidence
Recognition, exceptions and sources
An evidence-based question is answered with interchangeable pros and cons or a generic “both are valid” despite a material difference already established in context.
Possible cues: both options are equally valid despite contrary evidence; it depends without naming the dependency.
Exceptions: Genuine uncertainty, value tradeoffs and comparably supported alternatives require balanced treatment. A request to explain a viewpoint is not an endorsement request.
- OpenAI Model Spec, 18 December 2025
Describes evidence-proportionate treatment of viewpoints and constructive critique rather than reflexive agreement. A statement of intended behavior, not a measured account of model compliance or a mandate to choose a side when evidence is missing.
Words and sentence patterns
Mechanism without an operationGit is the preferred source of record.
Abstract labels such as “source of record” replace the file, actor or operation the reader needs to identify.
Reader cost: Readers still do not know what is stored, where it is stored or who changes it.
Pattern example
Git is the preferred source of record.
Possible revision
Commit the approved decisions as JSON files in the repository.
This is an explicit workflow instruction with an established JSON layout, not a claim that Studio commits automatically.
Can stay as written: Git is the preferred option when both supported storage adapters are available.
Counter-prompt
Ask which object, location and operation the sentence means. Prefer a verified file, field or action over an unexplained abstraction.Review aim: Name the object and operation
Recognition, exceptions and sources
An abstract label leaves the reader unable to identify what data exists, where it lives or what operation happens.
Possible cues: source of record; mechanism; ecosystem alignment; orchestration surface.
Exceptions: A defined domain term is useful in a specialist reference; a concept may require explanation before implementation detail.
- Avoid jargon
Specialist shorthand can help a specialist audience and obstruct other readers. Does not justify banning technical terms or identifying the author of a sentence.
Invented compound labelsInspect the exact-head evidence-ready review slice.
Several concepts are compressed into a new label that the product never defines.
Reader cost: Readers decode terminology rather than understand the operation.
Pattern example
Inspect the exact-head evidence-ready review slice.
Possible revision
Inspect the review result for this commit.
The compound has no documented meaning beyond the commit's review result.
Can stay as written: The sourceVersion field contains the file hash.
Counter-prompt
Reuse the project's established terminology. Define a new term only if it names a distinct concept the reader needs.Review aim: Use names the reader can resolve
Recognition, exceptions and sources
New compound labels, unexplained abbreviations or synonyms force the reader to infer concepts that already have stable names.
Possible cues: hyphen-stacked labels; multiple names for one field; private working shorthand.
Exceptions: Exact API identifiers, established specialist terms and intentionally defined concepts must remain stable.
- Avoid jargon
Specialist shorthand can help a specialist audience and obstruct other readers. Does not justify banning technical terms or identifying the author of a sentence.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
Not X, but YThis is not just feedback. It is a decision layer.
A dramatic opposition rejects an alternative that was never relevant to the reader’s task.
Reader cost: The invented alternative adds a distinction the reader did not need.
Pattern example
This is not just feedback. It is a decision layer.
Possible revision
Keep a passage or propose a replacement.
The general product introduction does not require the invented category contrast.
Can stay as written: A source hash identifies file content, not a Git commit.
Counter-prompt
Check whether both sides of a contrast answer a real reader question. If not, describe the supported behavior directly.Review aim: Keep comparisons relevant
Recognition, exceptions and sources
A rhetorical opposition introduces an alternative the audience did not need to distinguish, instead of explaining the actual behavior.
Possible cues: not X, but Y; this is not about; rather than merely.
Exceptions: Comparing version semantics, correcting a likely misconception or preserving approved positioning can require a contrast.
- OpenAI model guidance: Personality and writing style
Explicit control of prose, terminology and repeated rhetorical patterns. Rolling model documentation; not a comparative study or a rule for every OpenAI/Codex output.
Arrow-chain shorthandHEAD→hash→stale→reject; UI ok.
A result is compressed into arrows, identifiers and fragments whose relationships remain implicit.
Reader cost: Readers who missed the working session have to guess actors, conditions and outcomes.
Pattern example
HEAD→hash→stale→reject; UI ok.
Possible revision
The backend rejected the changed source version. The UI displayed the conflict.
A result report requires the actor and observed response; the short diagram has a defined purpose.
Can stay as written: Request → result → decision
Counter-prompt
Restore the missing relationship in ordinary sentences. Keep code and identifiers exact; shorten by selecting relevant facts rather than deleting grammar.Review aim: Explain relationships without private shorthand
Recognition, exceptions and sources
Compression removes the subject, condition or relationship needed by a reader who did not follow the working session.
Possible cues: dense arrow chains; unexplained identifier sequences; fragment-only result reports.
Exceptions: Diagrams, table cells, button labels and familiar notation can be concise fragments.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
Ornamental wordingWhich model and prompt earn their cost?
Decorative phrases obscure a subject or operation that readers could identify directly.
Reader cost: Readers interpret the flourish before finding the subject or action.
Pattern example
Which model and prompt earn their cost?
Possible revision
Model selection and prompt structure
The informational heading introduces a comparison without established results. The cache metaphor is useful when explaining response reuse to this audience.
Can stay as written: A cache works like a short-term memory for responses.
Counter-prompt
Name the supported subject or operation in familiar terms. In informational headings, compare an ornamental phrase with a concrete description and explain any loss of orientation. Keep precise technical terms, explanatory metaphors and accepted positioning language when they serve the audience; flag no phrase solely for its style.Review aim: Prefer a precise operation to ornamental wording
Recognition, exceptions and sources
Ornamental wording obscures the subject or operation, or asks readers to interpret a metaphor without useful explanatory detail.
