Skip to content
06 · Field Guide

Briefing and collaborating with AI

The original six-part brief remains useful. For building software, add a seventh part: Validation.

The seven-part brief

QuestionSectionWhat it contributes
Who should the AI act as?RoleRelevant expertise and perspective
Why does this exist?ContextProduct and problem understanding
Who is it for?UserNeeds, situation, and capability
What should be produced?RequirementsScope, behaviour, and deliverables
What limits apply?ConstraintsBoundaries, risks, and exclusions
What does good look like?SuccessQuality and outcome
How will we check it?ValidationTests, inspection, and evidence

Strong brief:

strong brief · markdown
# Role
You are a senior product designer and frontend developer creating a small educational prototype.

# Context
Creative Recharge helps desk-based workers choose a short break based on their available time and energy.

# User
A busy person who feels mentally stuck and wants a suggestion in under 30 seconds.

# Requirements
Create a one-page flow: introduction → three questions → one recommendation → start again. Use local sample data only.

# Constraints
No sign-in, database, health claims, location access, or external API. Use semantic HTML and do not add decorative UI that slows the flow.

# Success
The experience feels calm, playful, and complete. A first-time user can reach a useful result without explanation.

# Validation
Check the flow using keyboard only, at 320px and 1440px widths, with reduced motion enabled, and confirm the production build succeeds.

Discipline-specific layers

The shared structure stays the same. Add the layer that carries your craft.

Product design — interaction model. Describe sequence, decisions, states, navigation, system feedback, error recovery, and what should remain persistent.

Visual design — visual direction. Describe hierarchy, grid, typography, palette, density, imagery, component character, and what to avoid. Translate taste into observable properties rather than stacking adjectives such as “modern, premium, clean.”

Content design — language system. Describe user knowledge, tone, terminology, reading level, message hierarchy, content patterns, and examples. Specify what the interface must help the person understand or do.

User research — evidence layer. Separate observed evidence from interpretation and assumption. Include research questions, sample limits, known behavioural signals, and what the prototype is intended to learn.

Separate thinking from building

Do not force planning, selection, implementation, and critique into one request. Use four moments:

  1. 01
    Inspect
    What exists? What is unclear?
  2. 02
    Propose
    What approaches are possible and what are the trade-offs?
  3. 03
    Choose
    Which direction are we taking and why?
  4. 04
    Implement
    Build the agreed slice and validate it.

Useful prompt:

prompt
Do not edit yet. Analyse the request and current project. List assumptions and open
questions, then propose two approaches with trade-offs. Recommend one. After I choose,
produce a short implementation plan.

Build in small slices

A vertical slice is one thin, complete path through the experience. For the home project, that might be: answer three questions and receive one recommendation. It is more valuable than building five beautiful disconnected screens.

After the slice works, add visual refinement, more suggestions, motion, sharing, or extra states one at a time.

Four useful feedback moves

Feedback becomes more effective when it names the observation, impact, and intended outcome:

Debugging prompts

“It doesn’t work” gives the agent very little evidence. Use this structure:

debugging brief · markdown
# Expected
What should happen?

# Actual
What happened instead?

# Reproduction
Exact steps, viewport, device/browser, and data.

# Evidence
Error text, console output, screenshot, or recording.

# Recent change
What changed immediately before the problem appeared?

# Request
Diagnose the root cause before editing. Propose the smallest safe fix. After
implementing, repeat the reproduction steps and report the result.