9-Phase PRD Workflow for AI Coding Tools

Updated: 2025-07-25

AI coding tools are capable, but they are literal. If your requirements are vague, inconsistent, or missing key constraints, the code you get back will reflect that.

This post is a planning workflow I use to turn either:

a messy idea dump, or
an existing PRD that is still too high level

into an implementation-ready specification that both humans and AI agents can follow.

Key takeaways

The workflow is 9 phases and 9 artifacts. Each phase produces one concrete output you can file, review, and reuse.
You will write less code twice if you define entities, constraints, and acceptance criteria before prompting an AI.
After Phase 4, stop and define your MVP so you do not overload context windows and scope.
You can copy the prompt templates below into any AI tool. They are intentionally tool-agnostic.

Fixes when it breaks. Workflows when it doesn't.

OpenClaw guides, configs, and troubleshooting notes. Every two weeks.

What this workflow is (and is not)

What it is: a repeatable way to create requirements that are specific enough for AI coding assistants to implement with minimal back and forth.

What it is not: a guarantee that AI will write perfect code. The goal is to reduce ambiguity so that failures are obvious, fixable, and caught earlier.

The 9 phases, 9 artifacts

Each phase below includes:

Artifact name
Definition of done (a short checklist)
Prompt template (copy and customize)

If you already have a PRD, you can often start at Phase 3 or Phase 4.

Phase 1: Clarify the problem and user

Artifact 1: Problem Brief

Definition of done:

One paragraph describing the problem in plain language
Target user and their primary job to be done
One sentence describing the desired outcome
What is out of scope (at least 3 bullets)

Prompt template:

text

You are a product manager and requirements analyst.
 
Input (raw notes):
[PASTE YOUR BRAIN DUMP]
 
Task:
1) Extract the core problem statement.
2) Identify the primary user persona and their job-to-be-done.
3) List assumptions you had to make.
4) Ask me the 3 most important clarifying questions.
 
Output format:
- Problem statement
- Target user
- Desired outcome
- Out of scope
- Assumptions
- 3 clarifying questions

Phase 2: Define success criteria and constraints

Artifact 2: Success Criteria and Constraints

Definition of done:

3 to 7 measurable acceptance criteria (not implementation details)
Non-functional constraints (privacy, performance, accessibility, reliability)
Platform constraints (web, mobile, extension, API) stated explicitly
Known risks and unknowns listed

Prompt template:

text

You are a senior product manager.
 
Context:
[PASTE ARTIFACT 1: Problem Brief]
 
Task:
Create acceptance criteria and constraints for the first shippable version.
 
Include:
- Acceptance criteria (as testable statements)
- Non-functional requirements
- Explicit constraints (what we must not do)
- Open questions and risks

Phase 3: Choose the technical foundation (tool-agnostic)

Artifact 3: Technical Foundation Notes

This is not about picking the "best" stack. It is about documenting the stack you will use so an AI does not invent mismatched libraries, patterns, or storage approaches.

Definition of done:

Runtime and deployment target documented
Storage approach documented
Authentication/identity approach documented (even if "none")
Major libraries/frameworks documented (or "TBD")
Integration points documented (APIs, webhooks, third parties)

Prompt template:

text

You are a pragmatic software architect.
 
Context:
[PASTE ARTIFACT 1 and 2]
 
Task:
Propose a simple technical foundation suitable for a small team or solo build.
 
Rules:
- Prefer widely used, well-documented options.
- Keep it minimal.
- If you are unsure, mark items as TBD and list what must be decided.
 
Output:
- Architecture summary
- Data/storage approach
- Key libraries/frameworks (or TBD)
- Integrations
- Conventions (naming, folder structure, testing)

Phase 4: Model the domain (entities and relationships)

Artifact 4: Entity Dictionary

This artifact is where many "AI misunderstandings" disappear. If the agent knows what each entity is and what each field means, it stops guessing. Definition of done:

List of core entities
For each entity: purpose, field list, types, constraints
Relationships stated (one-to-many, many-to-many)
Validation rules and derived fields documented

Prompt template:

text

You are a data modeler.
 
Context:
[PASTE ARTIFACT 1-3]
 
Task:
Create an Entity Dictionary.
 
For each entity, include:
- Entity name
- Plain-language definition
- Fields: name, type, constraints, and purpose
- Relationships to other entities
- Validation rules
 
Output in Markdown.

