Practical guide

How to Build a Simple 4-Stage AI Workflow in a Browser Project

A short guide for business owners, consultants, operators, and non-technical users

This guide shows you how to create a reusable AI workflow inside a browser-based AI Project.

You can use this approach in:

• ChatGPT Projects
• Claude Projects

The goal is simple:

1. Create one business context Project. 2. Upload useful source files about the business. 3. Use the prompt in this guide to generate a workflow file pack. 4. Download the generated .md files. 5. Create a new Project for that specific workflow. 6. Upload the files. 7. Run the workflow one stage at a time.

This is not about building software.

This is about creating a repeatable business process that a non-technical owner can actually use.

1. Why Use a Browser Project Workflow First?

A 4-stage Project workflow is often better than starting with Claude Cowork, Codex, Claude Code, or another desktop-based agentic workflow.

Not because those tools are bad.

They are powerful when you need to build software, edit local files, work across folders, run code, connect systems, or automate a process that is already well understood.

But most business workflows should be designed manually first.

This is especially true when the workflow involves internet research.

A desktop or coded agentic workflow can quickly become difficult because internet research often requires:

• search tools,
• scraping or crawling tools,
• API keys,
• connectors,
• permissions,
• rate limits,
• error handling,
• and structured extraction.

That is why tools like Firecrawl can become necessary once you try to turn research into an automated system.

A browser Project workflow avoids that setup pain at the beginning.

You can use the AI assistant's built-in research and browsing features where available, keep the work tied to your uploaded business context, and make human decisions after each stage.

Use this first to prove the workflow.

Automate later.

Token, Usage, and Search Benefits

A desktop agent or coding agent often needs to inspect files, run commands, handle errors, and loop through tool calls. That can consume usage quickly, especially if the workflow is still vague.

A browser Project workflow is more controlled because you can:

• upload only the files that matter,
• run one stage at a time,
• stop after each stage for review,
• avoid sending the whole business problem into one giant prompt,
• prevent the AI from doing unnecessary work,
• and keep research, context, and decisions in one conversation.

This does not mean browser Projects are unlimited. Usage still depends on your plan, model, tools, and platform.

The point is simpler:

2. Why Context Matters

AI output is only as useful as the context it has.

If the AI does not understand your business, it will guess.

Before creating a workflow, build a business context Project for the company.

Upload source files such as:

• business plan,
• marketing plan,
• sales materials,
• ideal customer profile,
• customer personas,
• offer descriptions,
• pricing,
• brand voice,
• website copy,
• product or service documents,
• FAQs,
• past proposals,
• case studies,
• internal processes,
• examples of good work,
• examples of bad work,
• and important constraints.

The better the context, the better the workflow.

Do not upload random documents just to add volume.

Upload files that explain how the business works, who it serves, and what good output looks like.

3. What a 4-Stage Workflow Is

A 4-stage workflow breaks one large business task into four smaller steps.

The structure is:

1. Discover — Find possible candidates, options, ideas, or opportunities. 2. Verify — Check which ones are real, relevant, and supported by evidence. 3. Score — Rank the best options using a clear scoring model. 4. Generate — Create the final business output only for approved items.

This is better than one giant prompt because it creates review points.

The AI should not move automatically from research to final action.

The owner should approve what moves forward.

Other Use Cases

The same structure can be used for:

• hiring candidate screening,
• customer onboarding,
• grant research,
• vendor selection,
• proposal creation,
• content production,
• market research,
• operations improvement,
• compliance review,
• sales prospecting,
• project planning,
• and product research.

4. The Simple Setup Process

Step 1 — Create a Business Context Project

Create a Project for the business in ChatGPT or Claude.

Example:

ABC Company — Business Context

Upload source files about the company.

This is the Project where the AI learns the business.

You may already have a project for your business or a project with most of the context you need already set up, if so you can use

that one. Review it and add more context to it if you think it would be helpful.

Step 2 — Paste the Workflow Builder Prompt

Inside the business context Project, paste the prompt in Section 5 of this guide.

Fill in the placeholders.

Ask the AI to generate a full workflow file pack.

Step 3 — Download the Generated Markdown Files

The prompt will ask the AI to create files such as:

README.md
00_PROJECT_CONTEXT.md
01_MASTER_PROMPT.md
02_STAGE_1_DISCOVER.md
03_STAGE_2_VERIFY.md
04_STAGE_3_SCORE.md
05_STAGE_4_GENERATE.md
06_OUTPUT_TEMPLATES.md
07_SCORING_MODEL.md
08_RECORDS.md
09_EXAMPLES.md

Download the files.

If your platform cannot create downloadable files, ask it to output each file in a separate Markdown code block, then copy each one into a text file manually.

Step 4 — Create a New Workflow Project

Create a second Project for the workflow itself. only use this project for this workflow, do not add other workflows to this project in the future. instead replicate this

process for the other workflow. If you have a long workflow that the 4 stage prompt is too small for then the best thing to do is break it up into multiple project work flows.

there is a limit to what a oneshot flow like this can handel in a web based project like this so if you have run the project and its not giving you the results you want and you have

provided enough context and clear direction and made sure all of the correct project files are uploaded to the source files of the project and its still not working, this is an indication that you should break the workflow down into smaller multiple 4-stage project flows.

