Write AI prompts that keep the assistant on-topic and task-focused

TL;DR — If an AI assistant keeps wandering, the usual cause is a vague request with no boundaries. Fix it by giving one clear goal, a strict output format, explicit "do not" rules, and a short checklist the assistant must satisfy before it answers. Reading time: ~5 min

Goal

When you finish, you will have a reusable instruction template you can paste into an AI assistant so its replies stay focused on your task, use the format you asked for, and avoid unrelated suggestions.

Prerequisites

• Access to any AI assistant chat box or API prompt field
• A specific task you want done, written in one sentence, for example: "Summarize this meeting in 5 bullet points"
• The source material the assistant should use, if any, for example: meeting notes, an email, or a draft document
• A text editor or notes app so you can save your reusable prompt template
• Optional: a way to compare outputs side by side, such as two browser tabs

Steps

Step 1: Write the task as one observable result

Paste this into your notes app and replace the text in brackets:

Goal: [exact result you want].
Done when: [what the final answer must contain].

Example:

Goal: Turn these meeting notes into a client update email.
Done when: the answer contains a subject line, a 120-180 word email body, and 3 action items.

What you should see when this step succeeds: you can point to one final deliverable, not a general topic.

Step 2: Add boundaries so the assistant knows what to ignore

Paste this directly under your goal and replace the bracketed text:

Use only the information I provide below.
Do not add facts, tools, recommendations, or background that are not in my material.
If information is missing, write: "Missing information: [item]".
Do not ask more than 3 clarifying questions.

What you should see when this step succeeds: your instructions now tell the assistant what not to do, not just what to do.

Step 3: Force a specific output format

Paste one of these blocks under the boundaries.

For bullet points:

Output format:
- 5 bullet points
- Each bullet 12-20 words
- Plain English
- No intro sentence
- No conclusion

For an email:

Output format:
Subject: [one line]
Email body: [120-180 words]
Action items:
1. [item]
2. [item]
3. [item]

For step-by-step instructions:

Output format:
1. [step]
2. [step]
3. [step]
Rules:
- Use numbered steps only
- Include exact menu paths or exact commands
- Do not include theory unless I ask

What you should see when this step succeeds: the format is detailed enough that two different people would expect nearly the same shape of answer.

Step 4: Tell the assistant who the audience is

Paste this block and replace the bracketed text:

Audience: [who will read this].
Assume they know: [what they already know].
Avoid jargon. If you must use a technical term, explain it in one short parenthesis.

Example:

Audience: a non-technical customer.
Assume they know: how to use a web browser and copy text.
Avoid jargon. If you must use a technical term, explain it in one short parenthesis.

What you should see when this step succeeds: the assistant has a reading level and context to aim for.

Step 5: Add a self-check before the assistant answers

Paste this block at the end of your instructions:

Before answering, check your draft against this list:
- Does it answer the exact goal?
- Does it stay within the provided material?
- Does it follow the output format exactly?
- Did it avoid extra advice not requested?
If any answer is no, revise before sending.

What you should see when this step succeeds: your prompt now tells the assistant to catch its own drift before you have to.

Step 6: Use the full copy-paste template

Paste this template into the AI assistant and replace every bracketed section:

You are helping with one task only.

Goal: [exact result you want].
Done when: [observable end state].

Audience: [who will read this].
Assume they know: [what they already know].

Use only the information I provide below.
Do not add facts, tools, recommendations, alternatives, or background unless I ask.
If information is missing, write: "Missing information: [item]".
Do not ask more than 3 clarifying questions.

Output format:
[write the exact structure you want]

Before answering, check your draft against this list:
- Does it answer the exact goal?
- Does it stay within the provided material?
- Does it follow the output format exactly?
- Did it avoid extra advice not requested?
If any answer is no, revise before sending.

Material:
[paste your notes, draft, transcript, or data here]

What you should see when this step succeeds: the assistant replies in the structure you asked for and stays close to your material.

Step 7: If it still wanders, send one correction message

If the first answer drifts, do not rewrite everything. Send this exact follow-up message:

Revise your last answer using these corrections only:
1. Remove anything not supported by my material.
2. Follow the output format exactly.
3. Do not add background, tips, or alternatives.
4. Keep all useful content that already fits the goal.
Return only the revised answer.

What you should see when this step succeeds: the second answer is shorter, more literal, and closer to your requested format.

Verify it works

Use this quick test in your AI assistant. First, paste this prompt:

You are helping with one task only.

Goal: Summarize the text below.
Done when: the answer contains exactly 3 bullet points.

Audience: a busy manager.
Assume they know: the project names.

Use only the information I provide below.
Do not add facts, recommendations, or background unless I ask.
If information is missing, write: "Missing information: [item]".
Do not ask clarifying questions.

Output format:
- 3 bullet points
- 10-15 words per bullet
- No intro
- No conclusion

Before answering, check your draft against this list:
- Does it answer the exact goal?
- Does it stay within the provided material?
- Does it follow the output format exactly?
- Did it avoid extra advice not requested?
If any answer is no, revise before sending.

Material:
Project Atlas shipped on Tuesday. Two customers reported login confusion, but no outage occurred. The team will update the onboarding screen on Friday.

You should see all of these results:

- Exactly 3 bullet points
- No extra paragraph before or after the bullets
- No invented outage, root cause, or recommendation
- Only facts from the material

If any of those checks fail, return to Step 3 and make the output format more specific.

Common pitfalls

The mistake: asking for a topic instead of a deliverable

Symptom: the assistant gives a broad explanation, history, or extra options. Fix: replace "Help with onboarding emails" with "Write a 150-word onboarding email with 3 bullet points."

The mistake: not saying what source material is allowed

Symptom: the assistant invents facts, adds best practices, or mixes in outside knowledge. Fix: add this line: "Use only the information I provide below."

The mistake: leaving the format too loose

Symptom: you get paragraphs when you wanted bullets, or advice when you wanted steps. Fix: specify the exact structure, for example: "Exactly 5 bullet points, 12-20 words each, no intro, no conclusion."

The mistake: putting multiple jobs in one prompt

Symptom: the assistant answers one part well and ignores or blends the rest. Fix: split the work into separate prompts, for example one prompt for summary, one prompt for rewrite, one prompt for action items.

The mistake: correcting by starting over from scratch

Symptom: each retry changes everything, including parts that were already good. Fix: send a revision message that says what to remove, what to keep, and that it must "Return only the revised answer."

The mistake: asking for "concise" or "professional" without defining it

Symptom: the style changes, but the answer still wanders. Fix: replace vague style words with measurable rules, for example: "120-180 words," "plain English," and "no background section."

This article was written by an AI system and published pending human review. Verify anything you intend to act on.