Tell your AI once.
Put your work, projects and plans in one workspace your assistant reads before it answers. It learns more every week, never forgets what you decided, and tells you when something it knows has gone out of date.
Plan my week.
The release is on the 14th, so Monday and Tuesday go to the checklist. Thursday's client call needs the pricing deck, which your notes say is still in review. Two designer interviews fit Wednesday afternoon, and Friday morning stays blocked.
read: work/product/CONTEXT.md, work/hiring/CONTEXT.md, COMPASS.md, your task app
What changes when your AI has a workspace
You stop pasting context
No more “things to remind my AI” note pasted into every chat. Your assistant reads the relevant files before it answers, so “plan my week” is enough.
It knows more every week
Every decision, constraint and preference you settle is written down once and used from then on. The workspace compounds, and nothing has to be explained twice.
It reads only what the task needs
Working on one niche subject? The assistant opens that area's files and leaves the rest of your life out of it, so answers stay on topic.
One place for everything, or for one thing
Clients, projects, studies, health, money: each gets a folder. Run your whole life from it, or only your work. Start with one area and add the rest as they come up.
The catch: notes don't know how old they are
A workspace is only as good as it is current, and everything you write sounds current forever. A plan from April reads exactly like one from this morning. Developers call what happens next context rot.
Worse, a fact written in two places goes stale in one of them while both still look right. The assistant builds on the wrong one, confidently.
This started with one long conversation with Gemini. The longer it ran, the worse the answers got, and its memory kept pulling in things that had nothing to do with what I was working on. I needed to keep the context but start fresh, one subject at a time. That was the first version, in March 2026. The rules below are what keeps it from rotting.
--- domain: Newsletter status: active last_reviewed: 2026-03-03 --- ### Next Up - Launch the paid tier by 15 April
Your assistant
This file was last verified on 3 March, seven months ago. Before I plan around a 15 April launch: is that still the date?
Seven rules that keep it trustworthy
They came from six months of daily use, and a few from things going wrong.
One folder per domain, one hub per folder
Each area of your work or life gets a CONTEXT.md that says what is true in it right now. The assistant reads that first, every time.
Every file says how old it is
A status and the date it was last verified. Past its review interval, a file is suspect, and the assistant says so instead of trusting it.
One fact, one home
A deadline or a number lives in exactly one file. Everything else links to it, so there is never a second copy quietly going stale.
Learned when a summary file kept a client deadline for eleven days after it had changed.
Decisions are logged, not remembered
An append-only DECISIONS.md per domain. Settled questions stay settled, and nothing gets argued twice.
Learned when an agent filled gaps with decisions nobody had made, and a project's notes had to be rebuilt from scratch.
A linter catches the rot
One small script flags stale files, broken links and dates copied where they do not belong. No dependencies.
A compass for priorities
Domain files never say which area matters most. One file does, so “what should I do now?” has a real answer.
Tasks stay in your task app
Notion, Todoist, Things: the app holds dates and surfaces what is due. These files hold the why.
What's in the kit
Pick a file to see what it does.
AGENTS.md
The rules every agent follows, whatever the tool. Read context first, check freshness, never copy a fact, update what you invalidate.
2. **Check freshness before trusting a file.**
- status: dormant: real, but not being
worked. Never present it as live.
- last_reviewed older than the file's
review interval: say so rather than
trusting it.Works with the tool you already use
The rules live in AGENTS.md, the shared file most AI agents look for.
- Claude Codereads CLAUDE.md, which imports AGENTS.md
- Codexreads AGENTS.md directly
- Cursorreads AGENTS.md directly
- OpenCodereads AGENTS.md directly
Thirty minutes a week keeps it true
The linter catches shape. The weekly review catches the sentence that still reads fine and is no longer true. Your assistant walks you through it.
- 1Leftover workRead what the last session left uncommitted before anything else.
- 2Run the linterIts warnings are the agenda, not the verdict.
- 3Freshness and rotReread what is overdue, and check this week's decisions against every other domain.
- 4The Compass, lastOnly once the domains are true: pick what matters this week.
Start in three steps
- 1Copy the template
Press “Use this template” on GitHub, or download the zip.
- 2Set up your first domain
Open the folder in your AI tool and ask:
Read AGENTS.md and CONVENTIONS.md, then help me set up my first domain.
- 3Run the linter
It tells you what went stale. Run it before each weekly review.
uv run scripts/workspace-lint.py
Questions
Doesn't my AI already have memory?
Some tools remember things between chats, but that memory is hidden, stays inside one app, and tends to pull in things that have nothing to do with what you are working on. A workspace is plain files you own: you can read every word your assistant knows, fix what is wrong, keep each subject separate, and take it all to any tool.
Do I need to code?
No. It is a folder of Markdown files and one optional script. If you can edit a text file, you can run it, and your AI assistant can do most of the setup with you.
Which AI tools does it work with?
Anything that reads files in a folder: Claude Code, Codex, Cursor, OpenCode and others. The rules live in AGENTS.md, the shared format most agents already look for.
Do I need Notion?
No. Any task app works, or a paper notebook. The only rule is that dates you act on live there, not in these files.
How long does setup take?
About an hour to write your first domain properly. Start with one area you are actively working on and add the rest when they come up.
Is my data sent anywhere?
The kit is plain files on your own computer. Whether anything leaves it depends only on the AI tool you choose to open them with.
Want help setting it up?
If you'd like help fitting this to your own work, write to me. I'd also like to hear what you changed.
Hear about the next one
An email when I release something new: the next resource, or a big update to this kit. No schedule, and nothing else.
I'm Amin, based in Tunis. I build AI products, run a small startup and study, and all of it runs from one workspace like this with AI agents. I use it every day. More about me.