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

LLM Prompting Cheat Sheet

What the three vendors actually publish — including the long-context rule worth up to 30 percent and why step-by-step is now conditional.

What the three vendors actually publish — including the long-context rule worth up to 30 percent and why step-by-step is now conditional.

TL;DR — THE SHORT VERSION
  • Reasoning models and fast models want opposite prompts. OpenAI: reasoning models do better with "only high-level guidance"; GPT-class models need "very precise instructions".
  • "Think step by step" is now conditional. Redundant on reasoning models, still useful on fast ones.
  • Anthropic: put long documents at the TOP, your question at the BOTTOM. Reported to improve quality by up to 30 percent on complex multi-document inputs.
  • Google: "always include few-shot examples" — prompts without them are "likely to be less effective".
  • Google: leave sampling parameters at defaults on Gemini 3.x rather than tuning them.
  • Anthropic: 3–5 examples, varied enough that the model does not latch onto an unintended pattern.
VENDOR-DOCUMENTED

This sheet quotes the vendor's own published prompting guidance.

The formula

Role + Context + Task + Constraints · with long documents first and the question last

Template:

[Long documents / data, if any — placed FIRST]

[Role: who the model is]
[Context: audience, purpose, constraints]
[Task: the specific ask]
[Format: length, structure, what to exclude]
[2–5 examples of the output you want]

[Your question — placed 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

First, above everything. Anthropic report up to 30 percent better quality on complex multi-document inputs when the query sits at the end instead of the start.

02

Role

Who the model is. All three vendors recommend it; it activates vocabulary and framing without extra instructions.

03

Context

Audience, purpose, what surrounds the task. Google's guidance is not to assume prior knowledge.

04

Task

The specific ask, stated plainly. Anthropic's test: show the prompt to a colleague with no context — if they would be confused, the model will be.

05

Examples

Google recommends always including them. Anthropic suggests 3–5, relevant and varied, wrapped in <example> tags.

06

Structure

XML tags to separate instructions, context and input. Anthropic's documented approach; OpenAI also recommend markdown headers and XML for delineation.

Does it take a negative prompt?

No field — exclusions go in the prompt or system message.YES

State them plainly and specifically: "No filler phrases. No exclamation marks. Never use the words 'transform', 'leverage', 'seamless'. No introductory sentence explaining what you are about to do." Negative constraints work well on LLMs precisely because they override strong default tendencies that positive instructions struggle to displace.

What not to do

Do not tell a reasoning model to think step by step — it already does, and the instruction can push it into over-deliberation.
Do not paste your question before a long document; put the document first.
Do not tune temperature on Gemini 3.x — Google recommend defaults.
Do not give one example and expect a pattern; give 3–5, varied.
Do not assume the vendors agree — OpenAI place context near the end of a system message, Anthropic place long documents at the very top. Follow whichever you are prompting.
WHERE THE VENDORS DIFFER

OpenAI's documented system-message order is identity, instructions, examples, then context near the end. Anthropic's long-context guidance puts documents at the very top with the query last. These are not interchangeable rules — they describe different things (message composition versus document placement) but they will pull you in different directions if you treat either as universal.

The checklist

Before you send it:

1 — Which kind of model — reasoning or fast?
2 — Long documents at the top, question at the bottom?
3 — Role stated?
4 — 3–5 varied examples included?
5 — Exclusions stated explicitly?
6 — Sampling parameters left at defaults on Gemini?
7 — Would a colleague with no context follow this prompt?
SOURCE

OpenAI prompt engineering guide, developers.openai.com · Anthropic prompting best practices, platform.claude.com · Google Gemini prompting strategies, ai.google.dev. All checked 25 Aug 2026.

Model versions and vendor documentation both move. Re-read the source before relying on a specific number. Errors are logged at corrections.

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

ABOUTMETHODVERIFYCORRECTIONSPRIVACYCONTACTINDEXAI PROMPT GENEER · CHECKED 22 AUG 2026