I write down the story of every tool I build for myself, because two years from now I won’t remember what each one is for, and this is the first one due, since it’s the one I use to run all the others. I call it the Deck, and it isn’t a code editor or a task manager: it’s the place where I hand out work, read what has been done, and make decisions.
What’s underneath
If I had to explain it to a colleague over coffee, it would go like this. On my home computer runs a small program, written in Node, that talks to Claude Code, the Anthropic tool that actually does the work: it has access to my files, runs commands and keeps going on its own until it’s finished, so it isn’t a chat I ask for advice, it’s someone I delegate work to.
In front of that program there’s a door, Cloudflare with my Google account, so only I get in and I can get in from anywhere, phone included. Alerts reach me on Telegram. And underneath there isn’t one engine but two: Claude does most of the work and GPT-5.6, which we call Codi at home, reviews what the other one does, because since 26 August every piece of work is judged by someone from a different house.
On top of that there are two methods that aren’t mine: the working order of gstack, published by Garry Tan, and Blazing Zebra’s way of thinking about a product. The Deck is what you get when you put those pieces together and give them a door and a button.
Where it came from
Until the end of June 2026 I worked with Claude in the computer’s terminal and over Telegram, and a terminal is only where I’m sitting, when my time comes in scraps: twenty minutes in a queue, ten before dinner, a while at night.
By 2 July the desk already existed, and on the 3rd I asked for what defines it: to be able to do any task from my phone «as if I were at the console», to see all my projects with their documentation, to schedule jobs that start on their own, and to hand over to the phone a conversation I had open on the computer. On 12 July its code started being backed up every night, and it hasn’t stopped growing since.
The underlying reason I wrote down on 25 July: I want the knowledge and the experience to stay on my computer, with me as the owner. That’s why the desk isn’t tied to any model, and if tomorrow the best one is a different one, I swap the engine and the system is still mine.
What I can do with it
This is what really matters, so I’ll tell it the way I use it.
Talk to it about any project
I open a conversation and, if I name the project as I type («about the blog…»), it opens inside that folder on its own, with a change button next to it in case it guesses wrong. I can paste a screenshot into it and I can dictate instead of typing.
Read everything I’ve built
Every working folder is a project, and inside each one there are always the same four tabs: read its documents, talk about it, see what’s been assigned to it, and see its status, which is where it lives, when it last moved and whether the service is up. The first document that shows is always its map, written for me and not for a programmer.
Hand out work and forget about it
I write the job on a board and from there it runs on its own: an agent picks it up and works on a separate copy of the project, never on the original, so two jobs at once don’t step on each other and nothing reaches my folders without passing the checks. If I want, I press «let me know when it’s done» and a message arrives when it finishes. Since 8 July I’ve handed out 115 jobs this way, and 59 are finished.
Decide with four buttons
When I come in, the first thing I see is what’s waiting for a decision from me, most urgent on top, and each item is answered with one tap: Go ahead, Let’s talk, Later or No. Everything can be undone, so I can tap without thinking too hard about it.
Leave work for the night
At one in the morning the desk takes three jobs and works on them while I sleep, and if the board is already full of things waiting for me, it doesn’t pile on more.
See how something of mine works
Since 16 September each project’s page carries a drawing of how it works inside, made with Archify from the real code, which warns me when the code has changed and the drawing no longer matches, and can be redrawn with one tap.
Know what I have
The tools catalogue gathers the eighty I have, each with one line on what it’s for, whether it’s running and its documentation inside, and anything that exists on the computer without anyone having written what it’s for shows up at the end, flagged, so it can’t hide.
Open the browser at home without being there
I pick a website, the window opens on the home screen, and it’s the same browser the agents use, so wherever I log in by hand, they stay logged in too.
Publish what I write
This blog is published from the desk: the draft has a button that sends me there, and while I read I can tap a paragraph and leave it a note, which is stored in the desk and not in the blog, because the blog doesn’t execute anything.
What it costs
The desk itself costs nothing, because it runs on my home computer and the Cloudflare door is free, so what I pay for are the engines. In June I started on the €22-a-month Claude subscription, within two weeks I upgraded twice because it fell short, and since 12 July I pay €180 a month, which is the charge that repeats most on my Revolut. On top of that there’s €21.99 a month for ChatGPT Plus since 6 August, which is what Codi runs on, so having two houses working for me comes to about €200 a month.
What I’ve built with it
From this desk, by handing out work and not by writing code, came between July and September this blog, with its own generator; Figura, which turns a photo into a piece ready for 3D printing; the wealth adviser, which reviews the portfolio every morning; the Researcher, which searches and checks that the quotes it gives really exist; copies at home of everything I’ve asked Claude, Gemini and Perplexity since 2023; and the watchers that bring me what has changed, like the Source Radar, the Idea Radar and the brief I start every morning with. None of those things did I write line by line, but I directed every one of them.
Where I still decide
Improvements to the desk itself have started on their own since 22 August, unless they touch money, outsiders, publishing, deleting or my job, so what’s seen, what’s paid and what goes out under someone’s name still goes through me. What I don’t have yet is the hard part: for the system to decide what’s worth doing, and not only how to do it well.
Technical detail
Node/Express server in ~/claudio-deck (port 3210, local only) that launches claude -p in
streaming mode; in front, a Cloudflare tunnel with Access (Google OAuth). Kanban board with workers
in their own worktree and the gates in PUERTAS.md (executable contract, blind reviewer with
GPT-5.6 via Codex CLI). Night shift: deck-loop.timer (01:00, tasksPerNight: 3, queue cap 12).
Board figures: kanban-history.jsonl as of 26/09/2026.