A factory turns a fuzzy request into finished, double-checked work.
Most people ask an AI for everything at once: “Build me the whole thing.” It often misunderstands, skips a step, or invents something. A factory fixes that by breaking the job into small, labeled stations and laying them out as folders on your computer.
The AI opens the right folder, reads the short note inside, does only that one step, writes the result into an “out tray,” and then stops so you can look. You approve or correct it, and it moves to the next station. The folders are the machine — there’s nothing to install and nothing running in the background.
Because every step’s instructions live in their own little file, the AI only ever loads what it needs for the step it’s on. That’s the whole trick: small notes, clear order, and a human checkpoint between every step.
Example: If you ask an AI to "write me a landing page," it might miss your brand voice, skip mobile layout, or invent features you never mentioned. A factory breaks that into six small steps — explore what exists, clarify the offer, plan the layout, build it, review it, hand it off — and makes you check each one before the next begins. The result is closer to what you actually wanted, with a paper trail to prove it.
What’s inside the building.
Here is the whole factory as a floor plan. Every room is just a folder; every sign is just a plain text file. The next section explains each one.
Every folder has a plain-English job.
A note pinned to the front door. It tells the AI who it's helping, how the place works, the naming habits, and the rules it must never break. This is the one note that's always open.
Your AI tool may call this file something else (e.g. CLAUDE.md, AGENTS.md, .cursorrules). The name doesn't matter — the content does.
A tiny directory. It says, in effect, “If you want to do this, go to that workshop and start at station 01.” Its only job is pointing the AI to the right place.
The facts and rules you write down once: your tone of voice, your do’s and don’ts, any fixed values. Set it up at the start and reuse it on every job. Change a rule here and the whole factory follows it.
The underscore prefix means "set it up once, then leave it alone.
Reusable parts, finished examples, and reference material — kept in one place so there’s a single correct copy of each. The AI pulls from here instead of reinventing things.
"shared" = one correct copy. If you find yourself duplicating a file, put it here instead.
A few short “how-to” cards for jobs you repeat — like “how to build one” or “how to inspect one.” Each is a small note the AI can pick up and follow.
One workshop per kind of job. Each holds an in tray (input/), a row of numbered stations (stages/01…06), and an out tray (output/). This is where the work actually happens.
Six stations, with you between each one.
Inside a workshop, the job moves down a short assembly line. Each station does one thing and then hands its result to the next. The diamond between stations is you — the AI stops, shows you what it made, and waits for your “go.”
Explore
Read the request and look around the supply room. Figure out what already exists and what’s needed. No building yet.
Clarify
Ask you the few questions that actually matter — one at a time — and lock down exactly what’s in and out of scope.
Plan / Spec
Turn the answers into a one-page plan: what gets built, which parts get reused, what the rules require. Still no building.
Build
Build to the approved plan, reusing the supply-room parts. The AI checks its own work against the rulebook as it goes.
Review
A fresh pass, as an inspector — not the builder. It checks the result against the plan and flags problems by how serious they are.
Hand off
Produce the finished file plus a short summary of what changed and anything that still needs a human sign-off. Done.
When something goes wrong: If a station finds a defect that started upstream, it doesn't patch it in place. It reports the issue and routes it back to the station that owns it. This routing discipline is where the factory's real value lives — it keeps problems from compounding downstream.
The job card: three short lists.
Every station folder holds one small note (a CONTEXT.md file). It always has the same simple shape — three lists and a “stop.” Once you’ve seen one, you’ve seen them all.
A Inputs
What to read before starting — usually the previous station’s out-tray result, plus the rules that apply.
B Process
The steps to follow for this station, and nothing more. Short and specific, so the AI stays in its lane.
C Outputs
Exactly what to produce and where to put it — a named file dropped into the out tray for the next station.
Habits that make a factory reliable.
You only need a folder and a text editor.
A factory is just folders and plain text files, so you can make one by hand — no coding. A .md file is simply a plain text file (the kind you’d make in Notepad or TextEdit); the “.md” just lets it carry light formatting like headings. Follow these eight steps.
Make the main folder
On your Desktop, create one new folder and name it my-factory. Everything lives inside it.
