AI Prompt Writing Formula: 6 Parts for Better Content Drafts

If your AI drafts keep coming back bland, the prompt is usually missing information the model cannot guess: who it is writing for, what it is allowed to say, and what the finished piece should look like. A good prompt is not clever wording. It is a short set of decisions written down in a predictable order.

This guide gives you a six-part prompt writing formula built for content work (blog sections, emails, product copy, social posts), shows a weak and a strong version of the same prompt side by side, and ends with a template you can copy.

The Short Answer

A reliable AI prompt for writing has six parts, in this order: Role (who the model should write as), Task (one clear deliverable), Context (the reader and why the piece exists), Sources (the facts it must use, clearly separated from your instructions), Limits (length, tone and the claims it must not make) and Format (the exact shape of the output, ideally with a short example). Leave any one out and the model fills the gap with the most average answer it can find.

Where This Formula Comes From

None of the six parts is our invention. They are the points the major AI providers keep repeating in their own documentation, arranged into one checklist for writers:

• Google's Prompting Guide 101 for Gemini in Workspace organises effective prompts around four areas: persona, task, context and format.
• OpenAI's best practices for prompt engineering recommend putting instructions at the start, separating them from the text with markers such as ### or """, and being specific about context, outcome, length, format and style.
• Anthropic's Claude prompting best practices stress being clear and direct, giving the model a role, using examples, structuring inputs with tags, and telling the model what to do instead of what not to do.
• Google's Gemini API prompting strategies cover input, constraints, response format, few-shot examples, added context and breaking prompts into components.

The two parts writers most often skip are Sources and Limits. They are also the two that protect you from confident, wrong copy.

The 6-Part AI Prompt Writing Formula

1. Role: who is writing

One sentence that sets expertise and voice. "You are an experienced B2B copywriter who writes plainly for busy finance managers" does more than "You are a world-class expert". The role narrows vocabulary and tone; it does not add knowledge the model lacks, so it never replaces sources.

2. Task: one deliverable, one verb

Say exactly what you want back: "Write the 'How to choose' section of a blog post", not "help me with my blog". If you need an outline, a draft and a meta description, ask for them in separate turns. Smaller, single-purpose prompts are easier to check and easier to fix, which is the same reason we recommend generating a post in sections rather than all at once.

3. Context: the reader and the reason

Describe the specific reader, what they already know, what they are trying to decide, and where this piece sits (a blog post, a cold email, a landing page). This is the part a good brief has already settled. If you have one, paste the relevant lines from it; if you do not, our AI content brief checklist covers what to decide first.

4. Sources: the facts it must use

Paste the facts, figures, quotes, product details or notes the piece should rely on, and put them inside clear delimiters so the model can tell your material apart from your instructions. Triple quotes, ### lines or simple tags such as <notes> and </notes> all work. Then add one instruction: use only these facts for numbers, prices, dates and claims about the product. In our experience this single line removes most invented statistics.

5. Limits: length, tone and no-go claims

Give a word range, a reading level and a tone, then list what the piece must not do: no guarantees, no "best in the market", no medical or legal advice, no competitor names. Where you can, phrase limits positively, as both OpenAI and Anthropic suggest. "Write in short, plain sentences a non-native English reader can follow" works better than "don't be complicated".

6. Format: the shape of the output

Specify headings, bullet use, paragraph length and anything that must appear (a call to action, a table, a one-line summary). If the format matters, show a tiny example of it. Examples steer models strongly, so make sure the example has the structure you want, not wording you want copied.

A Weak Prompt vs a Formula Prompt

Here is the same request written two ways.

Weak:

Write a blog post about choosing accounting software for small businesses. Make it SEO friendly.

Using the formula:

Role: You are a practical small-business writer who explains software
without jargon.

Task: Write the section "How to choose accounting software" for a blog post.

Context: Readers are owners of 2-10 person service businesses in India who
currently use spreadsheets. They want to know what to compare before
starting a free trial. This section comes after an intro on why
spreadsheets break down.

Sources (use only these for any facts or numbers):
"""
- Our customers most often ask about GST invoicing, bank feeds and
  multi-user access.
- Typical setup time we see: one afternoon for basic invoicing.
"""

Limits: 350-450 words. Plain, friendly tone. Do not name or rank specific
brands. Do not promise savings or compliance outcomes. If a point needs a
fact that is not in the sources, write [CHECK] instead of guessing.

Format: One H2, then 5 H3 criteria, each with 2-3 sentences. End with a
one-sentence tip on running a trial with real invoices.

The second prompt is longer, but it takes a few minutes to write and it removes the usual rework: the generic intro, the invented "studies show" figure, the brand list you cannot verify, and the wrong length. (The sources block here is an illustration; use your own business notes.)

Copy-and-Paste Template

Role: You are a [type of writer] who writes for [audience] in a [tone] voice.
Task: Write [one specific deliverable].
Context: The reader is [who], knows [what], and wants to [decide/do what].
This piece is [where it will be used and what comes before/after it].
Sources (use only these for facts, numbers and product claims):
"""
[paste notes, data, quotes]
"""
Limits: [word range]. [Reading level]. Do not [claims to avoid].
If a fact is missing, write [CHECK].
Format: [headings / bullets / paragraph length / required ending].
Example of the format: [short example, optional]

How to Improve a Prompt That Still Misses

When the output is off, change one part at a time so you know what fixed it. A quick diagnostic for writers:

• Too generic? Context is thin. Name the reader and their situation more precisely.
• Made-up facts? Sources are missing or the "use only these" line is absent.
• Wrong tone? Strengthen the Role and add a two-sentence sample of the voice you want.
• Wrong length or structure? Tighten Format and give a skeleton example.
• Says things you cannot publish? Add those claims to Limits, phrased as what to write instead.

Save prompts that work as templates. Over a few weeks you build a small library for your recurring jobs (product descriptions, newsletter intros, FAQ answers), which is one of the habits that genuinely speeds up writing.

Common Mistakes

• Asking for a finished article in one prompt. Long single-shot outputs drift and repeat themselves. Outline first, then draft section by section.
• Relying on "act as an expert". A role shapes voice; it does not supply facts. Expertise has to come from your sources and your review.
• Mixing instructions into pasted text. Without delimiters, the model may treat part of your notes as instructions, or your instructions as content.
• Skipping review because the prompt was good. A strong prompt reduces errors; it does not remove them. Every factual line still needs a fact-check, and the draft still needs a human edit for voice (see our AI content editing workflow).

FAQ

Does the same prompt formula work in ChatGPT, Claude and Gemini?

Yes. All three providers' guides recommend clear tasks, context, examples and explicit output formats, so the six parts transfer. Tools differ in style and defaults, which is why it is worth testing an important prompt in more than one; we compare their writing behaviour in ChatGPT vs Claude for writing.

How long should a writing prompt be?

As long as it needs to be to remove guesswork. For a single blog section, 100 to 200 words of prompt plus your pasted sources is a sensible range. If the prompt is longer than the output you want, split the task.

Is the order of the six parts important?

Instructions first and pasted material clearly separated is the part that matters most. Keeping the same order every time also makes your prompts easier to reuse and compare.

Will a good prompt make AI content rank on Google?

No prompt guarantees rankings. Google's guidance is that it rewards helpful, reliable content however it is produced. A better prompt simply gives you a stronger first draft to add your own experience to, which we cover in AI content and SEO.

Sources

• Google Workspace: Prompting guide 101: writing effective prompts
• OpenAI Help Center: Best practices for prompt engineering with the OpenAI API
• Anthropic: Prompting best practices
• Google AI for Developers: Prompt design strategies (Gemini API)