How to Write ChatGPT Prompts: Your Complete Guide

Most prompt guides are maintenance manuals for models that no longer exist. This edition rebuilds the advice from the vendors’ current documentation, ranks the evidence, and flags the places where the three biggest labs contradict each other.

By Tom Morgan · First published April 12, 2026 · Fully revised September 30, 2026 · About 3,300 words · 14 min read

How this was produced. Research and drafting for this revision were AI-assisted. The vendor pages cited below were opened and read on September 30, 2026, not summarized from memory. No original benchmarks were run for this article, and nothing here should be read as a test result. Where a claim is my own analysis, it is labelled that way.


The short answer

Quick answer In September 2026, all three major vendors tell you to describe the outcome you want, supply the context the model cannot guess, state one or two hard boundaries, and stop there. They also tell you to delete instructions that were written for weaker models. They disagree on two things that matter: whether to include examples, and whether to script the steps. Treat any prompt you wrote before mid-2026 as legacy code and rebuild it from a smaller baseline instead of patching it.

Key facts at a glance

FactSource typeDate
OpenAI’s current family is GPT-6 (Astra, 6.1 Sol, Luna); its docs tell you to start from a fresh, minimal prompt when migratingVendor documentationRead Sept 30, 2026
Anthropic staff said they cut about 80% of Claude Code’s system prompt because the newest models “want a smaller system prompt”Trade press reporting a conference talkJuly 2, 2026
Anthropic reports that putting the question after long documents improved response quality by up to 30% in its own testsVendor documentation (first-party claim)Read Sept 30, 2026
OpenAI’s reusable “prompt objects” API is scheduled to shut down on November 30, 2026Vendor documentationRead Sept 30, 2026
Google’s own Gemini prompting page recommends few-shot examples in every promptVendor documentationPage updated Sept 17, 2026

Why the April edition of this guide needed rebuilding

The first version of this article got the direction right (shorter, clearer, outcome-first) and several specifics wrong or stale. Rechecking it against current documentation turned up five problems:

  1. It compared GPT-4o, o3, Claude 3.x and Gemini 1.5. None of those is the current generation. Google’s Firebase documentation notes that Gemini 2.0 Flash and Flash-Lite were scheduled to shut down on June 1, 2026.
  2. It called role prompts dead. Anthropic’s current guide still recommends a one-sentence role. What has faded is the elaborate persona, not the role line.
  3. It said pasted examples work best “across all models.” Google says so. An Anthropic staffer says the opposite for the newest Claude models. See the disagreement section below.
  4. It said Anthropic recommends splitting tasks into sequential steps. Anthropic’s current guidance for its newest models leans toward handing over the whole task and the definition of done.
  5. It contained a “sweet spot” chart (20–60 words) and an opening anecdote drawn from unpublished personal testing. That material couldn’t be documented or reproduced, so it has been removed. No vendor publishes a word-count target, and this edition doesn’t invent one.

That last point is the reason for the next section.


Which sources to trust, in order

Prompting is a field where confident advice outruns evidence. I ranked what this article rests on so you can weigh it.

TierWhat it isStrengthExamples used here
1Vendor documentation for the model in questionStrongest available, but first-party: it describes what the vendor observed on its own models and testsOpenAI “Using GPT-6”; Anthropic “Prompting best practices” and “Prompting Claude Fable 5.1”; Google “Prompt design strategies”
2Trade press reporting vendor staff or vendor guidesUseful for direction, one step removed from the sourceThe Decoder on OpenAI’s user guide (July 13) and Anthropic’s prompt cut (July 2)
3Third-party “rules that work” blog postsWeakest: often repackage tier 1 and sometimes contradict itSeveral ranked for Claude Fable 5; not relied on here

A disagreement worth naming. Some third-party guides for Claude recommend negative prompting (“tell it what not to do”) as a headline rule. Anthropic’s own documentation advises the reverse for controlling format: say what you want instead of what to avoid. Where tiers conflict, this article follows tier 1.

A limit on all of it. Every tier-1 source is a vendor describing its own product. None of the three publishes head-to-head comparisons against the others’ models, and I found no independent benchmark that isolates prompt style as the variable. Everything below is therefore “what the makers recommend,” not “what has been proven to win.”


Where all three agree

Reading the three sets of docs side by side, one pattern stood out: the vendors agree far more about what to remove than about what to add.

