Skip to main content

“Don't do that”, “VERY IMPORTANT”, “Can you…?”: what the evidence says about negation, capitals and questions in prompts

Few studies test two phrasings of the same instruction inside an actual agent prompt. We reviewed the 2026 vendor guides, the papers and the field reports on negation, capital letters, repetition and the question form, and kept only what they state: where the evidence is strong, where it is thin, and which circulating figures we discarded.

  • 12 min read
  • 23 sources
  • 3 diagrams
  • Paul-Antoine Tual
Lire en français
Contents

Guidance on crafting prompts usually relies on vendor manuals and limited trials. Rarely do these sources compare two versions of the same command within an actual agent prompt. We reviewed 2026 vendor guides, arXiv papers, and the field reports we could open on 6 November 2026. We retained only their explicit claims.

We tag each source as vendor documentation (unmeasured), measurement (studies with data and limits), or field report (blog posts or forum threads). The outcome is a collection of trends, not fixed rules.

  • Negation: brief avoidance rules work, but wide bans are applied too often, and mentioning the banned item can keep it active.
  • Capitals and repetition: emphatic wording causes recent Claude models to over-respond, and the only repetition with a measured gain is a brief restatement of a format rule at the end.
  • Questions: commands yield actions where queries yield advice, yet no evidence shows the question form makes an agent ask questions back.

Negation: simple avoidance rules hold, broad prohibitions misfire

A prohibition is not one thing. Short rules that forbid a word or a punctuation mark are held well, while broad bans, bans on a way of reasoning and rules that name the forbidden item tend to backfire. The vendors that give advice prefer the positive form, but they have not measured it.

  • Vendor guidance: Anthropic’s guide has a recommendation titled “Tell Claude what to do instead of what not to do”. The text suggests asking for smoothly flowing prose rather than using the phrase “Do not use markdown in your response” [1]. The page lacks a date, and no data supports this claim.
  • Names trigger associations: A study of 40,000 samples found that 87.5% of negative constraint failures were due to priming. Naming the banned word activated its mental representation instead of hiding it [2]. This preprint is from January 2026, which falls outside our six-month review period, but we use it to explain the mechanism.
  • Negated questions cause errors: Accuracy on questions with negation ranged from 45.2 to 57.6% across six open models. This compares to 91.8 to 96.5% for positive versions. Turning off specific late-layer attention modules improved negated accuracy by about 17 points for Llama-3.1-8B [3]. The test checks explicit negation in questions, not rules inside a prompt.
  • Polarity changes shift outcomes: In 14 moral dilemmas, small open models of 1 to 4 billion parameters supported an action 24% of the time when it was stated affirmatively. Support rose to as much as 100% when the action was negated. This created a swing of up to 76 points across 16 models. Agreement between models dropped from 73% to 59% under simple negation [4]. This measures how polarity affects ethical choices, not the impact of a prompt rule.
  • Simple avoidance works under load: When three models handled a concurrent task, avoidance rules like no commas or forbidden words stayed at or near 100% compliance. Rules placed at the end of the response dropped by up to 50% [5]. The sample size is six items per constraint type, so the result is only indicative.
  • Prohibitions fade in long chats: For Mistral Large 3, compliance with prohibition rules fell from 73% at turn 5 to 33% at turn 16. Rules requiring an action stayed at 100% compliance. These results come from 4,416 trials on 12 models (see Figure 1) [6]. This seems to conflict with the previous finding, but the pressure differs: one case uses a concurrent task, while the other uses conversation length. This preprint is dated 22 April 2026, slightly beyond our window. It is the only measurement we found of rule decay over a long exchange.
  • Judgments and guardrails: In a field test with three code samples, a look-alike false problem was flagged in 2 of 3 examples with no instruction. It was also flagged in 2 of 3 examples with an explicit list of what not to flag. By contrast, a positive inclusion criterion flagged 0 of 3 [7]. On a production code-review prompt, negative and positive phrasing found practically the same bugs. The model name is not given, and the sample is very small.
Compliance with prohibition rules and action rules over a long conversation Two lines for Mistral Large 3 between turn 5 and turn 16, compliance in percent. Prohibition rules fall from 73 to 33. Rules that require an action stay at 100. Only two points are reported per series, so the straight lines are a guide. Prohibitions decay in a long conversation, action rules do not Compliance by conversation turn, Mistral Large 3, in % 0 25 50 75 100 turn 5 turn 16 100% 100% 73% 33% Rules that require an action Prohibition rules (do not do this) Two reported points per series; the straight lines are a guide, not a fitted curve.
Figure 1. For Mistral Large 3, compliance with prohibition rules fell from 73% to 33% between turn 5 and turn 16, while rules that require an action held at 100%. Source: Gamage, arXiv:2604.20911, 22 April 2026, abstract (4,416 trials, 12 models).

