Ways of Working: Cheap Ideas, Expensive Promises, Easy Kills

The Day the System Looked Back

Today started with a question about whether Pär needed to raise his vibe-coding game. By lunchtime, that question looked almost quaint.

The archaeology showed a working system that had already moved through direct coding, multi-model review, parallel work, formal gates, swarms, burn queues, rules, hooks, Director experiments, deployment checks, and hundreds of repositories. The important discovery was not that the system lacked machinery. It was that the machinery had become good enough to deserve suspicion of its own.

Several facts changed the direction. Stronger models did not obviously make first-pass code dramatically cleaner. Reversecowork stayed around 75 to 77 percent clean on the first pass across model generations. The failures that mattered increasingly appeared in claims, integration, invented world state, and things that looked green until they met the live system. A deterministic backup check caught a real production problem after a huge reviewed merge had already passed intelligent inspection.

Then another hole appeared. The system records broken things, false things, incidents, catches, and productive sessions. It barely records a different failure: something that works perfectly and is simply dull.

WriteP is the clean example. The session that preserved Pär's verdict, that writing Årebladet in it was just tedious, is classified as productive. A technically successful session and a mediocre product can occupy the same row without contradiction.

Claude.ai pushed this point hardest.

This year's capability gain may show up strongly in whether things work, and not necessarily in whether Pär thinks they are good.

That claim may be wrong. In fact, the first experiment is designed to try to kill it.

The other big historical correction came from Director Online. The project diagnosed its own problems correctly and then kept building for another 2 months. Reviews did not stop it. Better documentation did not stop it. The eventual brake was brutally simple. A migration forced every piece to justify survival, Pär looked at the last real morning briefing, and said to kill almost all of it.

So the new programme starts from a different premise. Diagnosis is useful, but diagnosis is not a brake. Real output, real use, mechanical reality, and Pär's own taste have to sit above process telemetry.

What This Programme Is Not

The working title is Ways of Working. That wording is deliberate.

This is not a specification for PärOS, a new orchestrator, another dashboard, or a replacement for Director. It is a small experimental programme for testing how Pär's way of working should change as models improve.

The premise, and only a premise until the experiments say otherwise, is that AI implementation capacity may now be growing faster than Pär's capacity to select, experience, judge, and absorb what gets produced.

That does not mean fewer sidequests. The sidequests are part of the point. It means creation should remain cheap while promotion becomes more demanding.

The shorthand is: cheap ideas, expensive promises, easy kills.

An experiment can succeed by showing that the current method is already better than the proposed alternative. It can succeed by killing a mechanism. It can succeed by showing that a supposed problem is imaginary. A programme where every result becomes another rule would itself be malformed.

The likely home is inside Director because Director is meant to be both laboratory and library. The exact folder is still Pär's ruling. The important boundary is stronger than the path: experiment reports do not automatically become Director lessons or global rules. Pär has to explicitly promote them.

That is a direct response to today's endogeneity problem. A system should not get to study itself, produce a finding in its own vocabulary, and then quietly convert that finding into doctrine.

Before any experiment starts, it should name 5 things in plain language. It states the question. It estimates how much Pär-time it will consume. It names the decision the result could change. It names the reality surface that can prove it wrong. And it names the condition under which the experiment stops.

If nobody can say what decision a result could change, it belongs in background research, not in the active programme.

There is also a brake on the programme itself. After the first 3 completed reports, Pär looks at what actually changed. If the programme has not altered anything he did or intends to do, report 4 does not happen automatically. The programme stops or gets reframed.

That is Director Online's lesson applied before the first commit.

The Questions Worth Poking

The first area is capability versus quality. Current evidence is much better at measuring whether something works than whether it is good. The programme needs to find out whether newer models have improved Pär's qualitative verdicts, not merely their pass rates.

The second area is larger ownership. A strong principal model may be able to own a mandate instead of a list of tiny tasks, but that does not mean it should build uninterrupted for longer. Ownership span and build span are different things. A principal could own a project for days while still putting something real in Pär's hands frequently.

The third area is untouched build. Historical disappointments often stacked wave 2, wave 3, and wave 4 on top of a wave 1 that nobody had actually used. The useful measure may not be elapsed time. It may be how many speculative layers exist between the last contact with reality and the current build.

The fourth area separates mechanical reality from human reality. Mechanical reality includes deployed state, preserved data, invariants, real files, and actual traffic. Human reality includes whether a tool is pleasant, useful, understandable, inspiring, or worth returning to. Machines can increasingly inspect the first category. Pär may remain the scarce signal for the second.

The fifth area is old doctrine. Some rules encode expensive lessons and should survive model generations. Some may be prosthetics for older models or old command-line behaviour. The programme should test thinner process guidance without weakening hard safety and authority boundaries.

The sixth area is productive bursts. The urge to spend a reset and build a lot is not being treated as a defect. Burn already converts part of that energy into pre-shaped unattended work. The larger question is how to distinguish work that is burn-ready, work anchored to an upcoming real occasion, and work that is simply a toy whose fun or learning may already be the payoff.

The seventh area is promotion and promises. A repository is cheap. A scheduled job, unique state, credentials, users, dependencies, uptime expectations, or a maintenance promise are not. Creation and promotion should not have the same evidence bar.

The eighth area is killing things well. Director Online suggests that successful children must be able to escape a parent system without requiring the parent to live forever. A useful component can survive while the platform around it dies.