PrincipleOpenAIAnthropicGoogle
Lead with the outcomeDescribe what good looks like, the constraints and the evidence; avoid step-by-step guidance unless the exact path mattersPrefer general instructions over prescriptive step listsState the goal precisely and directly; skip persuasive padding
Give the model context it can’t guessAttach only sources that change the answer (user guide)Explain why an instruction matters; the model generalizes from the reasonSupply the material to answer from instead of assuming the model has it
Keep roles shortSuggested structure opens with a one-to-two-sentence roleA single sentence “makes a difference”Role lives in the system instruction, a couple of sentences
Reserve absolutes for real invariantsALWAYS/NEVER only for true rules such as safety or required fieldsDial back “CRITICAL: you MUST” phrasing on recent models, which now over-triggerConstraints should be explicit but not shouted
Structure long inputsPut stable content first, dynamic content lastDocuments on top, question at the end, tags to separate partsContext first, instructions at the very end; consistent delimiters

Two rows deserve emphasis because they are cross-vendor consensus you can act on today:

  • Long documents first, question last. Anthropic and Google give the same instruction independently. Anthropic attaches a figure (up to 30% better responses in its tests); treat that as a vendor claim, not a law.
  • Short roles. No current vendor page endorses a paragraph of persona. All three treat the role as a line or two.

Where they disagree (and what I’d do about it)

1. Examples: essential, or a trap?

  • Google: include few-shot examples in every prompt, and says prompts without them are likely less effective.
  • Anthropic’s docs: examples are among the most reliable ways to steer format, tone and structure; aim for three to five, varied, wrapped in tags.
  • Anthropic staff (reported): for the newest Fable models, examples tend to constrain output, because the model is more imaginative than the samples it is shown.

Those three statements cannot all be applied to the same task. I found no published head-to-head test that settles it.

My judgment (analysis, not data). The split is about the kind of task. When the job is matching a format, a voice or a classification scheme, an example beats a paragraph of description on any model, and Google and Anthropic’s docs both support that. When the job is judgment (a strategy, an argument, a design), an example becomes a ceiling: the model imitates its shape. For judgment tasks, describe the criteria and skip the sample. If you must include one, say it illustrates format only. Untested, so verify on your own inputs.

2. Scripted steps: helpful scaffolding or noise?

OpenAI and Anthropic now say to hand over the outcome and let the model choose the route. Google’s general guidance says to break complex work into chained prompts, and its own agent template is a nine-point procedure. The reconciliation is that Google’s advice targets pipelines you build (one prompt per stage so you can inspect each output), while OpenAI and Anthropic’s advice targets a single request to a reasoning model. Anthropic makes the same distinction: chain when you need to see intermediate outputs, not to help the model think.

A small inconsistency in Google’s own page is worth knowing about. Its prose says the model generally doesn’t need to outline its reasoning in the reply, yet its example template still ends with a line asking it to think step by step. Both can be true (the line is cheap), but it shows how much of this guidance is habit carried across generations.

3. Default style

Each vendor describes a different house style that you have to steer against:

  • OpenAI says GPT-6 Astra leans toward detailed, list-and-table-heavy replies and may repeat phrases across sessions; specify the writing style you want.
  • Anthropic says Fable 5.1 formats less than earlier models, so old anti-bullet rules can strip out structure the content needs. Its fix is to replace “never use lists” with a rule for when lists are appropriate.
  • Google says Gemini 3 answers directly and efficiently by default; ask explicitly if you want conversational or detailed output.

The practical consequence: a formatting instruction that fixed one model can break another. Carrying a “no bullets” rule from a GPT prompt into a Claude prompt is a common way to make output worse.


The non-obvious point: prompt debt

Most coverage frames prompting as adding skill. The current documentation frames it as removing baggage. OpenAI tells developers migrating to its newer models to treat them as a new family and begin from the smallest prompt that keeps the product working. Anthropic’s Fable 5.1 guide says to remove instructions that suppress behavior the model now does well, and warns that verification reminders carried over from older prompts can cause over-checking on Opus 5.

I’m calling this prompt debt: instructions that were a patch for a past model’s failure and now cost tokens, narrow the model’s options, or contradict its new defaults. (The term and the framework below are my own, offered as a working heuristic, not as reported research.)

The Prompt Debt Audit

Sort every line of a saved prompt into one of five buckets:

BucketWhat it isAction
OutcomeWhat the finished result must beKeep
ConstraintA boundary that would cause real harm if broken (audience, length, “keep the approved figures unchanged”)Keep, and keep few
ContextFacts, files or background the model cannot inferKeep only what changes the answer
ProcessHow to get there (numbered steps, “first analyze, then compare”)Delete, then re-add only if the output gets worse
RitualLines added to fix an old failure: stacked personas, “take a deep breath,” shouted rulesDelete

Then run the stripped prompt on five real inputs from your own work before restoring anything. If a deleted line’s absence causes a failure you can point to, put it back and write down why. That way every surviving line has a reason.

What this changes in practice. If you maintain a library of saved prompts or custom instructions, the highest-return hour you can spend this quarter is not writing new prompts. It is auditing old ones.


The model-by-model cheat sheet (September 2026)

Model lineups and defaults change every few months. This table reflects the vendor pages as read on September 30, 2026.