Specific bans work well. Wide-ranging rules often fail.

Capitals and repetition: loud rules over-trigger, a closing reminder helps

Strong language causes newer models to react too often. Repeating the entire prompt mainly aids models that lack reasoning ability. The only repetition with a measured benefit is a brief reminder of a format constraint placed at the end of the prompt.

  • Anthropic: The Claude Opus 4.5 and 4.6 models react more sharply to system prompts. A rule designed to stop under-triggering may now lead to over-triggering. The guide advises swapping “CRITICAL: You MUST use this tool when…” for “Use this tool when…” [1].
  • OpenAI: The GPT-5.6 guide states “Keep the policy in one place and state each rule once”. It warns that repeating commands like “ask first”, “do not mutate” or “wait for approval” can spark unnecessary approval requests. The guide reserves ALWAYS, NEVER and must for true invariants [8].
  • Capitals, measured: A study of 13 models found that placing a target word in capitals within lowercase text focused attention on it. Accuracy on factual questions increased by up to 8.95 points with the best methods. However, the discovery model, Qwen2.5-7B-Instruct, saw only a 1.85 point rise [9]. Alternating case dropped scores by up to 13.96 points on LLaMA-3.1-8B-Instruct. Reasoning models usually varied by less than 0.5 point (see Figure 2). These tasks involved questions with precise answers, not behavioural instructions. Therefore, this result does not conflict with Anthropic.
  • Repeating the prompt: Repeating the input won 47 of 70 benchmark-model combinations. There was no loss when the model did not reason. With step-by-step reasoning, the effect was neutral to slightly positive (5 wins, 1 loss and 22 ties out of 28) [10]. The tests used Gemini 2.0 Flash, GPT-4o, Claude 3.7 Sonnet and DeepSeek V3 in February and March 2025. The preprint dates from December 2025. We flag this as outside our window.
  • A closing reminder: A format combining explicit framing with a trailing reminder restored compliance to 90 to 100% in many settings. Average gains were 17.7 points for o4-mini, 16.9 for DeepSeek-V3.1 and 12.7 for Llama-3.3-70B [5]. This approach changes several elements at once. Thus, the reminder alone is not isolated.
  • Instruction files: Across about 1,650 Claude Code sessions on two TypeScript codebases, file size showed no detectable effect, with statistical support for the absence of an effect. Position and structure had no detectable effect either, but without proof of absence. The odds of compliance dropped by about 5.6% for each extra function generated in a session. The author identified this finding after the fact [11]. Emphasis was not a variable. A separate study found that context files did not generally improve task success. They raised inference cost by over 20% on average [12].
Effect of capital letters on accuracy, in points Horizontal bars around a zero line. Capitalised target word, best scheme: plus 8.95 points. Capitalised target word on Qwen2.5-7B-Instruct: plus 1.85. Alternating case on LLaMA-3.1-8B-Instruct: minus 13.96. Reasoning models: typically under 0.5 point either way. Capitals can help or hurt, and reasoning models barely react Change in accuracy, in points, on questions with a factual answer -15 -10 -5 0 5 10 Capitalised target word, best scheme (up to) +8.95 Capitalised target word, Qwen2.5-7B-Instruct +1.85 Alternating case, worst model (LLaMA-3.1-8B-Instruct) -13.96 Reasoning models, typical variation under 0.5 either way Tasks: MMLU-Pro, ARC-Challenge, SQuADv2. Factual questions, not behavioural instructions.
Figure 2. Capitalising the target word raised accuracy by up to 8.95 points, alternating case cost up to 13.96 points, and reasoning models barely moved. Source: Dillitzer et al., arXiv:2608.03711, 4 August 2026, 13 models.