Example output snippet (entity definition):

**Entity: ChecklistItem** - A single actionable task used to track implementation progress.
 
1. id (string, required) - Stable unique identifier for the item.
2. title (string, required) - Short description of the task.
3. status ("todo"|"doing"|"done", required) - Current progress state.
4. references (string[], optional) - Links to specs, files, or tickets.

Stop here and define your MVP

Before you turn your full feature set into implementation tasks, take a pass at defining the smallest version that proves value.

Read: How to Define Your Software MVP from a PRD

Phase 5: Break features into modules and responsibilities

Artifact 5: Module and Responsibility Map

This is where you turn features into a small number of modules that each have a clear purpose. Definition of done:

Modules listed with a one-sentence responsibility
For each module: inputs, outputs, and dependencies
Clear boundaries (what the module does not do)
Primary user flows mapped to modules

Prompt template:

text

You are a staff engineer.
 
Context:
[PASTE ARTIFACT 1-4]
 
Task:
Break the product into modules.
 
For each module, include:
- Responsibility statement
- Inputs/outputs
- Dependencies
- Data it owns (entities/fields)
- Error cases to handle
 
Keep the number of modules small. Output in Markdown.

Phase 6: Write implementation-ready requirements per module

Artifact 6: Module Specs (Implementation Notes)

This is the "AI executable" part. The key is to specify behavior, interfaces, and edge cases without turning the PRD into a full codebase.

Definition of done:

Each module has endpoints or functions defined (signatures or contracts)
Inputs and outputs documented with examples
Error handling expectations documented
File locations or package boundaries proposed

Prompt template:

text

You are a senior engineer writing implementation notes.
 
Context:
[PASTE ARTIFACT 3-5]
 
Task:
For each module, write a concise spec that a developer (or AI agent) can implement.
 
Include:
- Public API (routes, functions, commands)
- Inputs/outputs (types or JSON shapes)
- Key edge cases
- File structure suggestion
 
Do not generate full code. Focus on clarity and completeness.

Phase 7: Add a testing and validation plan

Artifact 7: Test Plan

If you want AI-generated code to be reliably correct, you need a way to check it. A test plan gives you that.

Definition of done:

Test types selected (unit, integration, E2E as applicable)
For each module: what must be tested
Test data/fixtures approach described
Definition of "done" for a module (passes tests, meets acceptance criteria)

Prompt template:

text

You are a QA lead and test engineer.
 
Context:
[PASTE ARTIFACT 2, 5, and 6]
 
Task:
Write a test plan.
 
Include:
- Test strategy by layer (unit/integration/E2E)
- Critical test cases per module
- Negative tests and error handling tests
- What must be mocked vs real
 
Output as Markdown checklists.

Example output snippet (checklist excerpt):

- [ ] Module: Auth
  - [ ] Reject requests with missing session token
  - [ ] Expire sessions after configured TTL
  - [ ] Prevent privilege escalation by role mutation

Phase 8: Create an AI context block you can reuse

Artifact 8: AI Context Block

This is the small chunk of text you paste into every AI prompt so the tool stays consistent. Definition of done:

Under 400 to 600 tokens (short enough to reuse)
Includes architecture, key entities, conventions, and constraints
Includes links or pointers to where the truth lives (artifacts)

Prompt template:

text

You are an expert engineering assistant.
 
Context:
[PASTE ARTIFACT 1-7]
 
Task:
Create a reusable AI Context Block.
 
Requirements:
- Keep it concise.
- Include: project goal, user flow, tech stack choices, key entities, conventions, constraints.
- Include a short list of "do not assume" items.
 
Output:
- AI Context Block (copyable)

Example output snippet (context block excerpt):

text

Project: Habit tracker web app.
Goal: Let users create habits and check in daily.
Constraints: No third-party analytics. Store timestamps in UTC. Do not invent new entities beyond the Entity Dictionary.
Conventions: TypeScript, strict types, functions return Result<T, E> style errors.

Phase 9: Produce a build checklist for humans and AI agents

Artifact 9: Implementation Checklist

This is where planning turns into execution. You want a checklist that you can hand to:

yourself tomorrow
another developer
an AI agent running tasks in sequence

Definition of done:

Checklist grouped by module and priority
Each task references the relevant module spec and tests
Each task has a clear completion condition

Prompt template:

text

You are a technical project manager.
 
