The three largest model vendors each publish a prompting guide, and each guide contains a list of things not to do. Read together on November 5, 2026, all three warn against telling a reasoning model to think step by step, and two of them, Anthropic and Google, warn against shouting at the model. They disagree, three ways, on how many examples a prompt should carry. And most of the remaining warnings are not general advice at all but rules about one model generation, several of which produce an error rather than a worse answer.
That last point is why a list organised by vendor is less useful than one organised by mistake. A team that prompts Claude, GPT and Gemini from the same template, which is most teams running a gateway, needs to know which rule bites on which model. A prefill that worked on one Claude generation returns a 400 on the next. A custom temperature that Google already advised against on Gemini 3 is ignored with a deprecation header on Gemini 4. A "think step by step" line that helped a non-reasoning model is, in OpenAI's words, unnecessary on a reasoning one and can hinder it.
Two scope lines. Everything here is what the vendors say, quoted where the wording matters, with no measurements of our own; none of the three pages publishes the experiments behind its rules, and we do not invent before-and-after numbers. And this is about prompts to a model, not the instruction files that load into a coding agent, which our instruction-file audit covers separately.
- 01All three vendors say a reasoning model does not need step-by-step instructions.OpenAI's reasoning guide calls chain-of-thought prompts unnecessary and says they can hinder; Google says outlining reasoning steps is generally not necessary on Gemini 2.5 and 3; Anthropic says a prompt asking its newest models to write out reasoning may be declined.
- 02Both Anthropic and Google say: stop shouting.Anthropic's guide says Claude Opus 4.5 and 4.6 may overtrigger on prompts written as CRITICAL: You MUST and recommends normal phrasing. Google's Gemini 3 guidance says to avoid unnecessary or overly persuasive language.
- 03The vendors give three different answers on examples.Google: always include few-shot examples. OpenAI, for reasoning models: try zero-shot first, then few-shot if needed. Anthropic: include three to five examples for best results. The right number depends on which model you are prompting.
- 04Several rules are errors, not quality tips.Prefilled assistant turns return a 400 from Claude 4.6 models onward and on Claude Mythos Preview. Gemini 4 and Gemini 3.6 and later ignore temperature, top_p and top_k and will return 400s in a future release. These are version-bound and belong in a gateway's config, not a style guide.
- 05None of this is measured in public.All three guides are prescriptive. None publishes the experiment behind a rule, and several Anthropic rules name Opus 4.5 and 4.6 specifically. Treat each as a vendor's current advice for a named model, test it on your own traffic, and expect it to change.
01 — The SourcesThe three guides, and what each one covers.
Anthropic's page is the broadest. Its Claude prompting best practices page describes itself as the reference for current Claude models and lists fourteen of them by name, from Claude Fable 5.1 and Mythos 5.1 down to Haiku 4.5, with several rules tagged to a specific generation. OpenAI splits its advice across a general prompt engineering guide and a reasoning best-practices page that still refers to o1-era models in places. Google's prompting strategies page carries general design advice plus a section specific to Gemini 3.
| Guide | Models it names | Character of the advice |
|---|---|---|
| Anthropic, Claude prompting best practices | Fable 5.1 and 5, Mythos 5.1 and 5, Opus 5.5 to 4.6, Sonnet 5.5 to 4.6, Haiku 5.5 and 4.5 | Many rules tagged to one generation; migration notes included |
| OpenAI, prompt engineering and reasoning guides | "Reasoning models" versus "GPT models"; some text still anchored to o1-2024-12-17 | Split by model type; several API-usage rules mixed in |
| Google, prompting strategies | General section, plus Gemini 3; sampling note names Gemini 4 and Gemini 3.6 and later | Design principles plus one hard parameter deprecation |
02 — Reasoning PromptsTelling a reasoning model to think.
The most widely copied line in prompt engineering, "think step by step", is now a documented mistake on every vendor's reasoning models, with three different nuances. The why is the same in each case: a reasoning model already allocates hidden thinking before it answers, so a prompt that asks for visible reasoning either duplicates that work or competes with it.
Unnecessary, and can hinder
OpenAI's reasoning best-practices page: since these models reason internally, prompting them to think step by step or explain their reasoning is unnecessary, and such instructions may not enhance performance and can sometimes hinder it. Its prompt engineering guide says reasoning models do better with high-level guidance, likening them to a senior colleague where GPT models are a junior one.
Generally not necessary
Google's page says Gemini 2.5 and 3 series models automatically generate internal thinking text, so it is generally not necessary to have the model outline, plan or detail reasoning steps in the returned response. It stops short of forbidding it; the same page notes that a direct instruction to think very hard before answering can still improve performance.