Repeat a constraint once, at the end, and nothing more.

Question or order: the imperative gets action, but nothing shows the question makes the agent ask

An imperative gets the action where a question can get advice, which Anthropic and OpenAI both document. No study supports the idea that a question form makes an agent ask questions back, and the interruptions we found documented have other causes.

  • Documented: Anthropic contrasts “Can you suggest some changes to improve this function?”, which may get suggestions only, with “Change this function to improve its performance.”, which gets the change, and gives two system-prompt blocks, one to act by default and one to hold back until clearly instructed [1]. It also suggests confirming before hard-to-reverse actions such as a force push.
  • Same direction at OpenAI: the GPT-6 Astra guide says to treat “can you…”, “I want to…” and “help me…” as instructions to do the work [13]. The GPT-5.6 guide sets three behaviours: inspect and report for requests to explain or plan, make in-scope local changes without asking, and require confirmation for external writes, destructive actions, purchases or a material expansion of scope [8].
  • Cause 1, the model: OpenAI says Astra is more likely to ask for clarification and can stop when the user expects reasonable assumptions, and advises completing the work already authorised before asking [13]. A trade-press report of 5 September 2026 relays that Astra asks more often than GPT-5.6 Sol, with no figure [14].
  • Cause 2, the tool and its mode: in a Cursor forum thread of 10 and 11 August 2026, a user reports question dialogs on every turn that the /nohuman command did not stop. Staff replied that there is no switch to turn them off, that a user rule forbidding the question tool usually reduces them but not to zero, and that pinning a model other than Auto may help [15].
  • Cause 3, the question policy: on a benchmark of ambiguous data-science tasks, permissive prompts made agents over-ask on clear tasks and conservative prompts made them default silently on ambiguous ones [16]. Under the conservative policy, Gemini 3.1 Pro asked on 94% of ambiguous target tasks and on 0% of fully specified ones, while GPT-5.4 silently defaulted on 100% of ambiguous metric runs and Claude Haiku 4.5 on 97%. In 88% of Haiku 4.5’s questions on target tasks, it was asking about style preferences.
  • Asking can pay: asking for one to three questions first cost 1.96 turns instead of 1.00 and scored 6.67 out of 8, against 5.67 for a raw prompt and 7.50 for a checklist, and for Claude the clarifying prompt (7.38) was just ahead of the checklist (7.25) [17]. The scores were given by the authors, with four task categories and no blind raters. Another study found that capping questions at one to three cut user-attention cost by 36 to 73% while keeping 86 to 94% of the reward, across eight models and four customer-service domains [18].
  • The question form on the user’s side: asking a model to turn statements into questions before answering reduced sycophancy more than a baseline instruction not to be sycophantic [19]. The abstract gives no figure, and it says nothing about how often an agent asks.
What makes an agent stop and ask the user a question Three documented causes point to a box where the agent stops to ask, which leads to an interruption for the user: the model, the tool and its mode, and the question policy. A fourth box, the wording of the request as a question or an order, is linked by a dashed arrow labelled no controlled test. Three documented causes of questions, one untested The model GPT-6 Astra asks more often than GPT-5.6 Sol (vendor guide) The tool and its mode Cursor: no off switch, /nohuman did not stop the dialogs The question policy Permissive: over-asks on clear tasks Conservative: silent defaults The agent stops to ask the user Interruption for the user Wording of the request question or order no controlled test
Figure 3. Three causes of agent questions are documented; the wording of the request, the dashed link, has not been tested. Conceptual diagram built from sources 13 to 16.

