ONLINEAGENT_OPS 2026.Q3 HOME ARTICLES CRAFT RECORD BLOG MAP HUBS FAQ SEARCH
HOMETHE CRAFTCHEAT SHEETSCHATGPT · CLAUDE · GEMINI
CHEAT SHEET · OPENAI · ANTHROPIC · GOOGLE · LARGE LANGUAGE MODELS

LLM Prompting Cheat Sheet

What OpenAI, Anthropic and Google actually publish — long documents first, question last, examples always, and no step-by-step for reasoning models.

READ5 min
WORDS899
SECTIONS5
SOURCES5
TYPECHEAT SHEET
CHECKED16 SEP 26

All three vendors publish prompting guides for their language models. They agree on less than people think. This sheet quotes each one and says where they differ.

TL;DR — THE SHORT VERSION
  • Reasoning and GPT models want different prompts. OpenAI: “reasoning models will provide better results on tasks with only high-level guidance. This differs from GPT models, which benefit from very precise instructions.”
  • Skip "think step by step" on reasoning models. OpenAI says it “may not enhance performance (and can sometimes hinder it)”.
  • Anthropic: long documents near the top, question at the end. “Queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs.”
  • Google: always give examples. “We recommend to always include few-shot examples in your prompts.”
  • Google: leave sampling parameters alone on Gemini 3.x — temperature, top_p and top_k were deprecated in the Gemini API on 21 July 2026.Google, Gemini API changelog, 21 Jul 2026, read at source 16 Sep 2026: “The sampling parameters temperature, top_p and top_k are now deprecated.” ai.google.dev
  • OpenAI: no sampling knobs on GPT-6 Astra. Custom temperature and top_p are not supported, and neither are log probabilities.OpenAI, API changelog, 3 Sep 2026, read at source 22 Sep 2026: “GPT-6 Astra does not support custom temperature or top_p values or log probabilities (logprobs).”
  • Anthropic: 3–5 examples. “Include 3–5 examples for best results.”
VENDOR-DOCUMENTED

Every rule on this sheet is quoted from the vendor's own published documentation, linked where it is used.

The formula

Long documents first + Role + Instructions + Examples + Question last

Reasoning: assembled on 16 Sep 2026 from the three vendors' guides linked below; no single vendor publishes this exact order.

Template:

[Long documents or data — near the top]

[Role: who the model is]
[Instructions: rules, format, what to exclude]
[3–5 examples, each in <example> tags]

[Your question — last]

Worked example:

<documents>…20k tokens of source material…</documents>

You are a technical editor for a general audience. Summarise the argument in the documents above in under 200 words, in prose, with no headers and no bullet points. Do not explain your process.

Breakdown — what each part does

01

Long documents

Anthropic: “Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.”

02

Role

Anthropic: “Setting a role in the system prompt focuses Claude's behavior and tone for your use case.” OpenAI's equivalent is the Identity section: “Describe the purpose, communication style, and high-level goals of the assistant.”

03

Instructions

Anthropic's test: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too.”

04

Examples

Google: “Prompts without few-shot examples are likely to be less effective.” Anthropic: “Wrap examples in <example> tags (multiple examples in <examples> tags) so Claude can distinguish them from instructions.”

05

Structure

OpenAI: “Use delimiters like markdown, XML tags, and section titles to clearly indicate distinct parts of the input”. Anthropic: “XML tags help Claude parse complex prompts unambiguously”.

Does it take a negative prompt?

No separate field — exclusions go in the instructions.NO FIELD

OpenAI's own template for instructions asks: “What should the model do, and what should the model never do?” Answer the second half plainly: "No filler phrases. No exclamation marks. No introductory sentence."

Corrected 16 Sep 2026: an earlier version of this sheet showed a YES badge beside the words "No field", and said negative constraints work "because they override strong default tendencies". The badge was wrong, and the explanation is not in OpenAI's or Anthropic's prompting guides; both were fixed.

TAKEAWAY

Documents first, question last, examples always — and on a reasoning model, give the goal rather than the steps.

What not to do

Do not tell a reasoning model to think step by step. OpenAI: “Since these models perform reasoning internally, prompting them to “think step by step” or “explain your reasoning” is unnecessary.”
Do not paste your question above a long document.
Do not change temperature, top_p or top_k on Gemini 3.x. Google: “Changing these parameters (for example, setting the temperature below 1.0) can cause unexpected behavior, such as looping or degraded performance, particularly in complex mathematical or reasoning tasks.”
Do not send a custom temperature or top_p to GPT-6 Astra. OpenAI: it “does not support custom temperature or top_p values or log probabilities (logprobs).”OpenAI, API changelog, 3 Sep 2026, read at source 22 Sep 2026: “GPT-6 Astra does not support custom temperature or top_p values or log probabilities (logprobs).”
Do not give one example and expect a pattern. Anthropic suggests 3–5.
Do not treat one vendor's order as universal. OpenAI says its own order “may vary by which model you are using”.
WHERE THE VENDORS DIFFER

OpenAI's developer message runs Identity, Instructions, Examples, then Context, “usually in this order (though the exact optimal content and order may vary by which model you are using)”. Anthropic puts long documents near the top and the query last. They describe different things — message sections versus document placement — but they pull in different directions if either is treated as universal. Follow the vendor whose model you are prompting.

TWO ORDERS · PULLING OPPOSITE WAYS
OpenAI’s developer message usually ends with context; Anthropic puts long documents at the top. They describe different things, so neither is universal.
OPENAIMESSAGE SECTIONSANTHROPICDOCUMENT PLACEMENTIdentityInstructionsExamplesContextLong documentsInstructions, examplesQuery, lastTOP TO BOTTOM
Reasoning — summarises this sheet’s box on where the vendors differ, from OpenAI’s and Anthropic’s prompting guides, both read at source 16 Sep 2026.

Free or paid? What each chat plan actually gets you is on free vs paid AI chat, and the no-subscription route is running AI free on your own computer.

Where this model sits among the others: which models take a negative prompt, and what each vendor publishes as its formula.

The behavioural differences between ChatGPT, Claude and Gemini are on prompting for LLMs.

The checklist

Before you send it:

1 — Reasoning model or GPT-style model?
2 — Long documents near the top, question at the bottom?
3 — Role stated?
4 — 3–5 varied examples included?
5 — What the model must never do, stated plainly?
6 — Sampling parameters left at defaults on Gemini?
7 — Would a colleague with no context follow this prompt?
SOURCES

OpenAI, Prompt engineering · OpenAI, Reasoning best practices · Anthropic, Prompting best practices · Google, Prompt design strategies (read in a browser) · Google, Gemini API changelog All read at source on 16 Sep 2026.

Model versions and vendor documentation both move. Re-read the source before relying on a specific number.

Every other sheet in this set: all model cheat sheets.

ABOUTMETHODVERIFYPRIVACYCONTACTINDEXAI PROMPT GENEER · EVERY ARTICLE CARRIES ITS OWN CHECKED DATE