AI context system

I carry several clients and my own company at once, and what we agreed is spread across call transcripts, PDFs and WhatsApp threads. An AI can quote any of them perfectly and still give me the wrong answer.

A client agrees to a website and an app on Monday. On Thursday they confirm they only want the website. On Friday someone sends a summary that still includes both. All three files are accurate records of what someone said, and only one reflects the current agreement.

That is two problems: getting the material in at all, and keeping track of what still applies.

A knowledge base of plain folders and files that Claude reads from and files into, in use since July. The framework was built by a fellow AI enthusiast: the Intake folder, the three layers per area, and the instructions Claude files by. The base inside it is mine, and so are the conventions it keeps for time.

One intake. Transcripts, PDFs and messages all go into a single folder without my deciding where they belong. When I ask, Claude files each item under its client, project or topic.

Three layers per area. Raw keeps the originals untouched, so I can always go back to what was actually said. Knowledge holds the maintained notes: facts, decisions and open questions, each with its source and date. Assets holds the work produced from Knowledge, proposals, emails and reports, each with a note of the facts that went into it, so when a fact changes I can see which pieces still carry the old one.

Keeping track of what applies. Updates are dated and kept in order. When a confirmed change replaces a decision, the earlier one stays on file and is marked superseded. Where two sources disagree and nothing has settled it, the conflict is flagged and I decide. In the example, all three documents stay in Raw, and Knowledge records Thursday's scope as current, Monday's as superseded and Friday's summary as conflicting. Claude is instructed to draft a revised proposal from that Knowledge and check the originals where needed, so the one saved in Assets should cover the website only.

The image below follows the same example through the folders. Its content is invented, because the real base is client material throughout.

It is what I work from. Call preparation, proposals and client follow-ups are drafted out of it, so the next task starts with a maintained account of what we agreed, where it came from and what changed. I don't have to reconstruct that history in every new chat.

It still takes judgment: I review the conflicts it raises, and it knows only what I have put into it.

How are you keeping changed decisions current in your own AI setup?