Most prompting advice that gets repeated as a rule is either true for one model only, or was never in any vendor’s guide at all. This page takes the common ones, one at a time, and sets each against what OpenAI, Anthropic, Google, Black Forest Labs and Midjourney actually publish. Where the vendors disagree with each other, it says so, because that disagreement is usually the real lesson.
- A role sets tone, not skill. Anthropic says a role focuses behaviour and tone. Give the task details as well.
- “Think step by step” is not universal. OpenAI says it can hinder its reasoning models.
- Length depends on the model. Midjourney favours short prompts; Black Forest Labs gives a medium range for FLUX.2; Anthropic wants long documents placed first.
- Negative prompts and quality tags are model-specific. FLUX.2 has no negative prompt, and Google’s Imagen guide recommends quality keywords where Black Forest Labs asks for camera details instead.
- Leave temperature alone on Gemini 3.x. Google says so, and has deprecated the parameter.
- No prompt is finished on the first try. Google and OpenAI both describe prompting as iteration, one change at a time.
Each myth below is stated in plain words, then answered with the vendor’s own sentence, then one line on what to do instead. For beliefs about AI itself, see common AI misconceptions; for words that get misused, see AI terms people get wrong.
Chat models: words that are supposed to work like switches
“Magic words, tips and threats make the model try harder”
None of the vendor guides read for this page recommends politeness formulas, promised tips or threats.Reasoning, September 2026 — based on the ten vendor pages listed under Sources; this site did not search every guide every vendor publishes. Google’s guidance for Gemini 3 points the other way: it asks for plain statements of the goal.Google, Prompt design strategies, read at source 22 Sep 2026: “Avoid unnecessary or overly persuasive language.” Anthropic’s guide says the thing that improves results is explaining why an instruction matters.Anthropic, Prompting best practices, read at source 22 Sep 2026: “Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals and deliver more targeted responses.”
Instead: replace the flattery or the threat with the reason the task matters and who the output is for.
“Telling it ‘you are an expert’ makes it more capable”
Anthropic, which does recommend roles, describes what a role does in narrower terms.Anthropic, Prompting best practices, read at source 22 Sep 2026: “Setting a role in the system prompt focuses Claude's behavior and tone for your use case.” This site made the bigger claim once and corrected it on 22 Sep 2026: prompting for LLMs used to say every model responds “significantly better” to a role, with no source.
Instead: keep the role if it helps the tone, and put the actual specifics of the task next to it.
“Writing it in capitals makes it obey”
Anthropic’s guide gives “NEVER use ellipses” as its less effective example, and prefers a version that gives the reason.Anthropic, Prompting best practices, read at source 22 Sep 2026, the more effective version: “Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them.” For two of its models it goes further, and says forceful wording can backfire. It notes that Claude Opus 4.5 and 4.6 respond more strongly to the system prompt than earlier models, so prompts written to push them into using tools may now make them overuse those tools.Anthropic, Prompting best practices, read at source 22 Sep 2026: “Claude Opus 4.5 and Claude Opus 4.6 are also more responsive to the system prompt than previous models.” and “The fix is to dial back any aggressive language.” That advice is written for those models; the same guide says model-specific advice should be re-tested before it is applied elsewhere.
Instead: write the rule once, in normal case, with the reason attached.
“One example is enough”
Anthropic gives a range, not one.Anthropic, Prompting best practices, read at source 22 Sep 2026: “Include 3–5 examples for best results.” Google is stronger still about including examples at all.Google, Prompt design strategies, read at source 22 Sep 2026: “We recommend to always include few-shot examples in your prompts.” The site’s CARE framework page carried the one-example version until it was corrected on 22 Sep 2026.
Instead: give several varied examples, so the model copies the pattern rather than one instance.
No word works like a switch. What the chat vendors ask for is context: the reason, the audience and, on most models, a handful of examples.
Reasoning models: advice that expired
“‘Think step by step’ always helps”
OpenAI says that on its reasoning models it may not help, and can make things worse.OpenAI, Reasoning best practices, read at source 22 Sep 2026: “Some prompt engineering techniques, like instructing the model to “think step by step,” may not enhance performance (and can sometimes hinder it).” The same guide asks for short, plain prompts on those models: “Keep prompts simple and direct”. How to ask AI well said asking for the steps “usually improves the answer” until 22 Sep 2026, and was corrected against this guide.
Instead: on a reasoning model, state the goal and what a good answer looks like, and leave the method to the model.
“More examples always help”
Here the vendors pull apart. Google says to always include examples; OpenAI says its reasoning models often do without them.OpenAI, Reasoning best practices, read at source 22 Sep 2026: “Reasoning models often don’t need few-shot examples to produce good results, so try to write prompts without examples first.”
Instead: on an OpenAI reasoning model, try without examples first and add a few only if the output format drifts.
Tricks that made older chat models reason out loud are, on reasoning models, at best redundant. Check which kind of model you are prompting before reaching for them.
Length: longer is not better, and neither is shorter
“A longer, more detailed prompt is always better” (or always worse)
Each vendor answers this differently, for its own model. Midjourney favours short prompts.Midjourney, Prompt Basics, read at source 22 Sep 2026: “Short and simple prompts typically generate the best images with Midjourney.” and “Avoid making long lists or detailed instructions; these can confuse the process.” Black Forest Labs gives a middle range for FLUX.2.Black Forest Labs, Prompting Guide — FLUX.2 [pro] & [max], read at source 22 Sep 2026: “Medium (30-80 words): Usually ideal for most projects”. Anthropic asks for precision, and for long material to go first with the question last.Anthropic, Prompting best practices, read at source 22 Sep 2026: “The more precisely you explain what you want, the better the result.” and “Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.” Google asks Gemini 3 users to “State your goal clearly and concisely.”Google, Prompt design strategies, read at source 22 Sep 2026.