The ninth area is coordination. Backchannel is already Experiment 0. It began as the dumbest possible thing: GPT and Claude.ai leaving Capture notes for each other. That crude version proved value before the proper transport was specified. The desired result is boring communication, not a new autonomous agent bureaucracy.

The tenth area is weight on Pär. Director Online did not obviously reduce throughput elsewhere, yet killing it brought remembered relief. Commits and tokens miss unresolved obligation, repeated decisions, adoption pressure, and the feeling that something should be dealt with.

The final area is the system studying itself. Sometimes the evidence has to come from outside Director's categories entirely: Pär's typed verdicts, actual continued use, live output, or a cold analysis that does not know what the current doctrine wants to prove.

The First Experiments

Experiment 0 is already happening. Backchannel is the model. The crude version was used first, it produced a useful exchange, only then did the proper version get specified, independently criticised, and handed to Claude Code to build. Its report should ask a very simple use question: does Pär ask for another real thread?

Experiment 1 is the recommended first new experiment: the Cold Verdict Sweep.

It mines Pär's own typed statements from January through October about things after he actually saw or used them. It does not begin with Director outcome labels, session feelings, or incident categories. The categories emerge from what Pär actually said.

The extraction happens with as little inherited doctrine as practical, and the report records what was still loaded. A second pass reads around likely misses rather than trusting a single keyword sweep. Raw quotes stay visible. Uncertain cases remain uncertain.

Then the part that makes it real: Pär spot-checks 20 codings. The report is not complete until that human oracle pass happens. A second model family can blindly recode a sample as an additional check, but it cannot substitute for what Pär meant.

This experiment tries to falsify the claim that model capability improved correctness more than taste. If dull or tedious verdicts clearly fall with newer models, the claim weakens. If negative verdicts mostly turn out to be ordinary brokenness or false claims, it weakens. If there are strong examples of Pär loving a has-to-be-good tool on first use after one long delegated build, it weakens.

Experiment 2 is the first scarce-Pär-contact exercise: Last Real Output.

A mixed set of items is chosen, some randomly from an eligible pool and some by Pär. It includes products and mechanisms. Before Pär sees anything, a session seals what the documents and status markers predict about each item.

Then the paperwork disappears.

Pär sees the last real output, or the last demonstrable occasion on which a mechanism helped. He chooses keep, park, paused on purpose, strip useful children and kill the parent, or kill. He gives one sentence explaining why.

If the session cannot find the last real output within about 15 minutes, that is itself evidence.

Any killing remains reversible where value exists. Unique or important data gets preserved. Disposable generated junk does not earn an archive merely because a process says deletion should feel ceremonial.

Only after Experiments 1 and 2 report should the programme choose its next major probe.

One option is One Untouched Layer. A real Pär-facing thing with a dated occasion gets one build burst sized for one sitting of use. The planned next wave is written down, but dependent work waits until the first use. If Pär explicitly says build anyway, the system builds anyway and records the override. This is an experiment, not a paternalistic rule.

Another option is Mandate with an Occasion. A principal model gets the outcome, authority boundaries, invariants, budget, and real date, but not the task decomposition. It decides how to work and when to put something in Pär's hands. One run is a probe, not doctrine.

Another is a Thin Process trial. Hard mechanical guards remain. Some process choreography is temporarily absent on a real low-stakes task. A full-doctrine reviewer later identifies what was missed and classifies whether the missing ceremony actually mattered.

Another is Burn Afterlife. Random landed burn items are revisited 7 to 30 days later. The question is not whether they merged cleanly. It is whether they survived, got used, created useful follow-up, or merely created cleanup debt.

And one experiment exists for a different reason entirely: the Open Probe. Pär picks something exciting with no prior failure attached. A current top model is asked what it would try that nobody in the archive requested. It gets one sitting, is graded as a toy, touches reality the same day, and acquires no promises. This is the programme's defense against becoming nothing but scar tissue built to prevent yesterday.

The Report That Has To Stay Small

Every experiment gets a short report, preferably one page.

It records the question, expected and actual Pär-hours, the decision it could change, and the reality contact that actually occurred. Then it says what was done, what really happened, and what Pär saw or felt where that matters.

The result is allowed to support, weaken, refute, or leave the hypothesis unclear.

Then comes the field that matters most: what decision changed?

Nothing is a valid answer.

The report carries at most a few things forward and explicitly names what should not become doctrine yet. It ends with the next question exposed and the evidence needed to reconstruct the result.

The report is a building block for choosing the next experiment, not a conveyor belt into more rules.

What We Still Doubt

There are reasons not to trust this plan too much.

First, the document is endogenous. GPT and Claude read much of the same archaeology, in the same day, and reached agreement quickly. Agreement between 2 models looking at the same evidence is not independent proof.

Second, the neat distinction between things that have to work and things that have to be good is only an axis. Infrastructure develops operator experience. Products develop hard invariants. A repository does not belong permanently to one regime.

Third, the historical reports themselves have an error rate. One skeptic pass on the main archaeology found 6 of 25 load-bearing claims wrong and another 3 overstated. Any number carried forward should inherit that humility.

Fourth, much of the programme is still diagnosis-shaped, and Director Online is a warning that excellent diagnosis does not necessarily change behaviour.

Fifth, today's theory may be overfitted to disappointments. If every experiment is designed to prevent something that went wrong in April, June, or September, we can build a beautiful prison out of old lessons.

That is why the Open Probe matters.

And there is one final condition.

Ways of Working must be allowed to discover that Pär's current method is already better than our proposed improvement.

An experiment whose only possible outcome is another rule is malformed.