# Designer + AI workflow

AI Code Design · Practical planning guide

## 1. Frame the problem

Input: An anonymized brief, observed difficulty, constraints, and the decision you need to make.

AI can help: Turn your notes into candidate problem statements, identify gaps, and suggest questions.

You own: Choose the real audience and verify the problem with evidence. AI-generated personas are hypotheses, not research participants.

Output: One user, one task, a scope boundary, and observable success criteria.

Review: Can you distinguish observed evidence from assumptions and name what you will not build?

Prompt: Separate supplied evidence from assumptions. Suggest three narrower problem statements and questions to validate them. Do not invent interviews, demand, or metrics.

## 2. Explore alternatives

Input: The accepted brief, constraints, and examples of what users need to understand.

AI can help: Generate different flows, sketch directions, content structures, and tradeoffs.

You own: Select a direction for a stated reason. Check whether familiar patterns fit this audience and context.

Output: Two or three alternatives and your decision record.

Review: Have you compared materially different approaches instead of cosmetic variations?

Prompt: Propose three different ways to complete the primary task. For each, state the tradeoff, failure case, and assumption to test. Recommend an experiment, not a claim of validation.

## 3. Specify screens and behavior

Input: Design references, tokens, sample records, and interaction rules.

AI can help: Translate your design into component states, draft microcopy, and implementation requirements.

You own: Own hierarchy, accessibility, scope, and rules. Explain which visual decisions must be preserved.

Output: A state inventory: empty, loading when relevant, success, invalid input, no results, and failure.

Review: Does every visible action have a defined result, error response, and data boundary?

Prompt: Create a behavior specification for the primary task. List inputs, rules, state changes, keyboard behavior, error recovery, mobile layout, and what survives a reload. Mark unspecified decisions as questions.

## 4. Build one complete interaction

Input: The approved behavior specification and the relevant source files, without credentials.

AI can help: Implement a small change, explain the relevant files, and propose verification steps.

You own: Review the change, run it, inspect the result, and keep a recoverable source checkpoint.

Output: One working flow from input through result, plus a record of the checks performed.

Review: Can you reproduce success and failure yourself, and explain where the data lives?

Prompt: Implement only the smallest complete flow first. Explain the files and how to run it. Preserve existing behavior. Include invalid-input and reload checks. Label simulated integrations; do not add external services or dependencies without explaining why they are needed.

## 5. Debug with evidence

Input: Expected versus actual behavior, reproduction steps, relevant code, and sanitized error messages.

AI can help: Trace the reported failure, suggest competing causes, and propose a minimal repair.

You own: Reproduce the issue, review the evidence, and rerun the original and adjacent cases.

Output: A small fix and a regression check tied to the original failure.

Review: Did the original failure stop occurring without breaking a nearby interaction?

Prompt: Do not guess from appearance alone. Identify the boundary where expected behavior stops, rank likely causes using the evidence, and propose the smallest fix. State what information is still missing and how to verify the repair.

## 6. Test with people

Input: The task and prototype limitations; afterward, actual anonymized observations.

AI can help: Draft neutral task scripts and organize anonymized observations into themes and open questions.

You own: Recruit appropriate consenting testers, observe real behavior, and decide what the evidence supports.

Output: Observed problems, separate interpretations, a prioritized change, and a recheck.

Review: Are findings traceable to actual observations? Did you record help given and limitations?

Prompt: Draft a task-based test without telling the participant which buttons to press. Separate observation from interpretation. Do not simulate participants or invent results. Suggest how to recheck one change after the session.

## 7. Release and tell the story

Input: Actual checks, known limitations, the source checkpoint, and your personal contribution.

AI can help: Draft release notes, a maintenance checklist, and a case-study outline from supplied evidence.

You own: Decide readiness, confirm ownership and data handling, verify the deployed URL, and approve public claims.

Output: A release decision, maintenance owner, working link, and honest portfolio story.

Review: Do public claims match implemented behavior and observed results? Is rollback documented?

Prompt: Use only the supplied evidence to draft release notes and a case-study outline. Distinguish prototype limitations, planned features, actual results, and AI assistance. Identify missing release decisions and leave missing outcomes blank.