We found no study that compares a question, an order and an indirect request on the rate of questions asked by coding agents, and none in French.

What interpretability adds, and what it does not

Interpretability work gives a plausible mechanism for why naming a forbidden thing keeps it partly active, with imperfect control and sensitivity to wording, but it offers nothing direct on emphasis, instruction repetition or clarification requests.

  • Anthropic, July 2026: when Sonnet 4.5 was told to focus on a specific concept while copying an unrelated sentence, a command to ignore it left the concept much weaker than the focus command, but not gone. The ignore command itself triggered some of the concept [20]. A simple mention could prime the concept almost as well as a direct focus order. The phrase “Don’t think about X” suppressed the concept less than the “ignore” phrasing used in the main tests. The task is a copying task, not a realistic prompt.
  • A preprint of May 2026: on Llama-3.1-8B, with Mistral-7B and Gemma-7B-IT as cross-checks, the banned idea remained recoverable from hidden states under a suppression order. Explicit leakage in the output dropped from 48.3% after a direct mention to zero [21]. The authors do not confirm a strict white-bear effect: their ordering is direct mention above suppression, and suppression far above no mention. The analyses are correlational.
  • The negation mechanism: late-layer attention that favours simple shortcuts, as in the six-model study above [3].
  • What is missing: we found no internal feature for emphasis, no work on repeating an instruction in a prompt, and no mechanism for asking a question, in Claude or elsewhere. Most of the work uses small open models.

Our inference, which is not a published result, is that naming a thing puts it in the context and partly activates it. That makes the vendors’ preference for the positive form plausible, but it does not explain by itself the failures seen in real prompts.

Eleven practices and how far each is supported

Vendor guides and source overlaps form the basis for these practices, not controlled trials. Therefore, every row lists its foundation and relative strength.

PracticeBasisStrength
Describe the wanted behaviour firstAnthropic guide [1]Vendor guidance, not measured
If you must forbid, give the alternative and the reasonAnthropic guide [1]Vendor guidance, not measured
For a judgment or classification, write a positive inclusion criterion instead of a list of exclusionsField test [7]Field report, three examples
For an action guardrail, keep the negative form and add the allowed scope[5], [7]Measurement and field report, small
Avoid capitals and CRITICAL or MUST on recent Claude models[1], [8]Vendor guidance, not measured
State each rule once, in one placeOpenAI guide [8]Vendor guidance, not measured
Restate a critical format constraint at the endTrailing reminder [5]Measurement, three models, several changes at once
Do not count on repeating the prompt with a reasoning modelPrompt repetition [10]Measurement on 2025 models
Use the imperative to get an action[1], [13]Vendor guidance, not measured
Say which kinds of ambiguity justify a question, and when to askOur inference from [13] and [16]Inference
If a tool’s mode generates the questions, change the mode or the modelCursor thread [15]Field report

Two further cautions. Do not deduplicate a prompt file that works without comparing before and after on a test set, because file size showed no detectable effect in one study [11]. And do not tell an agent never to ask, since a conservative question policy pushed some agents to default silently on ambiguous tasks [16].

Figures in circulation that we did not keep

We verified the data we received against the original papers. We removed any claim that the source did not state or that the source contradicts. If we read the full text to get a number, we say so. We checked the other numbers against the abstract.

  • “77%” for the polarity swing in the moral-dilemma study: the abstract gives 24%, up to 100% and a swing of up to 76 points.
  • A drop to 20% at turn 25, with content-imposing rules above 93%, for the rule-decay study: the abstract gives 33% at turn 16 and 100%.
  • A rise from 80% to 95% in safe tool calls for Claude Sonnet 4.5 with a safety suffix, attributed to a tool-call safety benchmark: the abstract of that paper does not contain it.
  • About 3% lower success for model-written context files and about 4% higher for human-written ones: the abstract of the study says neither type generally improves success, with a cost increase above 20%.
  • Thresholds of prohibition density (10, 25, 40 or 60%) in several blog posts: none comes with a method, so we do not use them.

