Deciding where the truth lives

Making a company AI-ready starts with something unglamorous: agreeing where files go. Not a platform choice, not a model choice, but a folder convention.

Every AI agent you point at your work does the same thing first: it reads what is lying around. It cannot tell your current plan from the draft you abandoned in March. If both are on disk, both are the truth.

The four rules

They are boring on purpose. A rule that needs a meeting to apply does not get applied.

One source of truth per fact. Priorities live in one file. Project state lives in one file. Dates and gates live in one file. Everything else links to that file instead of repeating what it says. If you find a second copy, it is stale by definition, and the correct move is to delete it and say so.

Date-prefix documents that represent a moment. 2026-09-02_decision.md beats decision_final_v2.md, because the first one tells a reader, human or model, exactly when it was true. The second one tells you somebody once thought they were done.

Superseded reasoning moves to an archive folder, with a pointer left behind. A live document states what is true now. The history is not deleted, it is moved, and the live file says where. Anyone who needs the old reasoning can follow the pointer. Anyone who does not is no longer reading it by accident.

One instruction file per folder. Different tools look for different filenames, so the temptation is to keep one of each. Where two exist, one is real and the other is a pointer to it. Never two copies.

The mechanism

All four rules come from one fact about how models read.

A model weighs text by presence, not by freshness. Whatever sits in its context is evidence, and a paragraph written in March carries the same weight as one written this morning. People know the old plan is old. They see the date, they remember the meeting where it died. The model does not. It sees two plans.

So two copies drift, and the stale one wins whichever tool happens to load it. Not always. Often enough that you stop trusting the output, without being able to say why.

That is why the rules are about location rather than content. You cannot fix this by writing better text. You fix it by making sure only one text exists.

I did not follow my own rules

In August I ran a sweep of my own workspace, about twelve project folders, and found the same superseded rule stated as live fact in six different files. Six. Every one of them written by me or by an agent working for me, every one of them confident.

The same week, a context cache I had specified as a 500-word overwrite-each-session summary turned out to be 59 KB. It had been duplicating two other logs, one of 52 KB and one of 433 KB, and my agents had been reading all three, every session, as if they were separate sources.

Cleaning that up did more for output quality than any prompt change I made that month. The agents did not get smarter. They stopped being handed three versions of the same week.

The method, and where it is weak

The numbers come from my own tree. The six-file count is from a text search for one specific retired rule on 5 August 2026, across the whole workspace. The 59 KB figure is the size on disk of wiki\hot.md on 3 August 2026, against its own written specification of roughly 500 words. The 52 KB and 433 KB are the two logs it was duplicating, measured the same day.

Three honest limits.

This is a one-person workspace, about twelve project folders, not a 200-engineer estate. I have not run these rules at that scale, and I would expect the archive discipline to be the first one that breaks there, because moving a document to an archive is a judgment call and judgment calls do not survive headcount.

The claim that the clean-up beat any prompt change that month is a judgment, not a measurement. I did not run the same tasks before and after. What I have is a month of agent output I stopped having to correct for stale context, and no other change I can point to.

And the rules describe a convention, not a control. Nothing in my setup physically stops me, or an agent, from writing a second copy tomorrow. The sweep that found six files is the control, and it only works if someone runs it.

Where to start

Pick the one fact your agents get wrong most often. Usually it is priorities or status. Decide which single file owns it. Find every other place it is stated, replace each with a link, and put the date of the decision at the top of the owning file.

That takes an afternoon. It is the cheapest upgrade I know.

Everyone budgets for models. The cheapest upgrade I know is deciding where the truth lives, and the truth lives in a folder and file structure.

Claude drafted this from my own workspace notes; the rules, the numbers and the mistakes are mine.

Source post: Post | LinkedIn