BuildSpark AI
Lesson 3 of 8Jump to Try it

Home · Start here · Lesson 3 of 8

Get better answers by asking better

Get an answer you can use on the first or second try.

Next on this goal, step 4 of 4: Know which parts of an answer you can rely on.

Lesson 3 of 8 · about 4–5 min reading + 4 min doing · last checked 2026-10-19

Recall

Recall (Lesson 2): What do you tell the tool to write when your note doesn't cover something?

Show the answerCheck your Lesson 2 answer before you build a full prompt.

"Not in the note." It stops the tool filling the gap with a guess.

The idea

The language model can only use what you give it.

A prompt that works states four things. They are the task, the facts it must use, the reader, and the shape of the answer. If the first reply misses, say what to change in the same chat.

Four slots and a follow-upA prompt card with four slots, Task, Facts, Reader and Shape. Facts says attach or paste the source. An arrow leads to Language model, then Reply. A curved arrow labelled Follow-up returns to the card.TaskFactsattach or pastethe sourceReaderShapeLanguage modelReplyFollow-up: say what to change
State the four slots. If the reply misses, say what to change.

The four slots

Task

Say what to make and what it's for. "Rewrite this note as an email to a client" works better than "help with this note".

Facts

Paste or attach the text the reply must use. Add a rule for gaps: "If it isn't in the text, write 'not in the text'." This is the slot that stops guessing.

Reader

Say who it's for, such as "a client project lead, not technical". The reader sets the tone and the words.

Shape

Say the length and layout, such as "3 bullets, under 50 words".

Google's guides use four parts too: persona (a role, such as "You are a program manager"), task, context and format. We use reader instead of persona, because who it's for sets the tone. Our "facts" is their context. A role can still help with tone. It just doesn't add facts.

Show the shape with a made-up example

Anthropic says, "Examples are one of the most reliable ways to steer Claude's output format, tone, and structure." Google says, "We recommend to always include few-shot examples in your prompts." Few-shot means a few examples of the answer you want. Make the example up. A real customer email would send that customer's details (Lesson 5).

Give it the document, not a description

Gemini Apps Help says you can add up to 10 files in one prompt ("subject to availability"). OpenAI Help says ChatGPT can "Find references to a topic or extract relevant quotes" from files you upload. Ask for page or section numbers, so you can check each point.

Follow up in the same chat

Google's Workspace guide says, "Make it a conversation… Use follow-up prompts and an iterative process of review and refinement to yield better results." Stay in the same chat, so the facts you gave are still there. Name the change: "shorter", "move the delay first", "less formal". "Be accurate" or "do better" add no facts and name no change. If two follow-ups don't fix a reply, the gap is usually in the facts. Add the missing text, then ask again.

EXAMPLE · the note and the date are inventedSee a vague prompt rebuilt in four slots: a shorter reply, no invented date.

Before

"Summarise the meeting." The reply runs to 300 words and adds "launch moved to June".

After

"Task: rewrite the note below for a teammate who missed the meeting. Facts: use only the note. If the note doesn't say, write 'not in the note'. Reader: busy, not technical. Shape: 3 decisions, then 1 open question." The reply is 60 words and adds nothing. One follow-up: "Shorter. Drop the open question."

Why it worked: the facts slot gave the tool something to use. The gap rule stopped the June guess. The reader and shape kept it short and plain.

Practice

Step 1 (done for you): The reply invented a deadline. Task, Reader and Shape were all stated.Step 2: Which slot was missing?

If the reply is… fix this slot

Reply problem Fix this slot Follow-up to send
Too long, or the wrong layout Shape "Same content, 3 bullets, under 50 words."
Wrong tone Reader "For an upset customer. Plain and polite."
Adds a fact you didn't give Facts (Lesson 2) "Use the text above. If it's not there, say so."
Answers a different question Task "The task is X, not Y."
Vague Facts Attach the source and ask for page numbers (Lesson 4).

Try it in your own AI tool (4 min)

Take a prompt you've sent before. Rewrite it as four lines: task, facts, reader, shape. Send it, then send one follow-up that fixes one thing. Save the prompt you end up with; Lesson 7 reuses it.

Show the model answerCompare your four-line prompt and follow-up with a filled-in example.

"Task: weekly update email. Facts: the 4 bullets below. Reader: client project lead, not technical. Shape: 3 short paragraphs, under 120 words." Follow-up: "Move the delay to the first sentence."

Check

Question 1 of 4. The reply is right, but three times too long for a text message. What's the best follow-up?

Question 2 of 4. Which prompt is most likely to get a refund reply you can send?

Question 3 of 4. A reply invents a meeting date. What do you add to the prompt?

Question 4 of 4. You need the payment terms from a 30-page supplier contract. Which request is most likely to give a reliable answer?

SourcesOpen the 7 source pages behind this lesson's claims, all checked 2026-10-19.
  1. Anthropic, Claude Platform Docs: "Prompting best practices"
  2. Google AI for Developers: "Prompt design strategies"
  3. Google, Gemini Apps Help: "Tips for creating custom Gems" (persona: "Tell your Gem what role to play")
  4. Google, Gemini Apps Help: "Upload & analyze files in Gemini Apps"
  5. Google Docs Editors Help: "Learn about prompts for Gemini in Docs, Sheets, Slides, Vids & Forms"
  6. Google Workspace: "Prompting guide 101"
  7. OpenAI Help Center: "Uploading files and audio to ChatGPT"