Model familyWhat the vendor documentsWhat to change in your prompt
OpenAI GPT-6 Astra / 6.1 Sol / LunaAstra is more likely to stop and ask a question than earlier models; it follows instructions in skill files and AGENTS.md more closely; default replies lean toward lists and tables. Astra and 6.1 Sol don’t support the none reasoning effort; temperature and top-p must be removed when reasoning is onIf it pauses when you expected it to act, tell it that “can you…” requests mean do the work. Audit any attached instruction files for conflicts. State your formatting preference. Move stored prompts off OpenAI’s reusable prompt objects before the November 30 shutdown
OpenAI GPT-5.5 (earlier generation)Default reasoning effort is medium; higher effort isn’t automatically better and can cause overthinking with conflicting instructionsIf you’re still on it, cut process instructions and raise effort only when your own tests show a gain
Anthropic Claude Fable 5.1 / Mythos 5.1Thinking is always on. Default effort is high; the medium setting roughly matches Fable 5 at lower cost. At low effort it searches less and answers from memory moreDelete anti-formatting rules. For anything about a recent name or product, raise effort or tell it to verify before answering. Test each effort level against your own tasks
Anthropic Claude Opus 5.5 / Sonnet 5.5Opus 5-class models verify their own work; verification reminders from older prompts can add costRemove “double-check your answer” boilerplate; re-add only if you see errors
Google Gemini 3.xAnswers directly by default; wants consistent delimiters (XML-style tags or Markdown headings, one or the other); critical constraints in the system instruction or at the top; questions at the end of long contextsKeep temperature and other sampling settings at their defaults. Turn on Google Search grounding for recent facts and code execution for arithmetic

Not covered: Meta, xAI, Mistral, DeepSeek and other model families. I did not read their current documentation for this edition, so nothing here applies to them by default.


Three templates you can actually use

These are starting points built from the vendors’ guidance. They have not been benchmarked.

1. The four-block prompt (matches OpenAI’s user guide: goal, context, output format, boundaries; every block is optional for small tasks)

Goal: [the result, in one sentence]Context: [audience, situation, and only the sources that change the answer]Format: [length, structure, tone]Boundaries: [one or two hard rules, e.g. "Keep the approved dates and budget figures unchanged"]

2. The long-document prompt (documents first, question last, per Anthropic and Google)

<documents>

[paste the material here, each source in its own tagged block]

</documents> Using only the documents above: [the task]. First pull the passages that bear on the question, then answer from them. If the documents don’t say, tell me that instead of guessing.

3. The audit prompt (turns the Prompt Debt Audit into a request)

Below is a prompt I use regularly. Sort each sentence into: outcome, constraint, context, process, or ritual (a line that looks like a patch for an old failure). Rewrite the prompt keeping only outcome, constraint and context. List the process lines you removed so I can test them one at a time.<prompt>

[paste your prompt]

</prompt>

A verification line for high-stakes work, adapted from OpenAI’s user guide: before finalizing, ask the model to check its own output against a concrete test, for example “confirm that every action item has an owner and a deadline, and list any that don’t.” A specific check beats “double-check your work.”


Myths, rated against current documentation

Common adviceVerdictWhy
“Act as a world-class expert with 20 years of experience”OverstatedA one-line role is still endorsed. Long personas aren’t recommended by any of the three vendors
“Think step by step”Model-dependentAnthropic prefers general instructions like “think thoroughly” over hand-written plans; OpenAI says avoid step guidance unless the path matters; Google says reasoning outlines in the reply are generally unnecessary
“Longer, more detailed prompts get better results”Wrong for instructions, right for contextLong documents are fine and well supported. Long instruction lists are what the vendors say to cut
“Always include examples”ContestedGoogle yes; Anthropic’s staff caution for its newest models. See the disagreement above
“Use CRITICAL / ALWAYS / NEVER so it obeys”Counterproductive on recent modelsOpenAI reserves absolutes for true invariants; Anthropic says recent models over-trigger on emphatic language
“Turn temperature down for accuracy”Often unavailable or discouragedOpenAI says to remove it when reasoning is enabled on GPT-6 Astra and 6.1 Sol; Google strongly recommends leaving it at default for Gemini 3.x
“One giant master prompt is efficient”Depends on your pipelineGoogle recommends chaining for multi-step workflows you build; Anthropic says chain only when you need to inspect intermediate output
“Add ‘never use bullet points’ for cleaner output”Model-specificHelpful for models that over-format; harmful on Fable 5.1 per Anthropic

Will any of this still hold in six months?

An important question is whether this advice will last, and the honest answer is: some of it, and you can tell which.

Perishable (expect to re-check every model release): effort-level defaults, how many examples to use, formatting defaults, which parameters are supported, exact model names. The vendors themselves say so. OpenAI tells you to treat each new model as a new family; Anthropic tells you to re-run an effort sweep even when moving from Fable 5 to 5.1, because the level names don’t map to the same amount of thinking.

