Stop re-explaining yourself to AI

The efficiency AI promised quietly disappears the moment you cannot be sure what actually got remembered, or find yourself rebuilding context anyway because retrieval missed it, guessed wrong, or never reached the tool you are using. Here is what that looks like in practice.

The system is built from three plain files and a short ritual at the start and end of every working session.

A context document solves the first-day problem by telling an assistant what matters. A Working Context System also solves the hundredth-day problem: it separates current rules from history, routes lasting facts to the right record and gives every session the same governed way to propose and approve change.

The same three layers hold the two things a working life actually runs on: outstanding actions and live relationships.

Native memory has got good, and this argument does not need it to be bad. Both of the main assistants let you see what they have stored, correct it and delete it, and one will cite the earlier conversation a claim came from. None of that is a black box.

You can see and edit a good deal of it. What you cannot see is all of it. One vendor states in its own documentation that the summary shown to you may not include everything it remembers, and that the sources it displays may not reveal every factor that shaped a response.

The distinction is narrower and it holds anyway. Native memory optimises continuity and relevance. A Working Context System governs currentness, provenance and authority. Recall asks what is likely to be useful here. Governance asks what is currently true, what replaced what, and who agreed to it.

Governance sounds like approving a proposal at the end of a session. That is the most visible part and the least important. In practice it is four things.

The assistant proposes, sorts and drafts. What stays yours is refusing, retiring, filing and ruling.

The obvious objection is that bigger context windows and better retrieval will make a curated record unnecessary. They will not. A large knowledge base, retrieved intelligently, still leaves the assistant to work out which of two documents is current and which one lost the argument. Scale is a retrieval problem and it is being solved. Authority is an editorial problem, and it stays with you.

Checked against both vendors' own documentation, most recently on 4 August 2026. What is available to you also depends on your plan, settings and surface, and these particulars change quickly.

These are concrete examples of the wider portability problem: continuity can still be scoped to a product, mode or project. A Working Context System keeps the record people need to inspect, correct and carry between those boundaries.

I have resumed a client thread weeks later without rebuilding the brief because the current decision, its history and the documents to open were already in the record. That is an observed use case, not a claim that every session will recover context perfectly.

Colleagues can use different assistants against the same record, provided each assistant has appropriate access to the shared workspace. What is shared is the deliberate organisational context, not anyone's private conversation history. Nobody has to standardise on a vendor, nobody pays for seats they do not want.

A shared source of truth needs ordinary access controls and an agreed editing routine. The append-only Log preserves history, while dated entries and approval make it clear what changed, who agreed it and when. The workspace itself still has to manage permissions, version history and simultaneous editing.

Four working files, a session ritual skill, a START HERE guide and a setup note. The complete introductory Working Context System, free to use, adapt and share.

You can hand the pack to your assistant and ask it to install the skill.

The free pack uses Markdown: portable plain-text files that people and AI assistants can both read. They work in an ordinary text editor, but a free Markdown application such as Obsidian will display the headings, links and emphasis as a formatted document.

Markdown uses a few simple markers, such as # for headings and ** for bold text. Because it is plain text, it is easy to search, update, compare and move between tools without locking the record to one product. Obsidian is free to use, including for work; its optional Sync and Publish services are paid.

The assistant needs both read and write access to the workspace.

The honest way to start is to try it on whatever you already pay for, or do not pay for, before changing anything.

The first version was inspired by MJ Jaindl's four-file Claude setup shared on LinkedIn. I adapted it through daily use to meet my own needs, separating current rules from history, durable reference from tasks, and useful material from the things that should be recorded nowhere.

This is a practitioner method, not a research finding. These sources explore adjacent ideas about persistent instructions, external memory and durable Markdown-based context.