Context:
[PASTE ARTIFACT 5-8]
 
Task:
Create an implementation checklist.
 
Rules:
- Use Markdown checkboxes.
- Group by module.
- For each task, include references to the spec sections and required tests.
- Keep tasks small enough to be implemented in one focused AI prompt.

Downloadable template pack (Markdown)

If you want this to be repeatable, keep each artifact as its own Markdown file in your repo so it can evolve alongside the code.

Download links (static files):

Folder path: /static/downloads/ai-ready-prd-starter-pack/
Starter pack README: /static/downloads/ai-ready-prd-starter-pack/README.md

Key templates:

PRD template: /static/downloads/ai-ready-prd-starter-pack/01-prd-template.md
Guided interview prompt: /static/downloads/ai-ready-prd-starter-pack/02-guided-interview-prompt.md
Entity catalog template: /static/downloads/ai-ready-prd-starter-pack/04-entity-catalog-template.md
AI context block template: /static/downloads/ai-ready-prd-starter-pack/07-ai-context-block-template.md
Implementation checklist template: /static/downloads/ai-ready-prd-starter-pack/08-implementation-checklist-template.md

Tip: Keep the artifacts short and link between them. The value is consistency, not length.

FAQ

Do I need a specific tool like Notion or Google Docs?

No. This workflow is intentionally tool-agnostic. Use Markdown in a repo, plain text files, or any editor you prefer.

Why entities and constraints first?

Because those are where ambiguity hides. If the AI has to guess what a field means or what is allowed, it will guess differently than you intended.

How detailed should the module specs be?

Detailed enough that you could hand them to a junior developer and get roughly the right shape of implementation. If you find yourself writing full code, you are usually past the point of diminishing returns.

What if I already have a PRD?

Treat your existing PRD as input. Often you can jump directly to Phase 4 (Entity Dictionary) and Phase 5 (Module map), then backfill constraints if gaps show up.

How do I use this with AI agents without blowing the context window?

Keep Artifact 8 (AI Context Block) small. Then work from the implementation checklist one task at a time, always linking back to the module spec and required tests.

Next step: turn this PRD into an MVP

Once you have Artifact 4 through Artifact 6, it is tempting to build everything. Instead, pick the smallest end-to-end flow that proves the product delivers value.

If you skipped the MVP step above, go back and define it before you start implementation tasks.

Updated: 2025-07-25

AI coding tools are capable, but they are literal. If your requirements are vague, inconsistent, or missing key constraints, the code you get back will reflect that.

This post is a planning workflow I use to turn either:

a messy idea dump, or
an existing PRD that is still too high level

into an implementation-ready specification that both humans and AI agents can follow.

Key takeaways

The workflow is 9 phases and 9 artifacts. Each phase produces one concrete output you can file, review, and reuse.
You will write less code twice if you define entities, constraints, and acceptance criteria before prompting an AI.
After Phase 4, stop and define your MVP so you do not overload context windows and scope.
You can copy the prompt templates below into any AI tool. They are intentionally tool-agnostic.

Fixes when it breaks. Workflows when it doesn't.

OpenClaw guides, configs, and troubleshooting notes. Every two weeks.

What this workflow is (and is not)

What it is: a repeatable way to create requirements that are specific enough for AI coding assistants to implement with minimal back and forth.

What it is not: a guarantee that AI will write perfect code. The goal is to reduce ambiguity so that failures are obvious, fixable, and caught earlier.

The 9 phases, 9 artifacts

Each phase below includes:

Artifact name
Definition of done (a short checklist)
Prompt template (copy and customize)

If you already have a PRD, you can often start at Phase 3 or Phase 4.

Phase 1: Clarify the problem and user

Artifact 1: Problem Brief

Definition of done:

One paragraph describing the problem in plain language
Target user and their primary job to be done
One sentence describing the desired outcome
What is out of scope (at least 3 bullets)

Prompt template:

text

You are a product manager and requirements analyst.
 
Input (raw notes):
[PASTE YOUR BRAIN DUMP]
 
Task:
1) Extract the core problem statement.
2) Identify the primary user persona and their job-to-be-done.
3) List assumptions you had to make.
4) Ask me the 3 most important clarifying questions.
 