Example:

ABC Company — Lead Research Workflow

Upload all the generated .md files into that Project.

This is now the reusable workflow Project.

Step 5 — Run the Workflow

Open README.md and follow the instructions.

Then run the first workflow using the master prompt from:

01_MASTER_PROMPT.md

Start with Stage 1 only.

Do not run all four stages at once.

5. Copy/Paste Prompt: Build My 4-Stage Workflow File Pack

Copy the prompt below into your business context Project.

Fill in the placeholders before running it.

You are helping me create a reusable 4-stage AI workflow system for my business.

Your job is to generate a complete set of Markdown files that I can download, upload into a new browser-based AI Project, and use as a reusable workflow.

This workflow should work in either ChatGPT Projects or Claude Projects.

Do not run the workflow yet.
Only create the workflow file pack.

Business name:
[BUSINESS NAME]

Business description:
[PLAIN-ENGLISH DESCRIPTION OF WHAT THE BUSINESS DOES]

Main offer:
[WHAT THE BUSINESS SELLS OR DELIVERS]

Ideal customer:
[WHO THE BUSINESS SERVES]

Workflow I want to build:
[DESCRIBE THE REPEATED TASK THIS WORKFLOW SHOULD HELP WITH]

Examples:
- lead research
- vendor selection
- hiring candidate screening
- grant research
- client onboarding
- proposal creation
- content production
- operations improvement
- market research
- compliance review
- product research

Workflow goal:
[WHAT A SUCCESSFUL RUN SHOULD PRODUCE]

User of the workflow:
[WHO WILL RUN THIS WORKFLOW — OWNER, CONSULTANT, STAFF MEMBER, ETC.]

How often the workflow will be run:
[WEEKLY / MONTHLY / PER CLIENT / AS NEEDED]

Tone:
[DESCRIBE THE TONE — PRACTICAL, EXECUTIVE, FRIENDLY, DIRECT, PREMIUM, ETC.]

Important constraints:
[LIST ANYTHING THE WORKFLOW MUST AVOID]

Examples:
- Do not invent facts.
- Do not use private data.
- Do not create final outreach before human approval.
- Do not use hype.
- Do not make legal, medical, or financial claims.
- Do not recommend tools without explaining why they fit.

Available source files:
Use the files already uploaded in this business context Project as background.
Infer what you can from them, but do not assume they are perfect.
If context is missing, create clear placeholders.

Required workflow architecture:

Stage 1 — Discover
Purpose:
Create a broad starting list of candidates, options, opportunities, ideas, or items.

Stage 2 — Verify
Purpose:
Check which Stage 1 items are real, relevant, evidence-backed, and worth continuing.

Stage 3 — Score
Purpose:
Rank verified items using a clear scoring model.

Stage 4 — Generate
Purpose:
Create the final business output only for items approved by the human user.

Human review gates:
The workflow must stop after each stage and ask the human user what should continue.
The AI must not automatically move from Stage 1 to Stage 2, Stage 2 to Stage 3, or Stage 3 to Stage 4 without approval.

Records rule:
The workflow must update a records file after every run.
No run is complete until the records update has been produced.

Create the following Markdown files:

1. README.md
Purpose:
A short plain-English instruction file explaining how to use the workflow Project.

2. 00_PROJECT_CONTEXT.md
Purpose:
Summarize the business, ideal customer, offer, workflow goal, tone, constraints, and success criteria.

3. 01_MASTER_PROMPT.md
Purpose:
The main reusable prompt the user will paste when starting a workflow run.

4. 02_STAGE_1_DISCOVER.md
Purpose:
Reusable Stage 1 prompt.

5. 03_STAGE_2_VERIFY.md
Purpose:
Reusable Stage 2 prompt.

6. 04_STAGE_3_SCORE.md
Purpose:
Reusable Stage 3 prompt.

7. 05_STAGE_4_GENERATE.md
Purpose:
Reusable Stage 4 prompt.

8. 06_OUTPUT_TEMPLATES.md
Purpose:
Define the output formats used by the workflow.

9. 07_SCORING_MODEL.md
Purpose:
Define the scoring categories, point values, decision bands, and exclusion rules.

10. 08_RECORDS.md
Purpose:
A living records file for workflow runs, decisions, lessons, rejected items, and prompt improvements.

11. 09_EXAMPLES.md
Purpose:
Include simple example outputs and examples of what to avoid.

File requirements:
- Each file must be complete.
- Each file must be written in plain English.
- Each file must be short enough for a business owner to understand.
- Avoid technical jargon.
- Use clear headings.
- Include copy/paste prompts where useful.
- Include human review gates.
- Include records update instructions.
- Include warnings against inventing facts.
- Include a simple first-run instruction.
- Avoid platform-specific language unless needed.
- Refer to the workspace as a Project.

