Context Management 102: เขียน task briefContext Management 102: write a task brief

A task brief is a small document that tells AI what job it is doing.

It does not need to be fancy. It needs to remove guessing.

Bad prompt:

Make this better.

Better brief:

Task: Rewrite this onboarding email.
Goal: Make new workshop learners feel clear about what to do before class.
Audience: Thai professionals who are new to AI tools.
Source: Use the draft below.
Constraints: Friendly, direct, no hype, keep under 180 words.
Output: Thai email with subject line and bullet checklist.

The second version gives the model a job, a reason, a reader, source material, boundaries, and a target shape.

The task brief template

Use this template when the task matters:

Task:
Goal:
Audience:
Source:
Current state:
Constraints:
Output format:
Definition of done:
Ask first if:

You will not always need every field. For quick tasks, use only the first five. For important tasks, fill more.

Task

Write the task as a verb.

Good tasks:

• summarize these notes
• compare these two options
• review this draft for unsupported claims
• turn this transcript into action items
• create a lesson outline from this source

Weak tasks:

• thoughts?
• improve
• analyze
• help me

Weak tasks are not forbidden. They just force the model to choose the job for you.

Goal

The goal explains what success means.

Goal: Help a beginner decide what to do next, not teach every detail.

This changes the answer. Without the goal, the model may write a textbook when you need a decision memo.

Audience

Audience changes language, depth, examples, and assumptions.

Audience: Thai small-business owners who use LINE and Google Sheets but do not code.

That is much more useful than:

Audience: everyone.

If there are two audiences, say which one matters most.

Source

Source tells the model what it should trust.

Source: Use only the pasted policy below. Do not add outside assumptions.

or:

Source: Use the meeting notes as primary source. Use your general knowledge only for structure, not facts.

This matters when hallucination would be expensive.

Constraints

Constraints are rules for the work.

Examples:

Constraints:
- Thai first
- no sales language
- keep product names unchanged
- avoid claims about certificates
- include uncertainty if the source does not say something

Good constraints are concrete. Vague constraints like "make it professional" are less useful unless you explain what professional means here.

Output format

Output format prevents cleanup work.

Output format:
- Title
- 3-sentence summary
- Table: issue, why it matters, suggested fix
- Final recommendation

If you want Markdown, JSON, table, email, checklist, or slides, say so.

Definition of done

Definition of done tells the model how to judge completeness.

Definition of done:
The answer is ready to paste into a learner handout without extra restructuring.

This is especially useful for agents and longer tasks.

Ask first if

Use this field when a wrong assumption would be costly.

Ask first if:
- the source does not include the target audience
- the requested format conflicts with the content
- you need current facts that are not in the source

This prevents the model from confidently filling gaps that should be human decisions.

Practice: upgrade three prompts

Rewrite these into task briefs:

1. Make this copy better.
2. Summarize this meeting.
3. What should I do with this spreadsheet?

For each one, include:

Task:
Goal:
Audience:
Source:
Constraints:
Output format:

Next: Context Management 103 - prepare a source bundle.