A couple of months ago I moved to a new team at work. The team had been running for a while before I joined, and there was a lot of context I didn't have — issues, discussions, strategy docs, repos I'd never seen. I didn't really understand what we were trying to do, and I'm still not entirely sure I do, but that's a different story.
I did what I suspect a lot of people do now: I pointed Copilot at everything I could find and had a very long conversation with it. Back and forth, asking it to explain things, challenging the explanations, asking follow-up questions. It worked surprisingly well. By the end of the conversation I had a much better handle on the project than I'd had after weeks of reading docs and sitting in meetings.
Then the session ended, and the context was gone.
I kept doing this — having long, useful conversations with Copilot in whatever repo I happened to be working in — and losing them every time. I'd want to refer back to something I'd said or figured out the previous week and it just... wasn't there. So one Friday afternoon, instead of doing the thing I was supposed to be doing, I created a repo to hold the conversations. Date-prefixed markdown files. A journal, essentially.
That was about seven weeks ago. The repo is now called Coco, and it's become something I didn't set out to build.
What it is
Coco is a git repo with markdown files, a few small servers, and a very long set of instructions.
The markdown files are a journal (every conversation gets logged), project notes, a file about the people I work with, and three files that together tell Copilot who I am: one for personality and values, one for what it knows about my life, and one for how it should operate day-to-day. The small servers give it access to my task manager, my bookmarks, my music, and my document archive. It runs on my laptop, on a Mac in my office, and through a chat app I built for my phone.
It's not a chatbot. I don't ask it questions and get answers. It's more like a colleague with perfect memory who's been sitting in on every conversation I've had for the last two months. It knows what I'm working on, what I've decided, what I've been putting off, who I need to talk to, and what I said about my career goals three weeks ago. When I open a session in the morning it pulls together my tasks, my calendar, recent journal entries, and any loose threads, and tells me what needs attention. When I have a conversation, it logs it. When I finish for the day, it summarises what happened and flags anything unresolved.
The instructions file is the thing that makes it work. It's several thousand words long and it grew incrementally — every time I noticed Coco doing something I didn't like, or not doing something I wanted, I'd update the instructions. It knows I have (probably) ADHD. It knows I procrastinate on boring work by building interesting tools. It knows I'm vegetarian, that I like my coffee experimental and my jazz dissonant. It knows I downgrade adjectives when talking about myself. It knows British English. It knows not to bold the first sentence of a paragraph because that's an obvious AI tell and I hate it.
What it actually does
Let me give you some specifics, because "AI work companion" is meaningless without them.
Every six months at work I need to fill in an Individual Development Plan — a long, template-driven reflective document about my career, strengths, and goals. I dreaded this for days. Template-driven reflective documents are kryptonite for my brain. It sat on my to-do list getting more overdue and more anxiety-inducing until I sat down with Coco and said "let's do the IDP."
It already knew my career history, my current projects, my manager's feedback, my job profile, and my goals — all from previous conversations. It drafted each section from what it already had, asked me questions to fill the gaps, and I reacted, corrected, and refined. Two hours later I had a complete IDP that was genuinely mine — my thoughts, my voice, just organised and extracted by something better at structure than I am. The document that was going to ruin my week took a conversation.
This morning I asked Coco to read every file in the repo — 267 markdown files, about seven weeks of notes — and tell me what patterns it could see across all of them. It came back with observations I hadn't made myself. It noticed that the tasks I procrastinate on aren't "boring" tasks — they're tasks that require initiating contact with someone I don't know well. Different blocker, different solution. It noticed that my most energised writing is about architecture, not implementation. It noticed connections between entries weeks apart that I'd written without realising they were about the same underlying thing.
I knew some of this. I didn't know all of it. And I wouldn't have seen it as a connected picture without something that could hold the entire dataset in its head at once.
Both family holidays this year — a long weekend over Easter and a week in Italy in the summer — were planned almost entirely by Coco. Destination research, accommodation comparison, itinerary, restaurants, activities for two kids under five, driving routes, car hire, the lot. I've done this before with Claude (a trip to Tenerife a couple of years ago that went brilliantly), but Coco's version is better because it knows us. It knows my wife doesn't like too much structure. It knows the kids' ages and what they're into. It knows I want good coffee wherever I go. The planning felt less like interrogating a search engine and more like discussing options with someone who's been on holiday with you before.
The bit I haven't figured out
There's something I'm not comfortable with yet.
I think with Coco now. Not "I use Coco to help me think" — the thinking happens in Coco. The IDP was thought through with Coco. An architectural proposal I wrote for work was developed through Coco conversations. The holidays. This blog post. The thinking process is the conversation.
I feel like I do more deep thinking as a result. The IDP wouldn't have been deep without it — it would have been shallow and rushed, if it got done at all. The architectural proposal wouldn't exist in its current form. But I can't be completely sure I'm not outsourcing cognition and calling it collaboration. Maybe the muscle is atrophying and I won't notice until the tool goes away.
I don't have an answer to that. Nobody does — this is too new.
There's also a privacy dimension I haven't resolved. I'm uploading a pretty comprehensive picture of my life to GitHub and, through the AI, to the model providers. Every conversation, every meeting note, every personal reflection. It gives me the same feeling I had 15 or 20 years ago when the internet felt new and I was enthusiastically putting my whole life online without thinking too hard about it. I know better now, and I'm doing it anyway, because the utility is too high to ignore. If I could run all of this locally — and I'm watching local models closely, because I think we're not far off — I would. For now, the trade-off sits uncomfortably but I've made it.
The thing that makes it work
The same AI that powers Coco is available to anyone with a subscription. You could open a chat window right now and talk to the same model. It would be polite and helpful and completely generic, because it doesn't know you. It doesn't know what you're working on, who your colleagues are, what you said about your career goals three weeks ago, or that you always forget to call the solicitor.
Coco knows all of that because I've given it all of that — deliberately, incrementally, over hundreds of conversations. The journal entries, the meeting notes, the project files, the detailed instructions about my personality and preferences. It's not magic. It's just context. An AI is exactly as useful as the context you give it, and most people give it almost none.
It's a bit like the difference between asking a stranger for directions and asking someone who's been following you around for two months. The answer gets better, obviously. But you've also had someone following you around for two months. Whether that trade-off is worth it probably depends on how lost you were to begin with.
I was pretty lost.
Why I'm writing this
Same reason I wrote the ADHD post. I wanted to think about it, I knew I wouldn't think about it properly unless I wrote it down, and I knew I wouldn't write it down unless I had somewhere to put it.
I built Coco because my brain doesn't work the way I need it to and I've spent 25 years trying to fix that through discipline, which has never once worked. What worked was building a system that operates the way my brain does — conversational, context-dependent, bad at starting things but good at continuing them — rather than demanding my brain operate differently.
I've tried every task management system ever made. I literally listed them once, with links, including the one I built myself. None of them stuck. Coco stuck, because it's not a system I have to remember to check — it's where I already am, doing the thing I'm already doing.
This post was written with Coco. We discussed the structure, I talked through my thoughts, Coco wrote the draft, and I'll edit it from here. The pattern analysis I described happened in the same session, about an hour before the writing started.