How AI Work Keeps Its Judgment
You are committed to outcomes, and AI increasingly makes the deliverables that serve them. The trouble is that the AI keeps its own track of what it knows, in a way you cannot easily see, check or correct. The workspace kit gives you and the AI a shared language and a shared set of places for what matters, in plain files you can both read.
01 / 0:00
The gap between your outcomes and your AI
The AI keeps its own record. You can’t easily read it.
You are committed to various outcomes, and often AI helps you with the deliverables that contribute to making those outcomes happen. The trouble is that the AI keeps its own track of what it knows, in a way you can’t easily see, check or correct.
However you track your outcomes stays as it is. The kit does not replace it; it gives each outcome the AI works on a folder you can both read.
02 / 0:16
A shared language, in plain files
From a single task up to the standards you hold everywhere.
The workspace kit is a free, open-source way to give you and the AI a shared language and a shared set of places for what matters, so you can check its work, correct it and keep it on track, from a single task up to the standards you hold everywhere.
It is plain text and folders: projects/ for work with a finish line, ongoing/ for what you keep up, and four shelves for what each session learns. It runs in Claude Code, and its commands can be added to Cowork.
A project is anything with a finish line. Type /projects:new, answer a few questions, and it writes down what done looks like: a spring workshop, say, with its booking page live and the first ten places sold.
It also records who else is involved and how: who is responsible, who is accountable, who is supporting, who is consulted and who is informed, and closeout reminds you who needs to hear what changed. Every session opened in that folder starts from that page, so you and the AI start from the same place.
Some things are never finished, only kept up, and /ongoing:new is for those. Take a rose garden, where the desired state is healthy, attractive rose bushes all season: it asks for the standards you hold it to, like deadheading weekly from May and treating black spot within the week, and how often to look.
A review marks how many standards were met, and turns the fix for a slipped one, when it has a finish line, into a project.
The deliverables are where the AI earns its keep. /projects:pair turns the project’s page into a brief and gives you one command that starts an engineer in Claude Code, while the chat you are in becomes its manager.
Before bed I asked for the booking page and the confirmation email and left the computer on. The two write to each other in plain files, and periodically the manager checks in, reviews what is new and replies.
By morning the mailbox told the story, and the manager had not taken the engineer’s word for it: it tested the booking form itself and found confirmations going to the wrong inbox, which the engineer fixed before breakfast.
Having agents take different roles and check each other’s work from that angle often produces better results, and the kit builds it in.
At the end of a session, /closeout works out what it learned and sorts each lesson onto one of four shelves. It keeps only what the AI could not find for itself, such as a decision and the reason for it, rather than a guide to work it can already read.
Working standards load in every session, so a lesson joins them only when you agree, because every session pays for it. General reference loads when its topic comes up, project reference stays with the project, and templates wait until you start something of that shape. Then it brings the project’s page up to date, and nothing is committed until you commit it.
Once a week, /workspace:hygiene walks the same shelves from the other direction. It measures how much every session has to load, flags files sitting outside the place they belong, and suggests moving a project decision that would help every project up to a shared shelf. If you test your standards, it also suggests moving one that made no difference down to a shelf that loads less often.
It writes what it found to a dated report. Once a month, /workspace:register-audit rereads the guidance for drift toward urgency and emphasis, because a line written in a hurry primes every session that reads it.
None of this asks for your trust, because it is all plain files you can open. The project’s write-up is projects/spring-workshop/README.md, with its decisions beside it in decisions.md and the night’s conversation in its mailbox folder. The rose garden lives in ongoing/rose-garden, the shelves are CLAUDE.md, docs and the templates, and the weekly reports are in audits.
Read the write-up against what you know happened, and if it has drifted from what you intended, ask the AI to bring it back rather than editing the files yourself: it can word the correction the way the model reads best, and check the cross-references between files so the whole record stays in step.
More context going in is loading, but judgment carried out of the session, on purpose and by you, is what makes the next session better than the last.
Follow the Quick start in the kit’s README: install Claude Code, download the kit, and run its setup to make your workspace. Run setup’s Cowork step if you want the commands in the chat app too.
Then /workspace:quick-start asks a few questions and makes it yours. Pick one real project and one area you keep up, and close out every session for a week. The kit is free and open source.
The everyday practice
DRIVE: five steps for working with AI
The kit is how I run DRIVE every day: Decide what done means before you start, Request with the context it lacks, Iterate while you still hold the thread, Validate before it leaves you, and Evolve what you learned into how the next piece of work begins.
AI Effectiveness
Find out where your AI work stands
The Cognition Survey is the free assessment at the start of my AI effectiveness programme. It takes about eight minutes, it is anonymous, and you get tailored feedback and suggestions: how you work with AI today, and where the opportunities are.
Or work on AI with your team
I run AI effectiveness programmes for teams and organisations, grounded in how our minds really work: a whole-team picture of where people sit, DRIVE in everyday practice, and a promotion path so that what one person learns becomes what everyone can use.
Related: AI Improves Itself, the small loop this kit turns; DRIVE, the method; The Trust Trap, what happens without the check
The kit: github.com/cyberscribe/agentic-workspace-kit. The apps in the film are simulations made for it; what they show follows the kit’s own documentation. Narration: my own voice, generated with a local voice clone.
