"Think step by step" is the most repeated AI advice on the internet. For a growing class of tools, it is now wrong.
Almost everything you have absorbed about writing a good prompt, you learned on one tool. We read the official documentation from the makers of more than 60 of them to find out where those habits quietly stop working — and, in six documented cases, start working against you.
Download the full report (PDF)Bhattacharyya, A. (2026). Horses for Courses: What AI Companies' Own Instructions Reveal About Why the Same Prompt Does Not Work Everywhere. AIToolSpecs.com Research Publication No. 03. Available at: https://aitoolspecs.com/publications/prompt-dialects
There is no such thing as a good prompt. There is only a good prompt for a specific tool.
Prompting advice spread the way folklore spreads. Someone found a phrase that worked, posted it, and it was repeated until it hardened into a rule. That worked well enough when there was essentially one tool that mattered and everyone was learning it at the same time.
There are now dozens of tools that matter, built on genuinely different machinery, and the folklore did not split when the tools did. So people carry habits learned on ChatGPT into Claude, habits learned on Midjourney into a video generator, and assume a prompt that reads well is a prompt that works.
If you have ever added "think through this step by step" to a prompt out of general good practice: on one entire category of current models, the companies that built them now tell you not to.
Both halves of the pattern are useful, in different ways. The contradictions tell you where a habit you trust is about to cost you quality without ever announcing itself. The agreements tell you which habits are genuinely safe to carry everywhere.
01 · The most repeated prompting tip on the internet has expired for some tools
"Let's think step by step" is prompting's greatest hit. It has been copied into more templates, courses and LinkedIn posts than any other phrase in the field, and for standard chat models it genuinely earns its reputation — those models do not reason before answering unless you ask them to, so asking is what makes it happen at all.
Then a new category arrived. OpenAI's reasoning models — the o-series and similar "thinking" models — run a hidden reasoning process before they produce a word of output. OpenAI's own guidance for them says plainly: do not tell them to think step by step or explain their reasoning. It is unnecessary, and it can make results worse.
The reason is mechanical rather than stylistic. Asking a model to narrate a process it has already completed internally is asking for the work to be done twice, once invisibly and once for show. The second pass adds nothing and can measurably degrade the answer.
If "think step by step" is muscle memory — something you add to every prompt because it is good practice — that habit now silently works against you the moment you are talking to a reasoning-labelled model. And tools increasingly do not make it obvious which kind you have been routed to.
02 · Two rivals, working separately, arrived at the same warning: stop shouting
This is the finding we would flag first if you only read one section, because it is the only place in the entire review where two competing companies independently documented the same failure.
Google's current Gemini documentation states it plainly: avoid unnecessary or overly persuasive language. Anthropic's documentation for Claude Opus 4.5 and 4.6 names the failure mode — these models are more responsive to forceful phrasing than earlier versions, and prompts like "CRITICAL: You MUST use this tool when…" now cause the model to overtrigger.
THE SAME INSTRUCTION, WRITTEN TWO WAYS
Two different companies, two different products, one underlying warning: shouting at a current frontier model does not make it more obedient. It makes it less predictable.
Anthropic ties the overtriggering finding specifically to Opus 4.5 and 4.6 — not to every Claude model ever shipped. On this subject, which exact version can matter as much as which company.
Treat all-caps, "you MUST", "this is CRITICAL" phrasing as a last resort rather than a default — the one habit two separate, competing companies have independently flagged as backfiring on their current models.
The other four contradictions
Four more documented conflicts, each with the vendors' own wording and worked before-and-after prompt examples, are set out in full in the PDF.
For long documents — Anthropic flags roughly 20,000 words of input and up — put the document first and your question at the very end, which Anthropic states can improve answer quality by up to 30% on complex, multi-document tasks. OpenAI's default points the other way: instructions first, material after.
Midjourney reads terse, code-like parameter flags. ChatGPT's image generation wants full conversational prose. A prompt written correctly for one reads as garbled or oddly vague to the other.
Kling gives you a dedicated negative-prompt field. Runway and Veo have no equivalent — exclusions go into the main prompt as ordinary instructions.
CrewAI puts a ratio on it: roughly 80% of your effort belongs in designing the task instructions, and only 20% in describing the agent carrying them out.
The habits that survived contact with every tool we checked
Contradictions make the better story, but they are only half the finding — and arguably the less useful half. When unrelated companies, competing with each other, independently land on the same answer without copying, that convergence is a strong signal you have found something structural rather than stylistic.
Two showed up. Three separate coding-agent tools — Anthropic's Claude Code, Cursor and Windsurf — each independently built the same standing instruction file, loaded automatically every session and kept separate from the task in front of you. And Anthropic's Claude Code guidance states a specific, concrete rule: after two failed correction attempts on the same problem, stop correcting and start over with a clearer initial instruction that folds in what you learned.