Output format:
- Problem statement
- Target user
- Desired outcome
- Out of scope
- Assumptions
- 3 clarifying questions

Phase 2: Define success criteria and constraints

Artifact 2: Success Criteria and Constraints

Definition of done:

3 to 7 measurable acceptance criteria (not implementation details)
Non-functional constraints (privacy, performance, accessibility, reliability)
Platform constraints (web, mobile, extension, API) stated explicitly
Known risks and unknowns listed

Prompt template:

text

You are a senior product manager.
 
Context:
[PASTE ARTIFACT 1: Problem Brief]
 
Task:
Create acceptance criteria and constraints for the first shippable version.
 
Include:
- Acceptance criteria (as testable statements)
- Non-functional requirements
- Explicit constraints (what we must not do)
- Open questions and risks

Phase 3: Choose the technical foundation (tool-agnostic)

Artifact 3: Technical Foundation Notes

This is not about picking the "best" stack. It is about documenting the stack you will use so an AI does not invent mismatched libraries, patterns, or storage approaches.

Definition of done:

Runtime and deployment target documented
Storage approach documented
Authentication/identity approach documented (even if "none")
Major libraries/frameworks documented (or "TBD")
Integration points documented (APIs, webhooks, third parties)

Prompt template:

text

You are a pragmatic software architect.
 
Context:
[PASTE ARTIFACT 1 and 2]
 
Task:
Propose a simple technical foundation suitable for a small team or solo build.
 
Rules:
- Prefer widely used, well-documented options.
- Keep it minimal.
- If you are unsure, mark items as TBD and list what must be decided.
 
Output:
- Architecture summary
- Data/storage approach
- Key libraries/frameworks (or TBD)
- Integrations
- Conventions (naming, folder structure, testing)

Phase 4: Model the domain (entities and relationships)

Artifact 4: Entity Dictionary

This artifact is where many "AI misunderstandings" disappear. If the agent knows what each entity is and what each field means, it stops guessing. Definition of done:

List of core entities
For each entity: purpose, field list, types, constraints
Relationships stated (one-to-many, many-to-many)
Validation rules and derived fields documented

Prompt template:

text

You are a data modeler.
 
Context:
[PASTE ARTIFACT 1-3]
 
Task:
Create an Entity Dictionary.
 
For each entity, include:
- Entity name
- Plain-language definition
- Fields: name, type, constraints, and purpose
- Relationships to other entities
- Validation rules
 
Output in Markdown.

Example output snippet (entity definition):

**Entity: ChecklistItem** - A single actionable task used to track implementation progress.
 
1. id (string, required) - Stable unique identifier for the item.
2. title (string, required) - Short description of the task.
3. status ("todo"|"doing"|"done", required) - Current progress state.
4. references (string[], optional) - Links to specs, files, or tickets.

Stop here and define your MVP

Before you turn your full feature set into implementation tasks, take a pass at defining the smallest version that proves value.

Read: How to Define Your Software MVP from a PRD

Phase 5: Break features into modules and responsibilities

Artifact 5: Module and Responsibility Map

This is where you turn features into a small number of modules that each have a clear purpose. Definition of done:

Modules listed with a one-sentence responsibility
For each module: inputs, outputs, and dependencies
Clear boundaries (what the module does not do)
Primary user flows mapped to modules

Prompt template:

text

You are a staff engineer.
 
Context:
[PASTE ARTIFACT 1-4]
 
Task:
Break the product into modules.
 
For each module, include:
- Responsibility statement
- Inputs/outputs
- Dependencies
- Data it owns (entities/fields)
- Error cases to handle
 
Keep the number of modules small. Output in Markdown.

Phase 6: Write implementation-ready requirements per module

Artifact 6: Module Specs (Implementation Notes)

This is the "AI executable" part. The key is to specify behavior, interfaces, and edge cases without turning the PRD into a full codebase.

Definition of done:

Each module has endpoints or functions defined (signatures or contracts)
Inputs and outputs documented with examples
Error handling expectations documented
File locations or package boundaries proposed

Prompt template:

text

You are a senior engineer writing implementation notes.
 
Context:
[PASTE ARTIFACT 3-5]
 
Task:
For each module, write a concise spec that a developer (or AI agent) can implement.
 
