Where to keep what your AI writes

#The situation

Two months ago you had a real question. Which of these two things should we use, which supplier should we sign, is this approach going to hold up. You gave an AI assistant the context, it did the work, and the answer was good enough to act on. You acted on it.

Today someone asks why.

You go looking. The chat is in there somewhere — you remember the shape of the conversation, roughly the week it happened. Fifteen minutes of scrolling later you give up and do the obvious thing: you ask again.

And here is the part that should bother you more than it does. The new answer is different. Not wildly — differently weighted, a caveat that wasn't there before, one option described more warmly than you remember. And you have no way to tell whether that is because the world changed, because you phrased the question differently, or because the model did. You cannot compare, because you do not have the first answer.

You did not lose a file. You lost the ability to know whether you have changed your mind.

That is the actual cost of leaving AI output in the chat window, and it is not a tidiness problem. Tidiness is about how a folder looks. This is about whether you can ever stand behind a decision you made with an assistant's help — and it gets worse, not better, the more of your thinking you do this way.

#Ask the AI

The fix is not "save your chats." Chat logs are transcripts, and nobody re-reads a transcript. The fix is to ask for an artifact — something written to be read a second time, by someone who was not in the room.

Paste this at the top of a project, or into your assistant's standing instructions:

When I ask you to research or decide something, don't answer in chat.
Write a decision record as a Markdown file in this folder, shaped like
this:

---
title: <the decision, stated as a decision>
date: <today>
status: decided | open | superseded
---

## The question
One paragraph. What is actually being decided, and what happens if
we get it wrong.

## What I looked at
A table: option, what it costs, what it gives up, who it suits.

## The call
What you'd do, and the single reason that decided it rather than a
list of five.

## What would change this
The specific, checkable conditions under which this answer stops
being right. Be concrete: a price, a version, a date, a threshold.

## What I'm unsure about
> [!WARNING]
> The parts you are least confident in, and why.

Rules:
- Read the folder first. Use what is already in it.
- Date everything. If you used a fact that has a shelf life, say when
  it was true.
- Name the file so I recognise it in six months.
- Never invent a number to finish a sentence. If you don't have it,
  put it under "What I'm unsure about".

The section doing the real work is What would change this. A normal research answer is a snapshot with no expiry printed on it, which is why re-running it feels necessary — you have no way to know if it went stale. A record that names its own expiry conditions can be checked in ten seconds instead of regenerated in ten minutes.

And What I'm unsure about is the section that makes the file honest. An assistant that has to fill it in stops smoothing over the gaps, and you get to see where the confidence actually is.

#What comes back

A file — payments-provider-decision-2026-09.md, say:

---
title: Which payment provider for the EU launch
date: 2026-09-24
status: decided
---

## The question

We need to take card payments in the EU by January. The choice is
mostly irreversible for 12 months because of the integration work,
and getting it wrong means either eating 0.4% on every transaction
or rebuilding the checkout in Q2.

## What I looked at

| Option    | Cost           | Gives up            | Suits            |
| --------- | -------------- | ------------------- | ---------------- |
| Provider A| 1.4% + €0.25   | No SEPA direct debit| Card-first, fast |
| Provider B| 1.2% + €0.30   | Weaker dispute tools| High volume      |
| Stay put  | 1.9%           | Nothing to build    | Under €40k/mo    |

## The call

Provider A. Not on price — the two are within noise at our volume —
but because their dispute tooling is the only one that doesn't need
a person watching a dashboard daily, and we don't have that person.

## What would change this

- Monthly volume above €40k, where B's 0.2% starts paying for the
  extra dispute work (~€80/mo at that point).
- If SEPA direct debit becomes a requirement, A is out entirely.
- A's EU pricing is reviewed each January — recheck before renewal.

## What I'm unsure about

> [!WARNING]
> The dispute-tooling comparison is from A and B's own docs, not
> from using either. I have not seen a real chargeback go through
> B's flow, and that is the entire basis for the recommendation.

Read the last block again. It is the most useful part of the file and the part a chat answer would never have volunteered.

#Save it as a file

In the folder for that decision, next to whatever fed it — the quotes, the call notes, the spreadsheet you exported.

That is the whole move, and its value is not filing. It is that the folder is now context. The next time you ask about payments, the assistant reads the folder and sees what you already decided and why. It stops proposing things you have already rejected. It can tell you that one of the conditions you wrote under What would change this has been met. You are no longer starting from zero each time; you are adding to something.

This is also where most people's version of this habit quietly fails. A folder of forty .md files is only an archive if you can get back into it. In Finder they are a list of names, and a decision is never named after the sentence that mattered — you will not find "the only one that doesn't need a person watching a dashboard" by scanning filenames.

So: connect the folder in MD Flow. Every file in it renders — the table stays a table, the frontmatter becomes a header instead of a wall of dashes, the > [!WARNING] block reads as a warning. Click down the list and actually skim them. On Pro, search reaches inside the documents and returns the matching line with its number, across every connected folder, which is the only kind of search that finds a decision. It never modifies what it reads, which matters when the thing writing to that folder is an assistant with file access.

Free covers connecting folders, the rendering and filename search; searching document contents is part of Pro, a one-time unlock rather than a subscription.

#Why that matters next week

You can diff your own thinking. Ask the same question in December and you will get a second file. Two dated records side by side tell you something one answer never can: whether the reasoning moved, and which parts. That is not available to anyone whose research lives in a chat log.

The expiry conditions get checked instead of ignored. "A's EU pricing is reviewed each January" is a line you can act on because it is written down in the place you will be looking. It was almost certainly said in the original conversation too, and it evaporated with it.

Someone else can pick it up. A decision record is readable by a person who was not there — a partner, a contractor, you in March. The chat that produced it is readable by nobody, including you.

And the folder keeps compounding. Every artifact you keep is context for the next one. The assistant that answers your December question has read your September reasoning, which means the answer gets better in a way that has nothing to do with the model getting better.

The thing people underestimate about giving an AI a folder is that it is not an organisational habit. Organisation is what you do to things you already have. This is about having them at all — and the generation was never the expensive part. Regenerating is cheap. Trusting the second answer is what costs you.