Add the rooms
Inside my-factory, make four more folders: _config, shared, skills, and workspaces. These are your settings binder, supply room, toolbox, and workshops.
Write the “Start Here” card
In my-factory, create a text file named INSTRUCTIONS.md. In plain words, tell the AI who it’s helping, that it should walk the stations in order and stop after each one, and list any rules it must never break. (Use the starter on the right.)
Write the reception sign
Next to it, create CONTEXT.md. Keep it tiny: a short list that says which workshop to use for which kind of request, and to start at station 01.
Build your first workshop
Inside workspaces, make a folder for your job (e.g. my-first-job). Inside that, make three folders: input, output, and stages. Inside stages, make six folders: 01-explore through 06-handoff. Drop one CONTEXT.md job card into each — three short lists and a “stop” (use the template on the right).
Write down your rules once
In _config, add a file or two for the things that never change — your tone of voice, your do’s and don’ts, any fixed facts. This is what makes the factory yours, and you only do it once.
Connect it to your AI
Open your AI assistant and start a session. Paste the contents of your INSTRUCTIONS.md into your AI assistant's custom instructions or project settings so it’s always loaded. If your plan lets the AI read files on your computer, point it at the my-factory folder; if not, you can upload the folder or paste files in as you go.
Run it
Drop your raw materials into the workshop’s input folder, then tell the AI what you want — for example, “Build the thing described in my-first-job, starting at station 01.” It reads the Start Here card, goes to station 01, does the work, and stops for you to check. Approve, and it walks the rest of the line.
Skipping checkpoints. The checkpoints ARE the value. Running straight through defeats the purpose.
Loading everything at once. The AI only reads what the current station asks for. Don't paste the whole factory into the chat.
The whole shape, ready to copy.
Here’s the full folder layout and a starter note for every file. Copy the layout to picture the finished factory, then use the templates to fill in each piece.
# Your finished factory — every line is a folder or a plain text file my-factory/ ├── INSTRUCTIONS.md # the "Start Here" card (paste into your AI assistant) ├── CONTEXT.md # the reception desk: which workshop for which job ├── _config/ # the settings binder — your rules & facts, set once │ ├── voice-and-rules.md │ └── facts.md ├── shared/ # the supply room — reusable parts & examples │ ├── examples/ │ └── parts/ ├── skills/ # the toolbox — short how-to cards │ └── build-it/ │ └── SKILL.md └── workspaces/ # the workshops — one per kind of job └── my-first-job/ ├── input/ # IN TRAY — drop raw materials here ├── output/ # OUT TRAY — each station's result lands here ├── CONTEXT.md # this workshop's station map └── stages/ ├── 01-explore/ → CONTEXT.md ├── 02-clarify/ → CONTEXT.md ├── 03-plan/ → CONTEXT.md ├── 04-build/ → CONTEXT.md ├── 05-review/ → CONTEXT.md └── 06-handoff/ → CONTEXT.md
Starter: INSTRUCTIONS.md
# My Factory — Start Here You are my work partner. Help me produce finished, double-checked work. ## How this place works - Walk the numbered station folders in order, top to bottom. - Do ONLY the current station's job. - Write each result into that workshop's output/ folder. - STOP after every station and show me the result. Wait for my "go." ## Where to start - Read CONTEXT.md to pick the workshop, then start at station 01. ## Rules you must not break - Recommend, then let me decide. - Never invent facts. If something is missing, ask me. - <add your own rules here>
Starter: a station's CONTEXT.md
# Station 04: Build ## Inputs (what to read first) - output/03-plan.md (the plan) - _config/voice-and-rules.md ## Process (what to do here) 1. Read the plan and the rules. 2. Build exactly what the plan says. 3. Reuse parts from shared/. 4. Check your work against the rules and note anything off. ## Outputs (what to produce) - output/04-draft.md (the result, for the next station) ## Stop Show me the draft. I approve or correct it before station 05.