Include:
- Public API (routes, functions, commands)
- Inputs/outputs (types or JSON shapes)
- Key edge cases
- File structure suggestion
 
Do not generate full code. Focus on clarity and completeness.

Phase 7: Add a testing and validation plan

Artifact 7: Test Plan

If you want AI-generated code to be reliably correct, you need a way to check it. A test plan gives you that.

Definition of done:

Test types selected (unit, integration, E2E as applicable)
For each module: what must be tested
Test data/fixtures approach described
Definition of "done" for a module (passes tests, meets acceptance criteria)

Prompt template:

text

You are a QA lead and test engineer.
 
Context:
[PASTE ARTIFACT 2, 5, and 6]
 
Task:
Write a test plan.
 
Include:
- Test strategy by layer (unit/integration/E2E)
- Critical test cases per module
- Negative tests and error handling tests
- What must be mocked vs real
 
Output as Markdown checklists.

Example output snippet (checklist excerpt):

- [ ] Module: Auth
  - [ ] Reject requests with missing session token
  - [ ] Expire sessions after configured TTL
  - [ ] Prevent privilege escalation by role mutation

Phase 8: Create an AI context block you can reuse

Artifact 8: AI Context Block

This is the small chunk of text you paste into every AI prompt so the tool stays consistent. Definition of done:

Under 400 to 600 tokens (short enough to reuse)
Includes architecture, key entities, conventions, and constraints
Includes links or pointers to where the truth lives (artifacts)

Prompt template:

text

You are an expert engineering assistant.
 
Context:
[PASTE ARTIFACT 1-7]
 
Task:
Create a reusable AI Context Block.
 
Requirements:
- Keep it concise.
- Include: project goal, user flow, tech stack choices, key entities, conventions, constraints.
- Include a short list of "do not assume" items.
 
Output:
- AI Context Block (copyable)

Example output snippet (context block excerpt):

text

Project: Habit tracker web app.
Goal: Let users create habits and check in daily.
Constraints: No third-party analytics. Store timestamps in UTC. Do not invent new entities beyond the Entity Dictionary.
Conventions: TypeScript, strict types, functions return Result<T, E> style errors.

Phase 9: Produce a build checklist for humans and AI agents

Artifact 9: Implementation Checklist

This is where planning turns into execution. You want a checklist that you can hand to:

yourself tomorrow
another developer
an AI agent running tasks in sequence

Definition of done:

Checklist grouped by module and priority
Each task references the relevant module spec and tests
Each task has a clear completion condition

Prompt template:

text

You are a technical project manager.
 
Context:
[PASTE ARTIFACT 5-8]
 
Task:
Create an implementation checklist.
 
Rules:
- Use Markdown checkboxes.
- Group by module.
- For each task, include references to the spec sections and required tests.
- Keep tasks small enough to be implemented in one focused AI prompt.

Downloadable template pack (Markdown)

If you want this to be repeatable, keep each artifact as its own Markdown file in your repo so it can evolve alongside the code.

Download links (static files):

Folder path: /static/downloads/ai-ready-prd-starter-pack/
Starter pack README: /static/downloads/ai-ready-prd-starter-pack/README.md

Key templates:

PRD template: /static/downloads/ai-ready-prd-starter-pack/01-prd-template.md
Guided interview prompt: /static/downloads/ai-ready-prd-starter-pack/02-guided-interview-prompt.md
Entity catalog template: /static/downloads/ai-ready-prd-starter-pack/04-entity-catalog-template.md
AI context block template: /static/downloads/ai-ready-prd-starter-pack/07-ai-context-block-template.md
Implementation checklist template: /static/downloads/ai-ready-prd-starter-pack/08-implementation-checklist-template.md

Tip: Keep the artifacts short and link between them. The value is consistency, not length.

More AI Tools & Reviews guides

Related Articles

OpenClaw Cron Jobs: Complete Command Reference with Examples

Unlock OpenClaw skills: 5 proven steps with ClawHub

OpenClaw Signal Setup: 5 Proven Steps, Here's Exactly How

More AI Tools & Reviews guides

Related Articles

OpenClaw Cron Jobs: Complete Command Reference with Examples

Unlock OpenClaw skills: 5 proven steps with ClawHub

OpenClaw Signal Setup: 5 Proven Steps, Here's Exactly How