What this is
Everyone is building with AI now. The problem nobody talks about: a month later, nobody can explain the wiring. The picture lives in your head, and the moment a project gets real, it slips.
Archify fixes that. It is a free, open-source agent skill (github.com/tt-a1i/archify, over 90,000 stars) that reads your project, or listens to you describe it, and draws a clean, interactive map. Click any part and see what it connects to. It runs inside Claude Code, Codex, Cursor and OpenCode, and the rendering happens on your machine.
Two things make it different from every diagram tool I have tried. You edit it in plain English, in the same chat: "add the database", "highlight the login flow". And it checks its own work: every map is validated before it reaches you, so no crossing arrows, no overlapping labels.
01Install, two minutes
git clone https://github.com/tt-a1i/archify ~/.agents/skills/archify ln -s ~/.agents/skills/archify ~/.claude/skills/archify
That is the whole install. If you want zero network calls, set ARCHIFY_UPDATE_CHECK_DISABLED=1 in your shell.
02The one prompt
Open Claude Code (or Codex) in the folder of the thing you want to see and say:
Use Archify to draw how this system works. Show 8 to 12 core components and one primary path. Put supporting detail in the cards, not on the canvas.
Then refine in the same chat: "add the database", "highlight the login flow", "move the approvals to the left". It keeps the source and redraws.
03Pick the right map
- Architecture: which systems talk to which. Fits one screen.
- Workflow: who does what, in order. Swimlanes. Tells the story but sprawls.
- You usually want both. Start with architecture.
04Six things I learned mapping my own business
These are not layout tips. They are the decisions that made the maps useful from the first one, learned on my own podcast pipeline and my own Telegram bot.
- Draw it from the folder, not from memory. Open the agent inside the project folder and let it read the files. A map drawn from your description shows what you think you built. A map drawn from the folder shows what you actually built. Mine differed by four boxes.
- One map per system, and the map lives in that system's folder. Every project folder gets a
maps/folder with the spec file and the HTML next to each other. When the system changes, you say "add the new step" and it redraws from the spec. Never edit the HTML by hand. - Put the people in lanes, including yourself. On the workflow map, every person and every bot gets a lane: you, your assistant, your editor, the bot that posts. The lane with the most arrows into it is your bottleneck. That is how I found out I sit in three of the twelve steps of my own podcast pipeline.
- Every AI bot gets a box with three facts: what it reads, what it writes, what it costs. My bot's map carries that month's spend on the card. The moment the bots are on the map, you can see which ones earn their keep and which ones are a hobby.
- Eight to twelve boxes, one primary path. More than that and nobody reads it, including you. Detail goes into the cards you click, not onto the canvas. If the map does not fit one screen, you are drawing two systems. Split it.
- The validator checks that the map is readable, not that it is true. Mine passed clean with four boxes inside a boundary they do not belong to, and two arrows for flows that do not exist. Fact-check every box against the folder, every time.
What it showed me
My podcast pipeline has twelve steps. I appear in three of them. The map is how I found out.
That is the real reason I draw these now. Not for documentation. To see where I am still the bottleneck in my own business, and to hand the rest to the system.
If you map something, send it to me on Instagram. I want to see it, and I will tell you what I would change.
Get the full guide
Drop your email and the rest unlocks right here. No newsletter, just the things I build and learn, when there is something worth sending.
One email, no spam. Unsubscribe any time.