Machinery diverges. Workflow converges.
Put both lists side by side and the shape of the split is not random at all. It is almost perfectly clean — each set traces back to a different kind of question.
Whether the model reasons invisibly before answering. How it distributes attention across a very long input. Whether a specific version was tuned to respond more strongly to emphasis. Whether exclusions are a field or a sentence. These are technical facts about how a particular tool was built — and different tools are genuinely built differently.
How a person separates standing conventions from one-off requests. When to stop manually correcting something that is stuck and start over. How to give an assistant a real pass/fail check. These are solved problems about how humans work with any capable assistant — so separate teams solving them keep arriving in the same place.
Which gives you a rule you can apply to a tool this report never examined: advice about how the model works is probably tool-specific; advice about how you work is probably portable. When you read a prompting tip somewhere, ask which of the two it is. That single question sorts most of the folklore.
The research behind our Prompt Generation engine
Everything above is the foundation for the part of AIToolSpecs.com's paid reports that writes the starting prompt for each tool in your recommended stack. It is built on the principle this entire report documents: the right prompt depends on how that specific tool actually works, not on a general-purpose template with the tool's name swapped in.
A prompt written for Claude and a prompt written for Kling should not look like the same document with a find-and-replace run over it. In our reports, they do not.
A note on scope and sources
A few honest limits on what this research is and is not, so you can weigh it correctly.
This covers AI tools we researched directly — not every tool available, and not every model version each vendor currently offers.
Every specific claim above about a named vendor's guidance was checked against that vendor's own current official documentation at the time of writing (August 2026) — not a summary, a blog post, or a third party's paraphrase.
Midjourney's parameter-flag style rests on widely repeated practitioner convention rather than formal published documentation, because Midjourney does not document with the same formality as Anthropic, OpenAI or Google. We flag that rather than present it with false certainty.
All worked "this way, not that way" prompt examples were written by us to illustrate the underlying documented principle. They are original illustrations, not reproductions of any vendor's own examples.
Vendor guidance changes as new model versions ship. A quirk tied to one specific version — the Opus 4.5/4.6 finding above is exactly this — may not apply to whatever that vendor releases next.
Naming a tool's documented quirk, whether it works in a prompt's favour or against it, has no connection to whether we have an affiliate relationship with that tool elsewhere on our site.
Bibliography
- Claude — long-context document placement, up to 30% improvement — Anthropic, "Long context prompting tips," official prompt engineering documentation.
docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/long-context-tips - OpenAI's instructions-first, delimiter-based default ordering — Prompt Builder, "ChatGPT & OpenAI Prompt Engineering Guide," citing OpenAI's own published guidance, 2026.
promptbuilder.cc/blog/openai-prompt-engineering-guide-best-practices-2026 - OpenAI reasoning models — avoid chain-of-thought prompting, try zero-shot first — OpenAI, "Reasoning best practices," official API documentation.
developers.openai.com/api/docs/guides/reasoning-best-practices - Gemini — avoid overly persuasive language — Google, "Prompt design strategies," official Gemini API documentation.
ai.google.dev/gemini-api/docs/prompting-strategies - Claude Opus 4.5 / 4.6 — forceful language causing tool overtriggering — Anthropic, "Prompting best practices," official Claude Platform documentation.
platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices - Kling's dedicated negative-prompt field vs. Runway/Veo's inline approach — Artlist, "Negative prompts for Kling, Veo, and Wan," 2026.
artlist.io/blog/negative-prompts-ai-video - Midjourney's parameter-flag dialect vs. DALL·E/ChatGPT's conversational prose — SurePrompts, "AI Image Prompting: The Complete 2026 Guide."
sureprompts.com/blog/ai-image-prompting-complete-guide-2026 - CrewAI's 80/20 task-vs-agent design rule — CrewAI, "Crafting Effective Agents," official documentation.
docs.crewai.com/en/guides/agents/crafting-effective-agents - Claude Code — the two-failed-corrections rule; verification-first workflow — Anthropic, "Best practices for Claude Code," official engineering documentation.
anthropic.com/engineering/claude-code-best-practices - Convergent persistent-context file pattern across Claude Code, Cursor and Windsurf — The Prompt Shelf, "AGENTS.md vs CLAUDE.md vs .cursor/rules: The Complete 2026 Three-Way Comparison," 2026.
thepromptshelf.dev/blog/agents-md-vs-claude-md-vs-cursorrules-three-way-2026
We did the homework, so you don't have to.
Tell us your budget, goal, skill level and hardware, and we will match you to tools that actually fit — with a starting prompt written for how each specific tool really works, not a template with the name swapped in.
Get your own plan at aitoolspecs.com