AI Prompting Formula

The prompt structure that produces consistent, high-quality output from any modern AI.

By James Schramko · Updated September 2026.

Background

I saw an Anthropic prompt engineer present this formula in a session on prompting Claude. It is the structure they recommend for consistent, high-quality output from any modern LLM. The order matters. The components matter. I have used it across content, sales and operator decisions. The difference in output quality is real.

This replaces the earlier RICESFQ playbook. The components are the same in spirit. The structure is tighter. The Q step (questions) stays as a James addition because it earns its place.

Why It Matters

The model does its best work when the prompt is built top-down: who it is, how to sound, what materials to use, what rules apply and what good looks like. Then what was said before, what you want right now, how to think and how to format, and how to start the reply.

Skip a layer, or put the layers in the wrong order, and the output drifts. The order is the discipline.

The Formula

Use these in this order, top to bottom, in your prompt.

1. Task context

Tell the model who it is and what it is doing in general. Role plus broad domain.

Example: "You are a content editor working with small business owners on plain-English copy."

2. Tone context

Voice, register, audience. Sets the personality of the reply.

Example: "Direct, confessional, no hype. Short sentences. Treat the reader as an experienced operator."

3. Background data, documents, and images

The reference material the model should use as source. Paste it here, before the rules and the request.

Example: "[paste source brief, transcript, or document]."

4. Detailed task description and rules

The specific constraints. What to do, what not to do. Rules go here, before the end.

Example: "Simplify the paragraph without changing meaning. Keep all numbers, names, and claims exactly as written. Do not add new ideas. No em dashes."

5. Examples

Show, do not only tell. One or two strong examples beat a paragraph of description.

Example: "[paste a paragraph in the target style that has worked before]."

6. Conversation history

Prior turns or carry-over context. If the prompt depends on what was discussed earlier, paste it here. If not, write "none" so the model knows.

7. Immediate task description or request

The specific ask, restated after all the context. This is what you want the model to do right now.

Example: "Rewrite the paragraph above in the target style."

8. Thinking mode on

Current models reason before they answer when thinking is switched on. In the chat apps it is a toggle. In the API the default depends on the model. On the Claude 5 models it is on when you leave the setting out. On older models you switch it on. Turn it on for anything that needs analysis. Writing "think step by step" into the prompt was the old way to get the same thing. It adds little now.

Example: "Before you rewrite, think through which sentences carry the meaning and which carry filler."

9. Output formatting

How to deliver the answer. Length, structure, what to include and exclude.

Example: "Return only the rewritten paragraph. No commentary. No headings."

10. Format lock (optional)

The old technique here was prefilling the first words of the reply so the model committed to a format. Claude 4.6 and later models do not support assistant-message prefill. Lock the format in the instruction instead. Name the structure and give one example of the exact output. Where your API supports schema enforcement (structured outputs on the Claude API), use it. The reply then has to match the schema.

Example (in API): start the assistant turn with "Here is the rewritten paragraph:" so the model continues from there instead of preambling.

11. Questions (James addition, not Anthropic)

Invite the model to ask before it produces the final output. Catches ambiguity and missing context.

Example: "If anything in the brief is missing or unclear, ask before producing the rewrite."

How to Use the Formula

Use it top to bottom. Order is the discipline.
One task per prompt. If you have two unrelated jobs, write two prompts.
Reference material goes early (step 3), never at the end.
Rules go before the request (step 4 before step 7), never after.
Higher reasoning effort (step 8) can help on hard tasks. It also adds wait time and may cost more. Use the lowest effort that gives a reliable result.
The Q step (step 11) is most useful when the brief is vague. It also helps when you are testing whether the model has enough to work with.

For everyday quick prompts, you do not need all 11 components. Use a short subset, in the same order. The discipline is the order, and a full set is optional.

Best Practices

Use verified, high-quality source material.
Never upload private or client data to a public AI platform.
Save prompts that work so you can reuse them and track versions.
Always check the facts before you publish or share AI-assisted content.
Treat outputs as drafts. A person checks them.

Worked Example

```

Task context: you are a content editor working with a small-business owner.

Tone context: plain English, short sentences, grade 6 readability.

Background data: [paste source text].

Detailed task description and rules:

Simplify the paragraph without changing meaning.
Keep all numbers, names, and claims exactly as written.
Do not add new ideas.
No em dashes.

Examples: [paste a paragraph in the target style].

Conversation history: none.

Immediate task: rewrite the paragraph above in the target style.

Thinking step: before you rewrite, think through which sentences carry the meaning and which carry filler.

Output formatting: return only the rewritten paragraph. No commentary. No headings.

Questions: if anything in the brief is missing or unclear, ask before producing the rewrite.

```

Platform Flexibility

The formula was developed on Claude. The order holds on any competent model. Two parts depend on the platform. Step 8 needs the model to have a thinking mode. Most current ones do, under different names. Step 10 depends on what the platform gives you to lock a format. Use schema enforcement where the API supports it. Everywhere else, use a strict instruction plus an example. Check both against the tool you are using rather than assuming.

Responsible AI Use

Treat outputs as drafts that a person checks. Protect privacy, check the data is right, and keep training materials and client data inside your own workspace.

Use this formula in order and you get precise, consistent results from any AI tool. Your data and your workflow stay intact.