A project harness for AI coding agents
One that remembers what you’re building, has opinions about how to build it, and tells you when you’re heading the wrong way, instead of helpfully getting you there faster.
The problem
Close the window and the conversation is gone. What survives is one file your agent reads at the start of every session, and everything you want remembered has to go in there: months of decisions, conventions, mistakes, and reasons, read start to finish before every single task.
So it gets enormous, half of it goes stale, and the agent gets slower and vaguer the more you put in. Worse: it fills the gaps, confidently describing decisions you never made, then building on top of them.
“The longer you work on the project, the dumber it gets.”
By month two you’re spending the first twenty minutes of every session re-explaining your own product to the thing that built it.
Who it’s for
The product person. The business person. The one who knows an industry inside out and can see the thing that ought to exist in it. The agent brings you real decisions: what this is for, who it’s for, what’s in and what’s out, and the smallest version that proves it. That’s the cofounder’s half of the deal.
Your half is to decide.“I’m not an engineer. I’ve spent twenty-five years in software as a designer, a UX person, a product lead, and a founder: the guy who can see the whole thing in his head and can’t build a line of it. I co-founded Kajabi. It started as three of us with an idea and some mockups, and it became a $2 billion company. I didn’t write any of that code.
“Cofounder is the partner I always wished for: knows the project, has opinions, tells me when I’m wrong, and never needs the story re-explained. Twenty-five years of needing someone else to build it. Now I describe what I want and watch it get made. I’ve stopped pretending that isn’t a little bit magic.”
The memory half
Instead of one giant file, each file has one job. The one read every session is the shortest on purpose, so the agent never gets slower or dumber as the project grows.
The bigger half
“I think this is a mistake, here’s why, your call.”
That sentence is the whole idea. COFOUNDER.md is the structure behind it: the working relationship written down, so your agent pushes back once, like a partner would, then builds what you decide, well.
Tweak, feature, or direction, called out loud before it starts. Money, user data, and login are never tweaks, no matter how small they look.
Builds for the stage you’re actually at (idea, prototype, first users), not the one you’re imagining. No clever architecture until the current stage hurts.
Payments, login, email, hosting are commodities. It picks the most widely used service, writes it down, and never asks you to choose from a list you’ve never heard of.
Security, money, user data, and anything irreversible get raised without being asked. It never touches them without telling you first.
A checkpoint after every verified change, so “undo that” and “what changed today” just work. You never type a git command.
Makes the technical calls and recommends with a default instead of handing you plumbing decisions you spent a career not having to make.
Why files, not an app
If someone technical asks: it’s docs-as-code and decision records, arranged as context engineering for a coding agent. We call it a project harness.
An app would be one tool’s app. A folder of Markdown works in every agent you’d use today, and whatever comes next will read it too.
Open COFOUNDER.md and see, in plain English, how your cofounder thinks. If it just pushed back on you, the reason is in a file you can open.
Disagree with a rule? Edit the line. You’re the authority and the files are input. That’s the whole idea.
Get started
Ten minutes in, you have a project that remembers itself and a cofounder with opinions about what you build next.
Get the template on GitHub Free. A folder of Markdown files. Yours the moment you copy it. Then go bring your idea to life.