Output requirement:
If you can create downloadable files, create all files as separate .md files and provide a download link.
If you cannot create downloadable files, output each file one at a time in a separate fenced Markdown block with the filename clearly shown above it.

Before creating the files, briefly summarize the workflow architecture you are about to create.
Then create the files.

6. What the Generated Workflow Project Should Contain

After you run the prompt, your new workflow Project should contain these files:

README.md

This tells the user how to run the workflow.

It should explain:

• what the workflow does,
• when to use it,
• how to start Stage 1,
• how to approve items,
• how to move through Stages 2, 3, and 4,
• and how to update 08_RECORDS.md.

00_PROJECT_CONTEXT.md

This gives the workflow its business background.

01_MASTER_PROMPT.md

This is the main prompt used to run the workflow.

02_STAGE_1_DISCOVER.md

This creates the starting list.

03_STAGE_2_VERIFY.md

This checks the list.

04_STAGE_3_SCORE.md

This ranks the verified items.

05_STAGE_4_GENERATE.md

This creates the final output.

06_OUTPUT_TEMPLATES.md

This keeps outputs consistent.

07_SCORING_MODEL.md

This makes prioritization clearer.

08_RECORDS.md

This is the living memory of the workflow.

09_EXAMPLES.md

This shows what good output looks like.

7. How to Run the Workflow After Setup

Inside the new workflow Project, start a new chat and paste this:

Use the uploaded workflow files as the source of truth.

Run the workflow described in 01_MASTER_PROMPT.md.

Start with Stage 1 only.

Workflow objective for this run:
[DESCRIBE THIS RUN]

Scope:
[GEOGRAPHY / INDUSTRY / CLIENT / TOPIC / DATE RANGE / OTHER LIMITS]

Output target:
[HOW MANY ITEMS OR WHAT TYPE OF OUTPUT]

Important constraints:
[WHAT TO AVOID]

Stop after Stage 1 and ask me which items should continue.

After Stage 1, review the output.

Mark each item as:

Continue
Reject
Park for later
Needs more information

Then continue stage by stage.

Stage 2 Prompt

Continue to Stage 2 with these approved items only:
[PASTE ITEMS]

Verify them using the workflow files.
Separate evidence from assumptions.
Do not score yet.
Stop after Stage 2 and ask me what should continue.

Stage 3 Prompt

Continue to Stage 3 with these verified items only:
[PASTE ITEMS]

Use 07_SCORING_MODEL.md.
Return a ranked list.
Do not create final outputs yet.
Stop after Stage 3 and ask me what should move to Stage 4.

Stage 4 Prompt

Continue to Stage 4 with these approved scored items only:
[PASTE ITEMS]

Use 06_OUTPUT_TEMPLATES.md and 09_EXAMPLES.md.
Create the final output.
Include an owner review checklist.
Then produce the update block for 08_RECORDS.md.

8. Update the Records File Every Time

At the end of each run, the AI should produce a block for:

08_RECORDS.md

Copy that block into the records file.

Then upload the updated records file back into the workflow Project.

The records file should track:

• date,
• run objective,
• inputs used,
• outputs created,
• accepted items,
• rejected items,
• parked items,
• decisions made,
• lessons learned,
• and prompt improvements.

This is what makes the workflow better over time.

Without the records file, the Project becomes just another chat.

9. When to Turn This Into a Programmed Workflow

Do not automate too early.

First, run the Project workflow manually several times.

Only move to Claude Cowork, Codex, Claude Code, Make, Zapier, n8n, or custom software when:

• the stages are stable,
• the scoring model makes sense,
• the output templates are clear,
• the human review gates are known,
• the records file shows repeated patterns,
• and the business owner understands what the workflow should do.

A good rule:

Use a desktop or coded agentic workflow when you need the AI to work across local files, folders, apps, repositories, automations, or production systems.

Use a browser Project when you are still proving the workflow logic.

10. Final Rule

The goal is not to create a perfect AI system on day one.

The goal is to create a simple workflow that can be run, reviewed, corrected, and improved.

Keep the structure simple:

1. Discover
2. Verify
3. Score
4. Generate

Keep the owner in control.

Keep the records file updated.

Improve the workflow after every run.

Source Notes

• ChatGPT Projects can store project files and project-specific instructions, and uploaded files can guide responses: https://help.openai.com/en/articles/10169521-projects-in-chatgpt
• ChatGPT Search can provide timely answers with links to relevant web sources: https://help.openai.com/en/articles/9237897-chatgpt-search
• Claude Projects are self-contained workspaces with chat histories and knowledge bases: https://support.anthropic.com/en/articles/9517075-what-are-projects
• Claude project knowledge can use uploaded content across chats in a project: https://support.anthropic.com/en/articles/9519177-how-can-i-create-and-manage-projects
• Claude web search can provide real-time web information when enabled: https://support.anthropic.com/en/articles/10684626-enabling-and-using-web-search
• Claude Cowork is designed for multi-step desktop knowledge work across local files, folders, and applications: https://www.anthropic.com/product/claude-cowork
• OpenAI describes Codex as a coding agent for reading, editing, and running code: https://developers.openai.com/codex