PREP.md

A 3-minute read · launched July 2026

PREP.md: switch AI models and continue your project without explaining it again

I'm a kitchen manager in London, not a programmer, but I use AI every day and I kept hitting the same wall.

I'd spend a lot of time building something in ChatGPT. A plan, a project, a piece of writing. Sometimes that's weeks of work. With code it can happen in a single day, because all the copying and pasting makes one conversation heavy enough to slow down and start forgetting the decisions I made at the start. Or I simply want to use another company's AI to carry on with the project.

Either way the knowledge I built stays trapped inside one chat, in one product, and it disappears when the tab closes. And when I tried to rescue it by copying text into notes as I went, I just made a mess. No structure, no method, nothing another AI could pick up cleanly.

The fix was giving the project its own memory. Kept outside the chat, in a structure any AI can read quickly: what the project is, where it stands, and only the part it needs to read next. That last part matters. The AI doesn't re-read everything, so it picks the project up in seconds, and it doesn't burn tokens doing it.

In a kitchen, prep means everything in its place before service begins, so the work can flow. That's where the name comes from, and it's what this does for a project.

The whole idea: a folder, not a chat

PREP is a convention, not a product. A project is an ordinary folder in your own cloud drive, arranged in a small, predictable way:

Your Project/
├── PREP.md      ← what this is, where it stands, what to read
├── LOG.md       ← append-only history, one dated line per session
└── memory/      ← a dated snapshot of each session

That's the whole core. Three files with agreed names. PREP.md is the entry point, and it carries a small MAP that tells the AI what each file is and when to read it. So the AI boots light: it reads the entry point and only what the task needs, instead of scanning the whole history. Clear, structured, and cheap on tokens, even years in.

LOG.md is the scannable timeline. memory/ holds the full snapshots. Any AI that can read the folder opens the project, sees exactly where things stand, and continues. Today in ChatGPT, tomorrow in Claude, without losing a thing. The conversation becomes a temporary interface over a permanent project.

It's plain markdown. No app required, no account, no lock-in. If a folder follows the convention, any tool can support it.

Why not just paste the old chat?

A long conversation is mostly noise. Jokes, dead ends, corrections, things you tried and dropped. Pasting it asks the new model to sift through all of that and work out for itself what still counts, and it will get some of it wrong, while paying tokens to read the parts that stopped mattering weeks ago. A snapshot is what matters, already separated: the decisions, the current state, what comes next. Smaller than the work, on purpose.

Two things you do: save and open

Save. When a session is worth keeping, you save it. Because chat assistants still can't write files reliably, I built a small tool for it, called PREP Save. You paste the conversation's summary and it builds the whole project for you in your own Google Drive: the folder, PREP.md, LOG.md and memory/, already filled in correctly and following the standard. Next time, it adds the new snapshot and updates the status. It runs entirely in your browser, asks Google for a single scope (drive.file, per-file access to what it creates, so it can't see the rest of your Drive), and keeps no copy of your content.

Open. To continue the project in a fresh tab of any AI model, or to get an agent up to speed on it, you say this in plain words:

In my Google Drive, open the «project folder» inside PREP and read PREP.md.

The AI reads the standard from the folder itself and continues from where you stopped. That one sentence is the only syntax you ever need.

The design choices (the opinionated part)

  • The memory belongs to the project, not the AI. Switch models or devices freely. Your work outlives any single vendor.
  • Boot light. The MAP tells the AI what to read and when, so it never scans the whole folder. Fast and cheap on tokens, even years in.
  • No secrets, ever. PREP files get read into a chat, so they never store passwords or keys.
  • Smaller than the work. A snapshot is conclusions, not the full transcript. If it grows into another archive, it stops working.

Honest status

This is day one in public, built by one person around a day job, so expect rough edges and tell me when you hit one.

The folder format doesn't need the app at all. You can create the three files by hand on any drive (Dropbox, OneDrive, Nextcloud, even a local folder) and any AI with access can read them. If you'd rather start that way, there's a starter folder as a zip with the three files ready to fill in. PREP Save just automates the Google Drive case for now.

Help shape it

This is v0.3, which means the shape isn't settled yet. PREP is an open standard under CC BY, so what it becomes depends on the people who use it, not on me alone. If you try it and something is wrong, that is worth more to me than a compliment.

Three things I'd genuinely like to hear:

  • Is the three-file core the right minimum, or is something missing?
  • Where does it break for the way you work?
  • What did you try to do that the standard had no answer for?

If you have an opinion, it can still change the standard. Bring it to GitHub Discussions, or write to hello@prep.md.

The last two changes to the spec came from things breaking in real use, not from planning ahead: one rule exists because a platform could create files but not edit them, and the whole save flow was rewritten once it was clear that chat assistants can't be trusted to write files. That's the kind of report I'm asking for.

Try it