reflow2

Why it works this way.

People and AI both make mistakes, and they make them in predictable ways. reflow2 is built so that each partner covers for what the others do badly. This page is for readers who want the details; you don't need it to use reflow2. New here? Start with what it is.

Who brings what

PartnerBringsCan't be trusted with
You the ideas, judgement, and how the real world actually is remembering everything; noticing slow changes; being asked the same thing twice
Your AI assistant writing, summing up and searching, fast; the hands that use the tools sums; staying consistent; remembering yesterday's chat; knowing how sure it is
reflow2 memory that outlasts every chat, model and person; counting done the same way every time; a trail of who said what meaning: it keeps claims, it can't judge them

In one line: the AI speaks, the record remembers and counts, and you decide.

Where an AI goes wrong, and what catches it

It makes things up.
It is asked to search the record before it suggests anything, and the record holds what was actually decided, by whom and when. reflow2 also refuses anything that doesn't fit its vocabulary, so an invented kind of record can't be written.
It sounds just as sure when it's guessing.
Every finding says whether it was measured or is only a forecast. reflow2's root-cause steps ask for a test that could prove a cause wrong before any fix is built on it.
It agrees with you.
reflow2's questions don't go away because the conversation moved on. A gap is raised again every time until the design answers it, or someone records why it doesn't matter.
It forgets between chats.
The record is the memory. A new chat, or a different AI, starts from what the project already knows.
It can't be trusted with sums.
reflow2 does its own counting, with no AI involved: what's missing, what a change touches, what's covered, and what was taken away beside what was added.

Where people go wrong, and what helps

We forget, and people move on.
The record outlasts them: who decided what, when and why.
We look for facts that agree with us.
Psychologists call it confirmation bias. Each idea is written down beside the strongest argument against it, and the options nobody chose are kept.
The first idea sticks.
That's anchoring. Options are laid side by side before anything is decided, and a decision can be reopened when the evidence changes, with the old reasons still there to read.
We don't notice slow drift.
Every time its check runs, reflow2 compares the design with the work and says what moved.
Feelings push our choices.
Fear, excitement, not wanting to lose what we've already put in. reflow2 can't take feelings out of a decision, and shouldn't. It makes sure the reason is written down at the time, so it can be looked at again in a calmer moment.

Fast and slow thinking: a comparison

Psychologists describe two ways of thinking: fast, automatic and intuitive, and slow, deliberate and careful. Daniel Kahneman made the names System 1 and System 2 widely known.

An AI's answers are a lot like the fast kind: fluent, quick, and confident whether or not they're right. reflow2 makes room for the slow kind at the moments that matter: before a decision is settled, before a cause is accepted, before a change is made.

That's a comparison to help explain the idea, not a claim about how brains or AI actually work.

What reflow2 can't do

Why decisions, and not only code

A robot, a building, a season plan: every project comes down to a run of decisions. Notice a choice is needed, look into it, weigh the options, decide, and learn from how it turned out. That is why reflow2 isn't only for software. On a software project it can also check the code against the design; on a project without code, you get everything else.

See one project go round that loop.