Run the three-wording test before you rewrite a prompt library

A test that would settle the open questions is simple to set up: the same task in three wordings, an imperative, a question and an indirect request, on clear and on ambiguous tasks, with at least two models.

  • Count: the questions asked, whether the task was completed, and whether the result was correct.
  • Control: the same length and the same politeness in the three wordings, and the same system prompt.
  • Extend: the same protocol applies to a prohibition and its positive equivalent, and to a rule with and without capitals.

Our training keywords are tested the same way [22], and a rule that matters is better enforced outside the prompt than worded harder [23]. Until the studies exist, treat every practice above as a tendency to check on your own prompts.

Sources

  1. Anthropic, Prompting best practices, Claude Developer Platform documentation, undated page, read on 6 November 2026.
  2. Shailesh Rana, ‘Semantic Gravity Wells: Why Negative Constraints Backfire’, arXiv:2601.08070, 12 January 2026. Outside the six-month window; cited for the mechanism.
  3. Zhejian Zhou, Tianyi Zhou, Robin Jia and Jonathan May, ‘How Language Models Process Negation’, arXiv:2605.03052, 4 May 2026, revised 29 May 2026.
  4. Katherine Elkins and Jon Chun, ‘Framing Instability in LLM Ethical Stance: Auditing Negation Sensitivity in Moral Dilemmas’, arXiv:2601.21433, 29 January 2026, revised 8 July 2026.
  5. Avni Mittal, ‘Did You Forget What I Asked? Prospective Memory Failures in Large Language Models’, arXiv:2603.23530, 7 March 2026, revised 11 June 2026.
  6. Yeran Gamage, ‘Omission Constraints Decay While Commission Constraints Persist in Long-Context LLM Agents’, arXiv:2604.20911, 22 April 2026. Slightly outside the six-month window; the only measurement of rule decay over a long exchange that we found.
  7. Anastasiia Mudra, ‘Telling AI Agents What Not to Do: Does It Actually Help?’, Mavka blog, 7 August 2026.
  8. OpenAI, Prompt guidance for GPT-5.6, API documentation, undated page, read on 6 November 2026.
  9. Maximilian Dillitzer, Tin Stribor Sohn, Jason J. Corso and Michael Auerbach, ‘Attention is Case-Sensitive’, arXiv:2608.03711, 4 August 2026 (ECCV 2026).
  10. Yaniv Leviathan, Matan Kalman and Yossi Matias, ‘Prompt Repetition Improves Non-Reasoning LLMs’, arXiv:2512.14982, 17 December 2025. Outside the six-month window; tests run on 2025 models.
  11. Damon McMillan, ‘Instruction Adherence in Coding Agent Configuration Files: A Factorial Study of Four File-Structure Variables’, arXiv:2605.10039, 11 May 2026.
  12. Thibaud Gloaguen, Niels Mündler-Sasahara, Mark Niklas Müller, Veselin Raychev and Martin Vechev, ‘Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?’, arXiv:2602.11988, 12 February 2026, revised 29 September 2026.
  13. OpenAI, GPT-6 Astra, API documentation, undated page, read on 6 November 2026.
  14. The Decoder, ‘OpenAI shares prompting tips for GPT-6 Astra, including a blocklist of slop words’, 5 September 2026.
  15. Cursor Community Forum, ‘Constant noop AskQuestions dialogs interrupt workflow’, thread of 10 and 11 August 2026.
  16. Josefa Lia Stoisser, Marc Boubnovski Martell, Sidsel Boldsen, Kaspar Märtens and Robert Kitchen, ‘Ambig-DS: A Benchmark for Task-Framing Ambiguity in Data-Science Agents’, arXiv:2605.09698, 10 May 2026.
  17. Saurav Ghosh, ‘Less Back-and-Forth: A Comparative Study of Structured Prompting’, arXiv:2605.20149, 19 May 2026, revised 4 October 2026.
  18. Xuanming Zhang and colleagues, ‘Ask Less, Ask Better: Allocating User Attention in Proactive LLM Agents with Interruption Budgets’, arXiv:2610.24411, 17 October 2026.
  19. Magda Dubois, Cozmin Ududec, Christopher Summerfield and Lennart Luettgau, ‘Ask don’t tell: Reducing sycophancy in large language models’, arXiv:2602.23971, 27 February 2026, revised 29 July 2026.
  20. Gurnee, Sofroniew, Lindsey and colleagues, Anthropic, ‘Verbalizable Representations Form a Global Workspace in Language Models’, 6 July 2026.
  21. Rebecca Ramnauth and Brian Scassellati, ‘The Attentional White Bear Effect in Transformer Language Models’, arXiv:2605.28639, 27 May 2026 (under review at EMNLP 2026).
  22. Junyr, Keywords for AI agents: what the evidence supports, and the ten we teach, 13 October 2026.
  23. Junyr, Harness engineering: the software layer that turns a model into an agent, 20 September 2026.