Durable (has survived every generation in these docs): know what done looks like before you type; supply the context the model can’t see; state boundaries that matter; put long material before the question; verify anything consequential with a concrete check.

The test I’d apply: if a piece of advice describes a model behavior, it expires. If it describes how to communicate a task to any capable collaborator, it doesn’t.


What I could not verify

  • No independent evidence isolates prompt style as the cause of better output across vendors. All effect sizes cited (such as Anthropic’s “up to 30%”) are vendor-reported.
  • OpenAI’s position on examples was not addressed in the pages I read, so the table above says nothing about it rather than guessing.
  • Behavior on your task may differ. Vendor docs describe the vendor’s own test conditions. Your inputs are the only ground truth that counts.
  • The Prompt Debt Audit is untested. It is a structured way to apply what the vendors recommend. I have no data showing it improves outcomes.

Frequently asked questions

How long should a prompt be in 2026? No vendor publishes a target length. Anthropic’s and OpenAI’s guidance points to length being driven by what the model cannot infer, not by a word count. A short prompt with a clear outcome and one boundary is a fine default; long context (documents, data) is a separate matter and is well supported.

Do I still need “prompt engineering” as a skill? The vendors’ documentation describes the skill shifting from clever phrasing toward clear specification: outcome, context, boundaries, and a check. That is closer to writing a good brief than to writing code.

Should I write different prompts for ChatGPT, Claude, and Gemini? Yes, at the margin. The core (outcome, context, boundary) transfers. The corrections do not: OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.1 have opposite formatting tendencies, and Gemini 3 asks you to state verbosity explicitly.

Do examples still help? For matching a format, voice, or classification scheme, yes, and Google and Anthropic’s docs both support it. For open-ended judgment on the newest Claude models, Anthropic staff reports they can constrain the output. Test both on your task.

Does “think step by step” still work? It’s unnecessary for models that reason by default, and none of the three vendors recommends scripting the reasoning. A general nudge such as “think thoroughly before answering” is what Anthropic’s guide suggests instead.

How do I get accurate answers about recent events or products? Use a search-connected mode. Anthropic notes that Fable 5.1 at low effort may answer from memory instead of searching, and suggests verifying names in fast-moving fields. Google recommends Search grounding for recent facts. OpenAI’s guidance recommends stopping rules for retrieval so the model knows when it has enough evidence.

Will this guide be out of date soon? Parts of it, by design. See the section above on what is perishable. The update log below records what changed and when.


Sources

Read and cited on September 30, 2026.

  1. OpenAI, Using GPT-6 (developer documentation): https://developers.openai.com/api/docs/guides/latest-model
  2. OpenAI, Prompt engineering and API changelog (developer documentation): https://developers.openai.com/api/docs/guides/prompt-engineering
  3. Anthropic, Prompting best practices (Claude Platform Docs): https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
  4. Anthropic, Prompting Claude Fable 5.1 (Claude Platform Docs): https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1
  5. Google, Prompt design strategies (Gemini API docs; last updated Sept 17, 2026): https://ai.google.dev/gemini-api/docs/prompting-strategies
  6. Jonathan Kemper, “OpenAI’s new prompting guide tells users to stop overthinking and start with the result,” The Decoder, July 13, 2026: https://the-decoder.com/openais-new-prompting-guide-tells-users-to-stop-overthinking-and-start-with-the-result/
  7. Matthias Bastian, “Anthropic says it cut 80 percent of Claude Code’s system prompt…,” The Decoder, July 2, 2026 (reporting remarks by Anthropic’s Tariq Shihipar): https://the-decoder.com/anthropic-says-it-cut-80-percent-of-claude-codes-system-prompt-because-fable-5-models-want-a-smaller-system-prompt/
  8. Google Firebase, Prompt design (notice on Gemini 2.0 Flash shutdown): https://firebase.google.com/docs/ai-logic/prompt-design

Keep reading on BestPrompt.art: 25 Prompt Templates Worth Stealing This Month · The 30-Day Prompt Challenge · Prompt Writing Tips & Techniques · High-Converting AI Prompts


Update log

  • September 30, 2026: Full rebuild against current vendor documentation. Replaced the GPT-4o / o3 / Claude 3.x / Gemini 1.5 comparison with the current model families. Corrected the claims on role prompts, examples, and sequential steps. Removed the 20–60-word chart and opening anecdote, which could not be documented. Added the source-ranking section, the disagreement analysis, and the Prompt Debt Audit.
  • April 12, 2026: First published.

Byline: Tom Morgan. Research AI-assisted, disclosed above. No original testing was performed for this article.

Leave a Reply

Your email address will not be published. Required fields are marked *