May be declined
Anthropic's guide says to rely on the models' own thinking, and that a prompt asking the model to write out its reasoning, for example in thinking tags, may be declined on those five models. Manual chain-of-thought remains a fallback only on earlier models with thinking switched off.
The failure case is a gateway template written in 2024. It carries a reasoning instruction because that was good advice for the models of the time, and it now sends that instruction to three vendors who each say it is at best redundant. The example of the fix is to delete the line and, where you need a specific output structure, ask for the structure rather than the reasoning. Our post on defining done rather than effort is the longer version of that argument for agent prompts.
03 — EmphasisShouting, and persuading.
Two vendors have arrived at the same warning from different directions. Anthropic's is specific and mechanical. Its guide says that Claude Opus 4.5 and Claude Opus 4.6 are more responsive to the system prompt than earlier models, so prompts designed to stop a model under-using tools may now cause it to over-use them, and the fix is to soften aggressive language. The example it gives: where a prompt said "CRITICAL: You MUST use this tool when...", write "Use this tool when..." instead. The migration section says the same thing for Claude 4.6 models: reduce any guidance that pushed the model to be more thorough or to reach for tools more aggressively.
Google's is stylistic and aimed at Gemini 3: be precise and direct, state the goal clearly and concisely, and "Avoid unnecessary or overly persuasive language." The why in both cases is that the models now take emphasis literally. A line written to overcome an older model's reluctance becomes an instruction a newer model obeys too well. The example that bites is a customer-service prompt stuffed with capitalised ALWAYS and NEVER lines, which on Opus 4.5 or 4.6, by Anthropic's account, may over-trigger tools the customer did not need.
04 — Instruction QualityVague, negative and unexplained.
Three related mistakes come mostly from Anthropic's page, with one place where OpenAI's template quietly pulls the other way. The first is vagueness, and Anthropic's test for it is the one worth memorising: show the prompt to a colleague with minimal context and ask them to follow it; "If they'd be confused, Claude will be too." Elsewhere the guide adds that the latest models are trained for precise instruction following, so a request like "can you suggest some changes" may produce suggestions rather than changes, even if changes were what you meant. Say what you want done.
The second is negative framing. Under its heading on controlling response format, Anthropic's guide says to tell the model what to do instead of what not to do, replacing "Do not use markdown in your response" with an instruction that the response should be composed of smoothly flowing prose paragraphs. OpenAI's prompt engineering page does not contradict this, but its own prompt template lists, under Instructions, what the model should do and what it should never do. That is a mild tension rather than a disagreement: one vendor recommends positive phrasing, the other provides a slot for prohibitions.
The third is a rule with no reason attached. Anthropic says that giving the context or motivation behind an instruction tells the model more about your goals, and its example is precise: tell the model the output will be read aloud by a text-to-speech engine and it will understand why it should never use ellipses. Google makes the neighbouring point for Gemini under "Define parameters": explicitly explain any ambiguous terms. The failure case for all three is the same prompt, a page of bare prohibitions with no stated purpose, which a capable model follows to the letter and no further.
05 — Few-ShotHow many examples, three answers.
This is the one place the three guides give incompatible defaults, and it is the clearest case for reading the guide that matches the model rather than a generic checklist. Each vendor's wording, routed by the model you are prompting:
The disagreement is less strange than it looks once the model types are separated. Google's rule is written for Gemini in general; OpenAI's is written specifically for reasoning models and says GPT models benefit from very precise instructions; Anthropic's is written for Claude. There is no single rule because there is no single model. The practical consequence is that a prompt library with one examples policy is wrong for at least one vendor, which is an argument for the per-model templating our prompt pattern library is organised around.
06 — Version-Bound RulesRules that apply to one model and not the next.
Most of what remains in the three guides is not advice but configuration: a rule that produces an error, a header or a silent behaviour change on a named model. These belong in a gateway's per-model settings, and a team that reads them as general wisdom will apply them where they do not hold. The table lists each with the model the vendor names.
