← AI Workflow Guides

From Vague Ask to Clear Prompt

Turn fuzzy ideas into useful task briefs by naming the goal, evidence, constraints, approval boundary, and output shape before you ask the model to do the work.

Level Beginner
Time 12 minutes
Tools covered: Claude , ChatGPT , Gemini
prompting clarity briefing getting-started output-format context-packs
Updated April 4, 2026

What This Guide Is For

Many people say “AI does not work” when what they really mean is “I gave it a fuzzy task and hoped it would guess the missing parts.” That is not a character flaw. It is usually a task-packaging problem.

Freshness note: This playbook was reviewed against current OpenAI, Anthropic, and Google prompting guidance on April 4, 2026.

Use this guide when your first instinct is to type something short like:

  • “make this better”
  • “write a summary”
  • “help with this code”
  • “give me ideas”

Those prompts are not useless, but they force the model to invent too much of the task on your behalf.

The Five Parts A Clear Prompt Usually Needs

1. Goal

State what success looks like.

Bad:

Help me with my meeting notes.

Better:

Turn these meeting notes into an action plan with owners, due dates, and unresolved decisions.

2. Context

Tell the model what situation it is operating inside.

Examples:

  • who the audience is
  • what material it should use
  • what stage the work is in
  • whether this is a draft, plan, critique, or final deliverable

For many modern workflows, context is really an evidence pack:

  • meeting notes
  • linked docs
  • current status
  • source records
  • logs, diffs, or tickets

3. Constraints

Name the limits early instead of correcting them later.

Useful constraints include:

  • length
  • tone
  • no invented facts
  • no code changes yet
  • preserve existing structure
  • use bullet points, table, JSON, or plain prose
  • only use the supplied evidence
  • stop before writeback

4. Approval Boundary

Name what the model may draft versus what a human must approve.

Examples:

  • “Draft the status update, but do not send it.”
  • “Propose the task list, but do not assign owners automatically.”
  • “Return the change plan only. Do not implement yet.”

5. Output Shape

If you already know the shape you want, ask for it.

Examples:

  • “Return exactly 5 bullets.”
  • “Use a table with risk, owner, and next step.”
  • “Give one recommendation, two tradeoffs, and three follow-up questions.”

The more reviewable the output needs to be, the more explicit the structure should be.

A Fast Rewrite Formula

When your first prompt is too vague, rewrite it using this frame:

I need help with [goal].

Evidence and context:
- [what this is]
- [who it is for]
- [what material the model should use]

Constraints:
- [must do]
- [must not do]

Approval boundary:
- [what the model may draft]
- [what still needs human approval]

Please return:
1. [section or output shape]
2. [section or output shape]
3. [section or output shape]

This is enough for a surprising number of tasks. You do not need exotic prompt tricks before you can get much better results.

Before And After Examples

Business

Weak:

Write an update for stakeholders.

Clearer:

Write a stakeholder update for a cross-functional product launch.

Evidence and context:
- audience: product, engineering, support leads
- source material: notes below
- goal: communicate progress, risks, and next decisions

Constraints:
- keep it under 250 words
- do not invent dates or commitments

Approval boundary:
- draft only
- do not imply final approval or send anything

Please return:
1. subject line
2. concise update
3. open risks
4. explicit asks

Development

Weak:

Help debug this bug.

Clearer:

Help debug a checkout regression in a TypeScript service.

Evidence and context:
- error only appears in staging
- recent change set included a pricing refactor
- stack trace and reproduction steps are below

Constraints:
- do not jump straight to a fix
- rank hypotheses first

Approval boundary:
- return the diagnosis and lowest-risk fix path
- do not suggest shipping or merging anything yet

Please return:
1. incident summary
2. top hypotheses with confidence
3. fastest evidence path
4. minimal-risk fix path

Writing

Weak:

Summarize this article.

Clearer:

Summarize this article for non-technical executives.

Evidence and context:
- they need the business impact, not every detail
- source text is below

Constraints:
- under 180 words
- avoid jargon

Approval boundary:
- summary only
- do not add recommendations not supported by the source

Please return:
1. key takeaway
2. three important implications
3. one decision question we should discuss next

When To Stop Prompting And Gather More Input

Sometimes the prompt is not the real problem. Stop and gather more context if:

  • you do not know the audience
  • you do not know what “better” means
  • the source material is missing
  • several different tasks are bundled into one request
  • you would reject the output no matter what because the brief itself is unclear

In those moments, the right move is not “prompt harder.” It is “clarify the task first.”