Starter: _config/voice-and-rules.md
# Voice & Rules ## Tone - Be direct and specific. - Use plain language — no jargon unless the user asked for it. - Be honest about uncertainty. ## Do's - Ask before assuming. - Show your work. - Flag risks early. ## Don'ts - Never invent facts. If something is missing, ask. - Never skip a checkpoint. - Never modify _config/ files without asking. ## Formatting - Use headings to organize long outputs. - Use bullet points for lists. - Keep responses under 300 words unless detail is requested.
Starter: _config/facts.md
# Facts ## Product - Name: [your product] - What it does: [one sentence] - Audience: [who uses it] ## Key URLs / paths - [Add any fixed references here] ## Preferences - Preferred stack / tools: [list] - Deployment target: [e.g. local, cloud, specific platform]
Starter: skills/build-it/SKILL.md
# Skill: Build It ## When to use Use this when you have an approved plan and need to produce the deliverable. ## Steps 1. Read the plan from output/03-plan.md. 2. Read the rules from _config/voice-and-rules.md. 3. Reuse parts from shared/ where possible. 4. Build exactly what the plan specifies. 5. Check your work against the rules. 6. Write the result to output/04-draft.md. ## Check - Does the output match the plan? - Did you reuse shared parts instead of reinventing? - Did you follow all rules from _config/?
Starter: workshop CONTEXT.md
# Workshop: my-first-job ## What this workshop does [One sentence describing the job this workshop handles.] ## Station map | Stage | Name | What it does | |-------|----------|------------------------------------| | 01 | Explore | Read the request, look around. | | 02 | Clarify | Ask key questions, lock scope. | | 03 | Plan | Write the one-page plan. | | 04 | Build | Build to the approved plan. | | 05 | Review | Inspect the draft as fresh eyes. | | 06 | Hand off | Produce the finished file. | ## How to use 1. Drop raw materials into input/. 2. Start at stage 01. 3. After each stage, stop and show the result. 4. Continue on "go."
Starter: 01-explore CONTEXT.md
# Station 01: Explore ## Inputs (what to read first) - input/ (the raw request) - shared/examples/ (existing examples, if any) - shared/parts/ (reusable parts, if any) ## Process (what to do here) 1. Read the request carefully. 2. Look through shared/ for anything that already exists. 3. Summarize what exists, what's missing, and what's unclear. 4. Do NOT build anything yet. ## Outputs (what to produce) - output/01-exploration.md (summary of what you found, for the next station) ## Stop Show me the exploration summary. I confirm you understood the request before moving to station 02.
Starter: 02-clarify CONTEXT.md
# Station 02: Clarify ## Inputs (what to read first) - output/01-exploration.md (the exploration summary) - _config/facts.md (product context) ## Process (what to do here) 1. Read the exploration summary. 2. Identify the 2-5 questions that matter most. 3. Ask me one question at a time. 4. After all answers, lock the scope in writing. ## Outputs (what to produce) - output/02-scope.md (final scope: what's in, what's out, acceptance criteria) ## Stop Show me the scope. I confirm it before you write the plan.
Starter: 03-plan CONTEXT.md
# Station 03: Plan ## Inputs (what to read first) - output/02-scope.md (the locked scope) - _config/voice-and-rules.md (the rules) - shared/parts/ (reusable components) ## Process (what to do here) 1. Read the scope and the rules. 2. List exactly what will be built. 3. Note which shared parts to reuse. 4. Write a one-page plan with clear deliverables. ## Outputs (what to produce) - output/03-plan.md (the plan, ready for approval) ## Stop Show me the plan. I approve or correct it before you build.
Before you trust it with real work
A factory organizes and documents the work beautifully — but it can’t force anything. Every rule is a polite instruction the AI follows, not a lock. That makes your judgment the most important part of the system, especially for anything sensitive or regulated.
- Keep your human sign-off. The factory is a helper and a paper trail, not an approval. Final review still belongs to you and the right people.
- Mind what you put in. Don’t feed it confidential or regulated material until you’ve confirmed your tool’s data-handling and cleared it with whoever needs to clear it.
- Still test the real output. Whatever it produces, verify it the way you normally would — the factory documents the work, it doesn’t guarantee the result.
- Grow it from real misses. Each time the AI gets something wrong, add one line to a rules file. The factory gets smarter every time you use it.