We read the arXiv abstracts, and the full texts where a figure was not in the abstract, on 6 November 2026. The vendor guides are undated pages that will change. Three preprints (2, 6 and 10) fall outside our six-month freshness window and are cited for a mechanism or for the absence of a better measurement. Quotations from the guides are short extracts, and the French mirror translates them. Figures 1 and 2 reproduce numbers from the cited papers, and Figure 3 is a conceptual diagram. The wording tests proposed in the last section are our recommendation and have not been run.

Keywords
prompt engineering
negation
capital letters
prompt repetition
clarifying questions
AI agents
instruction following
interpretability
Claude
GPT-6 Astra
Share on LinkedIn Markdown version

Frequently asked questions

Is it a mistake to tell an AI model what not to do?

Not always. Short avoidance rules such as 'no commas' held at or near 100% under task load in one study, but broad prohibitions are over-applied, and naming a forbidden word can prime it. Anthropic advises saying what to do instead.

  • In a 40,000-sample study, 87.5% of violations of a negative constraint were priming failures, where naming the word activated it.
  • Across six open models, accuracy on negated questions was 45.2 to 57.6%, against 91.8 to 96.5% on positive ones.
  • In a small field test, a positive inclusion criterion removed a false-positive error that an explicit prohibition did not.
Do capital letters and words like CRITICAL make a model follow a rule better?

On recent Claude models, Anthropic says they cause over-triggering and advises plain wording. A 13-model study found capitals raised accuracy on factual questions by up to 8.95 points, but no study isolates capitals on behavioural rules.

  • Alternating case cost up to 13.96 points on LLaMA-3.1-8B.
  • Reasoning models typically varied by less than 0.5 point.
  • OpenAI advises reserving ALWAYS and NEVER for true invariants.
Does repeating an instruction help?

Repeating the whole prompt helped models that do not reason, in a study run on 2025 models. A short reminder of a format constraint at the end of the prompt restored 90 to 100% compliance in many settings in another. OpenAI advises stating each rule once.

  • Prompt repetition won 47 of 70 tests with no loss when reasoning was off, and was neutral to slightly positive when it was on.
  • The end-of-prompt reminder changed several things at once, so its effect alone is not isolated.
  • Repeating 'ask first' can trigger needless approval requests.
Does 'Can you…?' make an AI agent ask more questions?

No study shows it. 'Can you suggest…' can get advice where an imperative gets the action, but the interruptions we found documented come from the model, the tool, the mode and the question policy.

  • OpenAI says GPT-6 Astra asks for clarification more readily than earlier models.
  • In Cursor, a user report says the /nohuman command did not stop the question dialogs, and staff said there is no off switch.
  • A conservative question policy pushed some agents to default silently on ambiguous tasks.
Paul-Antoine Tual

AI Transformation Leader · Junyr Method™ · Transition manager specialising in AI for French SMEs and mid-caps. Engineer from the École des Mines de Nantes, lawyer, developer since 1993.

← All articles

A2UI v1.0

What an agent sees

This card is rendered here from the same A2UI v1.0 file that agents receive. Our renderer dresses it in the site’s design; another client would dress it in its own.

Loading the surface…