Instead: include every detail that matters and nothing that does not, then check the length guidance for the specific model on what the vendors actually say.
The right length is a property of the model, not of prompting. The rule that holds everywhere is to cut words that carry no instruction.
Image and video models: habits carried over from one tool
“Negative prompts work on every model”
Some models have no negative prompt at all.Black Forest Labs, Prompting Guide — FLUX.2 [pro] & [max], read at source 22 Sep 2026: “FLUX.2 does not support negative prompts.” Midjourney warns that naming the unwanted thing in the prompt can backfire: in its example, a party with “no cake” means “a cake might still appear”.Midjourney, Prompt Basics, read at source 22 Sep 2026. Google’s image guide asks you to “describe the intended scene positively” instead.Google, Nano Banana image generation, read at source 22 Sep 2026. Which models have a field, and which do not, is on the negative prompt library.
Instead: check your model first; where there is no field, describe what should be there rather than what should not.
“Quality tags like ‘8K masterpiece’ make any image more realistic”
The vendors do not agree, which is exactly why the tags cannot be universal. Google’s Imagen guide recommends quality keywords, with examples made on Imagen 3.Google Cloud, Prompt and image attribute guide, read at source 22 Sep 2026: “Certain keywords can let the model know that you're looking for a high-quality asset.” and “Photos - 4K, HDR, Studio Photo”. Black Forest Labs prefers a named camera and lens, which it says “produces more authentic results than just” asking for a professional photo.Black Forest Labs, Prompting Guide — FLUX.2 [pro] & [max], read at source 22 Sep 2026. OpenAI treats quality as a setting on the request and advises against special syntax.OpenAI, Image prompting, read at source 22 Sep 2026: “Choose the format that makes the requirements easiest to read and update rather than relying on special syntax.” and “Compare medium or high quality for small text, dense information, or multiple fonts.” The case for deleting quality stacks, and its limits, is on why AI output looks fake.
Instead: describe physical things (the light, the lens, the surface) and use the model’s own resolution and quality settings where it has them.
Image prompt habits usually come from one tool. Before copying a keyword list, find out whether the model you are using reads it at all.
Settings: the dial that moved
“Turn the temperature down for factual answers”
That was general advice, and Google’s own page still explains it in general terms: “Lower temperatures are good for prompts that require a more deterministic or less open-ended response”. The same page then makes an exception for its current models.Google, Prompt design strategies, read at source 22 Sep 2026: “Although you can modify these parameters, we strongly recommend keeping them at their default values for Gemini 3.x models.” and “Changing these parameters (for example, setting the temperature below 1.0) can cause unexpected behavior, such as looping or degraded performance”. Google’s release notes for 21 July 2026 went further.Google, Gemini API release notes, entry dated 21 July 2026, read at source 22 Sep 2026: “The sampling parameters temperature, top_p and top_k are now deprecated.” The site’s prompting glossary carried the old advice until 22 Sep 2026. The details are on the Gemini temperature deprecation.
Instead: on Gemini 3.x, leave the defaults and ask for accuracy in the prompt; for other vendors, read their current docs rather than carrying the old rule over.
A settings rule is only as current as the model it was written for. On Gemini 3.x, this one reversed.
Workflow: the prompt is not the finished product
“A good prompt works the same on every model”
OpenAI says results can shift even between versions of one model.OpenAI, Prompt engineering, read at source 22 Sep 2026: “Even different snapshots of models within the same family could produce different results.” Anthropic tells readers to treat model-specific advice as specific.Anthropic, Prompting best practices, read at source 22 Sep 2026: “Where a technique names a specific model, treat it as measured on that model and re-check it against your own evals before applying it to another.” The order of a prompt differs between vendors too, as the LLM cheat sheet shows.
Instead: when you move a prompt to a new model or version, re-test it before you rely on it.
“Write one perfect prompt and you are done”
Google’s prompt guide opens with a note: “Prompt engineering is iterative.”Google, Prompt design strategies, read at source 22 Sep 2026. Its image guide says the same about pictures.Google, Nano Banana image generation, read at source 22 Sep 2026: “Don't expect a perfect image on the first try.” OpenAI’s image guide lists “Iterate deliberately.” among its fundamentals, one change at a time.OpenAI, Image prompting, read at source 22 Sep 2026.
Instead: change one thing, compare against a standard you wrote down first, and keep a log, as set out on how to test a prompt.
A prompt is a draft you test, not a spell you find. The vendors say so themselves.
What this page could not verify
- Whether “please”, tips or threats change output. No vendor guide read here tests them. The page reports only what the vendors recommend instead.
- Whether capitals hurt on every model. Anthropic’s warning about aggressive language is written for two named Claude models and for tool use.
- Whether quality tags help on current image models. Google’s Imagen guide recommends them, with examples made on Imagen 3; this site has not tested them on newer models.
- xAI, Runway, Kling and other vendors were not re-read for this page. Their negative prompt rules are sourced on the negative prompt library.
Anthropic, Prompting best practices · OpenAI, Reasoning best practices · OpenAI, Prompt engineering · OpenAI, Image prompting · Google, Prompt design strategies · Google, Gemini API release notes · Google, Nano Banana image generation · Google Cloud, Prompt and image attribute guide · Black Forest Labs, Prompting Guide — FLUX.2 [pro] & [max] · Midjourney, Prompt Basics. All read at source on 22 September 2026. Vendor guidance changes with each model; re-read the source before relying on it.