Possible cues: facilitate the execution of; robust seamless workflow; powerful mechanism.
Exceptions: A defined technical adjective or an approved quotation can be exact and useful; word-list membership is not a defect. An explanatory metaphor can make a mechanism easier to understand; accepted positioning may deliberately use rhythm or figurative wording.
- Use simple words, concise sentences
Prefer precise verbs and consistent terminology; remove words that add no information. Context and meaning still determine whether a modifier or longer explanation is needed.
- Can AI writing be salvaged? Mitigating Idiosyncrasies and Improving Human-AI Alignment in the Writing Process through Edits
Distinguishes distracting ornament, vague detail and awkward phrasing in expert edits. Creative-writing study. Applying its categories to informational headings is an editorial derivation, not a measured heading detector.
Narration and working history
Development diaryThe original readiness folder mixed several different things. Only some belong to the product. The maintained structure gives each kind an explicit owner and retention policy.
A public guide retells earlier drafts, cleanup work or feedback rounds instead of describing the current files and behavior.
Reader cost: Readers reconstruct development history to find the current answer.
Pattern example
The original readiness folder mixed several different things. Only some belong to the product. The maintained structure gives each kind an explicit owner and retention policy.
Possible revision
Product code lives in frontend/, backend/ and packages/. Reusable commands live in scripts/. Dated verification reports live in docs/verification/.
The public guide needs the current file locations. The earlier folder cleanup does not affect the reader’s task. The before passage is the user-provided Voice documentation example, not a provider output capture.
Can stay as written: Version 2 renamed the field; existing integrations must migrate.
Counter-prompt
Explain what readers can use now. Retain earlier choices only when they explain a relevant compatibility constraint or design tradeoff.Review aim: Separate the result from working history
Recognition, exceptions and sources
The deliverable recounts abandoned approaches, internal filenames or conversation turns that do not explain the present result.
Possible cues: we initially considered; after your earlier feedback; during this session.
Exceptions: Decision records, incident reports, changelogs and research notebooks need history and provenance.
- Prompting Claude Fable 5
Connect status to observed results and explain outcomes without private working shorthand. Guidance for a named model and agentic setting; not a general claim about all Claude versions.
- Microsoft Learn style and voice quick start
Organize technical documentation around the customer's task; limit distracting qualifiers and asides. A documentation style policy, not empirical evidence about LLM prevalence or factual truth.
Unnecessary interpretive framingModel choice: why the smartest route changes everything.
A heading or introduction adds significance, interpretation or a result promise that the supplied material does not establish.
Reader cost: Readers must separate the actual information from added interpretation.
Pattern example
Model choice: why the smartest route changes everything.
Possible revision
Choose a model for the review.
The section gives model-selection steps; the supplied material does not support a wider outcome claim.
Can stay as written: This reference covers protocol version 1.
Counter-prompt
State the useful fact or task directly. Compare headings and introductions with the supplied brief: remove invented interpretive premises and result promises before evidence is available. Retain necessary scope, supported conclusions and accepted positioning; a framing choice requires a specific reader-impact reason to change.Review aim: Replace empty narration with information
Recognition, exceptions and sources
Framing adds an interpretive premise or result promise unsupported by the supplied material, or announces importance without useful scope.
Possible cues: it is worth noting; this section clearly explains; the key thing to understand.
Exceptions: A short scope sentence can distinguish a protocol reference from a tutorial or identify excluded versions. A result heading can state a supported conclusion. An accepted positioning headline is not automatically an unsupported factual claim.
- OpenAI model guidance: Personality and writing style
Explicit control of prose, terminology and repeated rhetorical patterns. Rolling model documentation; not a comparative study or a rule for every OpenAI/Codex output.
- Problem Patterns in AI: Beyond Hallucinations
Practitioner examples distinguish requested information from added interpretation. Editorial observations and an informal poll; no measured failure rates or established training causes.
- Measuring AI ‘Slop’ in Text
Reader judgments depend on relevance, information and structure in the evaluated domains. English news and QA with a small expert panel. Context for review criteria, not proof that these prompts improve models.
Review several rules in one prompt
Individual counter-prompts target a particular problem. For a document, the Library combines several rules in one prompt. The same source context does not have to be repeated for every rule.
Our local comparison uses 30 English/German cases and measures input volume at equal rule coverage. Detection quality, actual model tokens and cost need separate evaluation.
Measurements, model candidates and token economics ↗Combine rules for the task
Four profiles select six rules each. Add the audience, goal, facts and accepted decisions. The composed prompt requests exact quotes, a short reason and up to three alternatives; keeping the wording remains an option.
See the profiles and rule selections
- Public documentationTaxonomy before the task · Roadmap inside the usage guide · Development diary · Unnecessary interpretive framing · Hedge stacking · Optional-detail detours
- Technical referenceRoadmap inside the usage guide · Mechanism without an operation · Invented compound labels · Hedge stacking · Formatting on every sentence · Ornamental wording
- Product explanationTaxonomy before the task · Mechanism without an operation · Not X, but Y · Invented cause and benefit · Claim without evidence · Summary loops
- Agent result reportDevelopment diary · Unnecessary interpretive framing · Hedge stacking · Arrow-chain shorthand · Optional-detail detours · Unsupported completion claim
Claude, Codex and other models
The catalogue works independently of the model. Providers document controls for length, structure and word choice; that guidance is not a measured failure rate for a model. Studies compare linguistic features across text collections. The reader’s task determines whether a particular expression is a problem.
Each rule identifies its source and what that source supports. Examples are editorial illustrations, not unedited model responses. Prompt effectiveness has not been measured in a model comparison.