| Mistake | Model the vendor names | What happens | Vendor's fix |
|---|---|---|---|
| Prefilling the assistant turn | Claude 4.6 and later models, and Claude Mythos Preview | Request returns a 400 error; earlier models still accept prefills | Structured outputs or direct instructions |
| Setting temperature, top_p or top_k | Gemini 4 and Gemini 3.6 and later | Values ignored with a deprecation warning header; 400 errors in a future release | Remove the parameters; earlier Gemini 3 models keep defaults |
| Using the word "think" with thinking disabled | Claude Opus 4.5 | The model is particularly sensitive to the word and its variants | "consider", "evaluate" or "reason through" |
| Expecting effort to control length | Claude Opus 5 | Default responses run longer; raising or lowering effort does not reliably change visible length | Prompt explicitly for conciseness |
| Expecting markdown | OpenAI reasoning models, o1-2024-12-17 and later per the page | Markdown is avoided by default | Start the developer message with "Formatting re-enabled" |
| Expecting a conversational answer | Gemini 3 models | Answers are direct and efficient by default | Explicitly request a more conversational or detailed reply |
The markdown row is the one that trips cross-vendor prompts most often, because the defaults point in opposite directions. OpenAI's reasoning models suppress markdown unless the developer message re-enables it; Anthropic's guide, by contrast, supplies a prompt for avoiding excessive markdown. A shared template therefore gets too little formatting from one vendor and too much from another, and the only fix is testing the same prompt on each. The sampling row is the one that will break loudest: a request carrying a temperature value to Gemini 4 is ignored today and will be rejected later, on Google's own statement.
07 — Code PromptsCoding prompts that over-deliver.
Anthropic's guide carries two coding-specific warnings that describe behaviour rather than syntax, and both are the kind of mistake a developer makes by not saying anything. They apply to the models the page names, and the fixes it gives are sample prompt lines, not settings.
Extra files, abstractions and flexibility nobody asked for
The guide states that Claude Opus 4.5 and Claude Opus 4.6 have a tendency to overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that was not requested. Its remedy is a prompt line to avoid over-engineering and only make changes that are directly requested.
Solutions that make the tests pass and nothing else
Claude can sometimes focus too heavily on making tests pass at the expense of more general solutions, the guide says; its sample prompt tells the model not to hard-code values. The mistake on the human side is a prompt that names the test and not the behaviour the test stands for.
An implied action is not a requested one
If you say "can you suggest some changes", the guide notes, Claude will sometimes provide suggestions rather than implementing them. Precise instruction following cuts both ways: it does exactly what the sentence says, so the sentence has to say what you want done.
The failure case is a coding agent asked to "make the failing test pass" with no further constraint. The vendor's own guidance predicts the outcome: a change that satisfies the assertion and generalises nowhere. The equivalent advice for OpenAI's newest models lives in its model-specific guidance rather than the general guide, which our GPT-6 Astra prompts and skills upgrade guide covers.
08 — Beyond the PromptMistakes that live outside the prompt text.
OpenAI's guides fold three API-usage warnings into their prompting advice, and they are worth separating because no amount of prompt wording fixes them. The first is the unpinned model: OpenAI says even different snapshots within the same family can produce different results and strongly recommends pinning production applications to a specific dated snapshot. The second is assuming where the answer lives: the Responses API output array often has more than one item, and it is not safe to assume the text is at the first item's first content element.
The third is the one with a cost attached. If you call a reasoning model through the Chat Completions API, OpenAI says reasoning items are never carried into the model's context between turns, which results in slightly degraded performance and greater reasoning-token usage in complex agentic cases with many function calls. A team that migrated prompts carefully but kept the older endpoint is paying for reasoning it then throws away. Google's equivalent is on the tools side: its page says grounding with Google Search should be enabled whenever the model may need obscure or recent facts, and code execution whenever it needs to do any kind of arithmetic, counting or calculation. Asking the prompt to do what a tool should is the last mistake on the list.
If you run prompts across more than one vendor and want the per-model rules turned into gateway configuration rather than tribal knowledge, that is part of what our AI transformation engagements set up, alongside the evals that tell you whether a rule helped.
09 — ConclusionRead the guide for the model you are actually calling.
Delete the reasoning instructions and the capitals, set the examples policy per model, and move the version-bound rules into gateway config.
On November 5, 2026, the prompting guides from OpenAI, Anthropic and Google agree that a reasoning model should not be told to think step by step, and Anthropic and Google add that aggressive emphasis now backfires. They disagree on examples: always for Gemini, zero-shot first for OpenAI's reasoning models, three to five for Claude.
The rest of the warnings are version-bound. Prefills return a 400 from Claude 4.6 models onward. Temperature and top_p are ignored on Gemini 4 and Gemini 3.6 and later, with errors to follow. OpenAI's reasoning models suppress markdown unless re-enabled. Opus 4.5 is sensitive to the word "think" with thinking off, and Opus 5 needs an explicit request for brevity. None of these is a style preference; each is a fact about one model that a shared template has to know.
And none of it is measured in public. Treat the guides as the vendor's best current description of its own model, apply the rules to the models they name, and keep an eval that can tell you when a rule stops being true.