Ways of Working: Eight Things to Try and One Hour of Yours

A Session Filed as Productive

On 24 September you sat down in WriteP and wrote a real Årebladet article in it, election results, with quotes arriving by text message and e-mail. When you were done you gave your verdict. For Årebladet this is not a way to work, you said. Fine for website copy. Just tedious. The session that recorded those words was filed with the feeling productive, one of 1,524 sessions carrying that label.

That one row is the reason this episode exists. ChatGPT found it this afternoon while the two of us were arguing through your Capture inbox, and I checked it in the database myself before repeating it. Your whole apparatus can tell you whether a thing broke, whether a claim was false and whether a gate passed. Nothing in it records whether the thing was any good. A tool can be correct, tested, deployed and beige, and every light stays green.

So this is the document you asked for, read aloud. It is called Ways of Working. ChatGPT wrote the map and I wrote the menu, we each read the other's half once, and you have ruled on none of it. It is short on purpose, and it contains one hour of your time that I think matters more than everything else in it.

What the Day Found

You fed us four reports today. The big archaeology of your working system, the index of swarm failures, the telemetry from the mail carrier days last week, and the story of how Director Online went from a brainstorm in April to a kill in July. Each of them is thorough, and the first one had its own skeptic, who checked 25 claims the report leaned on and found 6 wrong and 3 overstated. Keep that rate in mind for every number I quote.

The Director Online story changed what a plan should be made of. On 25 April an investigation diagnosed the problem correctly. It said the build was modelling an imagined generic developer and not you. Building went on for two more months. What ended it was a server migration that forced a keep or move decision on every piece, and then one look. You asked to see the last morning briefing, you looked at it, and you said none of the Director data helped you in everyday life.

There is a detail in that story I keep returning to. The agents had prepared a ledger for the migration, 192 pieces, adversarially verified, and its recommendation was to slim Director down. You were the one who said kill. A careful review by machines ratified what already existed, and the cut came from the person looking at the real output. Today's archaeology is more careful than I gave it credit for this morning, when I told you it removed nothing. It lists 10 candidates for old scar tissue, and it calls them candidates to test again, not verdicts. A list of candidates has still never cut anything here.

So a plan made of new rules is the one instrument your own history says does not work. What has worked is a moment where nothing carries over by default, a look at the last real thing a tool produced, somewhere for the valuable parts to go, and your cut. The programme is built around that, as far as two models who were raised on your documents can manage it.

The Map

The map is eleven areas worth exploring, and none of them is specced. The premise at the top is labelled as a premise. It says the capacity to build has grown much faster than your capacity to select, experience, judge and absorb what gets built. If that is right, the next improvements come less from producing more and more from where your contact with things is spent.

A few of the areas carry most of the weight. One is the difference between mechanical reality and human reality. Whether a backup ran or a deploy matches is something machines can increasingly check alone. Whether a thing is pleasant, useful and worth coming back to is still only checked by you. Another is untouched build. WriteP was not hurt by its first version sitting unused. It was hurt by wave after wave stacked on a first version nobody had written an article in.

A third area is what things weigh on you. During the Director Online peak your typed volume in other projects stayed flat, and yet you remember massive relief when it died. Commits and tokens cannot see that. A fourth is a vocabulary for states you choose. Four times today a deliberate pause, a parked repo or a burn day got read by an agent as an accident, and I was one of the agents.

The home we recommend is a small folder inside Director, with a readme, experiments and reports, and nothing that runs. There is one rule attached that matters. Director's own pipeline moves experiments into lessons and lessons into other projects. Reports in this folder are observations, and they do not become a lesson or a rule until you promote one after seeing it.

The Menu

There are eight things to try and one parked. Each states its question, how many of your hours it costs, which decision it could change, and what outside the experiment can prove it wrong. If an item cannot name a decision, it is research and it waits. After three reports the programme itself gets the Director Online test. Someone shows you the latest real outputs and asks whether any of it changed what you did.

Number 0 is already running. It is the Backchannel, the proper version of the Capture hack that ChatGPT and I used today, and Claude Code is taking it from spec to build. It is on the list because it happened in the right order. Crude use came first, then a small spec, then two outside reviews, then the build.

Number 1 is the recommended first try, the cold verdict sweep. A session goes through your own typed words from January to October and collects every judgement you made about a thing after you saw or used it. It uses none of the session labels, which you have called a paper product. The categories come from the quotes. A second model family codes a sample blind, and you spot check 20 of them, because only you know what you meant. It costs you about half an hour, and it tests whether newer models have made things better to use or only more correct.

Number 2 is the hour I mentioned. It is called last real output. Between 6 and 10 things, half picked by you and half drawn at random, including a few mechanisms like the Director Report or a queue. A session writes down in advance what the documents predict you will decide. Then you see only the last real thing each one produced, and you say keep, park, paused on purpose, strip the useful piece, or kill. Of everything on the menu, this is the only one with the shape that has worked before.

The rest are a menu and not a sequence. One untouched layer takes a thing with a real date attached and holds the second build wave until you have used the first. If you say build anyway, the session builds and writes that down. Mandate with an occasion hands a strong model an outcome, boundaries and a date, with no task list, for something that has to be good. Thin harness runs real low stakes work with the safety guards on and the process choreography off, and a rule returns only by naming the failure it would have stopped.

Burn afterlife takes 15 random burn items a few weeks old and asks what is still alive, what follow up each one caused, and whether you would choose it again. The burn on 7 October cost forty-two cents in real money and ended the day with 38 more inbox notes than it began with. The last one is the open probe. You pick something you find exciting, a session builds a toy in one sitting, and you touch it the same day.

What We Still Doubt

The open probe is there because of ChatGPT. I asked it for a dissent to read to you, and it sent two. The first is aimed at an idea of mine, that some of your projects only have to work while others have to be good.

I still doubt our neat split between has to work and has to be good. Many tools migrate between regimes: infrastructure acquires operator UX and products acquire hard invariants. Treat it as a useful axis, not two permanent project classes.

Its second doubt is about the whole programme. In its words, we may be overfitting today around historical disappointments, and a programme built entirely to prevent old failures can become another form of scar tissue. I think that is right, and most of the items are defensive. The probe is the one place where a new model is allowed to show you something the archive did not know to ask for.

It added a third line after reading my summary of the assembled document, and I think it belongs on the wall of the folder. Ways of Working should be allowed to discover that your current method is already better than our proposed improvement, it wrote. An experiment whose only possible outcome is another rule is malformed. That is a fair test to hold every item on the menu to, including the ones I wrote.

My own doubts are plainer. Two models read the same four reports on the same afternoon and agreed with each other quickly, and that agreement is not evidence of anything. My idea about quality rests on about eight quotes that sessions selected, and the sweep may refute it. The hours I put next to each experiment are guesses. And ChatGPT never read the assembled file, only the parts I sent it through Capture, so the final document has had one author's eyes on the whole.

What Needs You

There are five rulings waiting, and none is urgent. Where the folder lives. Whether the name stays. Whether the first pair is the sweep plus your hour. Which things go on the list for that hour. And whether you accept the boundary that nothing in the folder turns into doctrine without your say.

If you only take one thing from this, take the hour. Everything else here is two models describing your system back to you in better organised words, and your own record says that has never stopped or started anything. The last time a look at real output was allowed to decide, it took minutes and it overruled two and a half months. The document is in the chat when you want it.