How to write prompts for Google Gemini: the PTCF framework

Most disappointing AI answers are not a model problem. They are a prompt problem. When you ask Gemini to write an article about marketing, you are leaving almost every decision that matters, the audience, the angle, the length, the tone, to a coin toss. The PTCF framework fixes this by giving you four slots to fill before you hit send, so the model spends its effort on your actual intent rather than guessing it. This guide explains each part, shows it in action, and covers the advanced moves and mistakes that separate a usable answer from a great one.

What is the PTCF framework?

PTCF is a four-part prompt structure promoted in Google's own Gemini guidance: Persona, Task, Context and Format. Each part answers a different question the model would otherwise have to assume. Persona answers who the model is being. Task answers what it should produce. Context answers what it needs to know. Format answers what the finished thing should look like. Fill all four and you have removed most of the ambiguity that drags an answer toward the generic middle.

You do not have to label the parts inside the prompt, although doing so helps Gemini parse longer requests cleanly. The discipline is in the thinking: before you ask, decide each of the four, then write them down. Google's prompting guidance for Workspace puts the average effective prompt at around 21 words, which tells you the goal is precision rather than length (Google Workspace, 2025).

If you have come across PARTS instead, that is a related five element framework taught in Google educator training. We break it down in what PARTS stands for in Gemini prompts.

How do Persona and Task work?

The Persona sets the role, expertise level and voice the model adopts. Telling Gemini to respond as an experienced conversion copywriter writing for small business owners shifts its vocabulary, assumptions and depth before it writes a single sentence. A good persona is specific about field and seniority, not just a job title.

The Task is the explicit action and deliverable. This is where most prompts quietly fail, because verbs like help with or look at are not deliverables. Name the output, the scope and any limits: draft three subject lines under fifty characters, summarise this report in five bullet points, rewrite this paragraph at a year-eight reading level. The clearer the verb and the boundary, the less the model has to invent.

Why do Context and Format matter so much?

Context is the background that makes the output yours rather than anyone's. It covers your audience, your goal, your constraints and any reference material the model should lean on. Gemini supports very large context windows, so you can paste a brief, a transcript or a full document and ask it to work from that source instead of from generic training knowledge. The more relevant context you supply, the less the model fills gaps with assumptions.

Format controls the shape of the response: a table, a numbered list, a short email, valid JSON, a particular tone, a word count. If you do not specify it, the model picks one for you, and it rarely picks the one you needed. Stating the format up front saves a round of back-and-forth and makes the output easy to drop straight into your workflow.

What does a PTCF prompt look like in practice?

Compare a weak prompt with a PTCF version of the same request. The weak one is write something about our new coffee subscription. The PTCF version names all four parts and leaves almost nothing to chance:

Persona: You are a friendly email marketer for a specialty coffee brand.
Task: Write a launch email announcing our new monthly coffee subscription.
Context: Audience is existing customers in Australia who buy single bags.
   Benefit is fresh roasts delivered monthly with free shipping over $40.
Format: 120 words, warm and conversational, one clear call to action,
   subject line under 45 characters.

The second prompt will produce a near-usable draft on the first try, because every decision the model needs has already been made. That is the whole point of PTCF: you move the thinking to the front, so the output lands closer to done.

What advanced techniques lift results further?

PTCF gets you most of the way. A few extra techniques handle the harder tasks:

• Few-shot prompting. Include two to five example input and output pairs so the model copies a pattern instead of inferring it. This wins whenever format or tone consistency matters.
• Chain-of-thought. Ask the model to work through the problem step by step. Research on this method shows meaningful gains in accuracy on multi-step reasoning tasks (arXiv, 2022).
• Prompt chaining. Split a big job into stages and feed each output into the next prompt, rather than overloading one request.
• Iterative refinement. Start simple, read the result, change one element, and test again so you can see what actually moved the output.

What mistakes should you avoid?

Five habits cause most weak Gemini outputs, and all of them are easy to fix once you notice them:

1. 1.Vague instructions. Write an article about marketing has no audience, angle or length. Name them.
2. 2.Missing constraints. Without a word count, tone or boundary, the model invents its own.
3. 3.Too little context. The model cannot use a brief you never gave it, so paste the source.
4. 4.Overloaded prompts. Ten nested instructions in one request bury the priority. Chain instead.
5. 5.Ignoring documents. Gemini can read images, PDFs and long files, so use them rather than retyping.

Get into the habit of asking one question before you send a prompt: have I named the persona, task, context and format? If any slot is empty, that is usually where the next disappointing answer is hiding.