reflow2

You decide. The AI helps. reflow2 remembers why.

reflow2 is a shared record of a project: what it has to do, what was decided and why, and what was learned along the way. You and the AI assistant you already use work on it together, and it is still there tomorrow, next season, and after the people who wrote it have moved on. Every piece is linked to what it rests on and what it affects, so the record can also say what a change will reach before anyone makes it.

See one project, start to finish Set it up

What it is, and what it isn't

It is

  • A shared, connected record of your project: what it has to do, what you decided and why, what you measured, and what went wrong, each linked to what it rests on and what it affects.
  • A partnership between you, the AI assistant you already use, and the record itself. Each covers for what the others do badly.
  • A memory for working together. Everyone can see who decided what, when, and why, so the next conversation starts from there.
  • For any project made of decisions: a robot, a building, a season plan.

It isn't

  • An AI. reflow2 runs no AI of its own. When words need writing, it asks the assistant you already use.
  • A replacement for people, or for talking to each other. It gives those conversations a memory.
  • The one who decides. The AI can suggest anything; a decision is only marked as made with a person's name on it.
  • Just a place to write things down. A list of decisions can't tell you which ones a change will reach, or which need nothing delivers yet. The links between them can.
  • Needed for everything. A one-person job finished in an afternoon doesn't need a record that outlives it.

On a software project reflow2 also checks the code against the design. On a project with no code you still get the questions, the decisions, their reasons, and the history.

Three partners, each covering for the others

None of the three is good at everything. Together they are.

You

Bring the ideas, the direction and the final say, and you know the real world.

Can't be expected to remember everything, notice slow changes, or enjoy being asked the same thing twice.

Your AI assistant

Brings speed and breadth: it writes, sums up, searches and asks.

Can't be trusted with remembering yesterday's chat, doing sums, staying consistent, or knowing when it's guessing. It also tends to agree with you.

reflow2

Brings memory that outlasts every chat, model and person; counting and checking done the same way every time; and a trail of who said what.

Can't judge whether something is true or a good idea. It keeps claims; it doesn't understand them.

Why it's built this way: the mistakes people and AI make, and what catches each one.

Every project is a run of decisions

Whether it's a robot, a building or a season plan, the work comes down to the same five steps, over and over. reflow2 is built around them.

  1. 1 Notice a choice is needed

    What goes wrong: something important is noticed late, or a change quietly breaks something else.

    reflow2 points out what the design doesn't say yet, and shows what a change would touch before anyone makes it. reflow2 AI

  2. 2 Look into it

    What goes wrong: people forget what was learned last time, and an AI starts every chat from nothing and fills the gaps by guessing.

    Before it suggests anything, the AI searches the project's record: past decisions, what broke, what was measured. AI reflow2

  3. 3 Weigh the options

    What goes wrong: the first idea wins, and we look for facts that agree with us.

    Ideas are written down as ideas, each beside the best argument against it. The ones you don't pick are kept. AI You reflow2

  4. 4 Decide

    What goes wrong: the decision happens in a chat or a hallway, and the reason is lost.

    You decide. It's recorded with your name, the date, the reason, and what you didn't choose. You

  5. 5 Learn from how it turned out

    What goes wrong: the same mistake comes back, or a fix treats the symptom and not the cause.

    Measurements are marked as measured, so they are never mistaken for guesses. When something fails, the cause is found before the fix, and the decisions the result calls into question are pointed out. You AI reflow2

What the record keeps, and what the links let it answer

Small facts, linked to each other and to the work.

A decision here isn't a line in a list. Each fact is a record of a known kind (a need, a decision, a finding, a check, a change), joined by named links to what it rests on, what delivers it, what checks it, what it caused and what it replaced. reflow2 refuses a record or a link that doesn't fit, so the links stay something it can reason over. That is what lets it answer questions a list of notes can't:

  • If we change this, what else moves? It follows the links out from the change, before anyone edits.
  • Which needs does nothing deliver yet, and what has never been checked?
  • What has been built, and is still waiting for someone to say so? It reads what the evidence shows was built, and lists the work whose record still waits for an owner's word.
  • Which decisions rest on a result that just failed?
  • How is it made, step by step, and which join does no step make? A process recorded as an assembly reads back as instructions: each step's parts and joins, who does it, and how it is checked.
  • Does this conclusion follow from what it rests on, what does it assume, and who checked each step? A step says how its premises support it, what it supposes until a later step discharges it, and what it holds on purpose, such as a model's idealizations. reflow2 checks the form, never whether a premise is true. A conclusion that rests on a report, graded by its source, or on a statistical test reads supported, never proven, and names the model it is supported under. One with an objection still standing against it reads contested, and rival positions are kept side by side. An experiment's result on its own only contests a claim; the claim reads refuted when its owner signs a ruling that says so. reflow2 notes, by its own clock, when each record was written, so a prediction reads whether it was stated before the result that tests it. A design choice is read by what its owner has accepted, not by what is built yet. A claim a named party made and nobody has checked reads claimed by them, unverified. A source named but never read cites nothing, and two lines of reasoning that share a premise are not read as independent. Ask what happens if a premise goes, and it names the conclusions that lose their support.
  • What hasn't anyone asked yet? It finds what the design doesn't say, and turns each gap into a question for you. An agent can put its own question to a named person, and it waits on the record until they answer.
  • Who decided this, when, and what did they not choose?

It adds up

No one designs a whole project at once. It grows one small decision at a time.

reflow2's own design was built this way, and grew past what anyone could hold in their head. Nobody planned that whole. Each decision was small, cheap to make and cheap to undo, and kept with its reason and its links. It's the tinkering Nassim Taleb describes in Antifragile: many small tries, each one kept, adding up to more than any plan could have laid out. And reflow2 isn't a large project. A building or a spacecraft would hold far more. reflow2's part is keeping every piece connected, so the whole stays something you can ask questions of, however large it grows.

One project, start to finish

Follow the Gearheads, a made-up school robotics team, as they design next season's climber: from the first thing it has to do, through the questions, the options and the decision, into the code, and back again when the coach changes the rules and the first test comes in short. Each step says who does it.

Read the walkthrough

Next: several designs, one effort

A team with many repositories gives each its own design. A hub, or project brain, links them and holds no copies: a record can point into another design, a change is followed from one into the next, and a requirement is handed to the design that owns it. How to make a hub.

Try it

It runs on your own computer, in VS Code where MCP is blocked, or as a server for a team.

On your own computer VS Code, where MCP is blocked A server for a team