Chaos Development
Chaos Development is the half of the method that decides how work is cut and built: how big a piece is, what is written down before it starts, who owns it, and where that owner’s freedom ends. Inside the team it is also called chaos programming. It is the same thing.
The trunk, not the ticket
A trunk is a large piece of work with exactly one responsible person and full autonomy inside it. It is cut at the problem’s natural seam, not at the size a sprint can hold. If a trunk needs smaller slices, its owner slices it for their own head, and those slices are not a shared artefact.
An invented example using MojoPaste's four parts; MojoPaste itself was built by one person. On a sprint board, a seam can end up spread over several sprints and people, each piece with its own estimate. As trunks, each seam is one piece of work with one owner and a status line, and no estimates.
The example is MojoPaste, our clipboard app, with invented tickets. Its code has four natural seams: capturing what you copy, keeping the history, showing the panel, and pasting back. On a sprint board each seam would be cut into small tickets, which could end up spread over several sprints and people. As trunks, each seam stays whole, with one owner from start to finish.
Why now: what AI changed
A user story is a compression format for two human limits: small enough to hold in one head, self-contained enough to survive being handed between two people. An AI has neither limit inside a trunk. It re-reads the spec at the start of every session and holds more of it at once than any of us, so the slicing step buys nothing and costs a real fraction of every week.
What does not go away is the other half, and it gets sharper. An AI needs a crisp definition of done more than a person does, because its characteristic failure is building the wrong thing quickly and with total confidence. So the time saved on slicing is spent on invariants and acceptance rather than pocketed.
A story is two things bolted together, and the method keeps one. The intent survives: what the work is for, in the user’s terms. The allocation unit is rejected: a story sliced to fit a sprint, estimated, assigned and burned down is sized by the reporting cadence, not by the problem.
Four artefacts replace the board
- Ideas. An ideas file is where a thing goes before it is a plan. Nothing in it is committed to and nothing in it is scheduled. Depth of description is near zero on purpose.
- Specs. One document per trunk, written when the trunk is committed to, not before. Well described: enough that someone entering cold knows what must be true when it is done, and no more.
- Epics. A document that indexes several specs under one intent, when the work genuinely spans several trunks.
- A header. The first thing in every spec: a status line in plain words, the name of the one person responsible, and their own few sentences on why they took it.
specs/capture.md
- Status
- specced; the first slice built
- Owner
- The one person who took it
- Why me
- Two or three sentences, in their own words, on why they want this trunk and what they bring to it.
What must be true when this is done
- Clipboard contents are never written to disk.
- What a password manager marks as secret never enters the history.
- Gates: one test for each line above.
Header
A fact about the present, written by the person it describes. It needs nobody's approval, so taking a trunk never waits for a meeting.
Body
The what. Team property. It changes only by team decision.
An invented spec, built on two of MojoPaste's real invariants. The header is the whole tracking layer; the body is the decision.
The status line is the entire tracking layer, and it lives inside the artefact rather than beside it. It is written in plain words, such as V1 built 2026-08-14, or specced 2026-08-14; the first slice built 2026-08-15. Trunk status is not review status: a spec that also carries an audit state keeps it on a line of its own, because merging two facts makes the header lie on the first day.
An idea earns a spec when it is small and well understood, or a place on the list of hard problems when being wrong about it would be invisible from the interface. Promotion is the only scheduling event the method has. Depth of description scales with commitment: a beautifully specced thing nobody builds is the most expensive artefact the method can produce.
Well described, not exhaustively described
A spec states what must be true. It does not state every step to get there.
- Write invariants, not procedures. Clipboard contents are never written to disk and what a password manager marks as secret never enters the history are two of MojoPaste’s. They survive every refactor. A numbered list of steps is stale in a week.
- The rest is spoken. Detail that did not make the spec is worked out in the daily and in review meetings. A spec is not the transcript of a decision; it is the decision.
- Acceptance is worth more than description. “Done” needs a definition that is not one person’s opinion: a test, an audit line, a changelog entry written for a person.
Done is defined per trunk, never globally. A single checklist every piece of work must satisfy paces the fast work to the slowest rule, and it feels like being managed, which is the one thing this method spends its whole budget avoiding. Each trunk states what must be true when it is done and the gates that prove it, written by the people who will have to meet it. A project’s safety floor, the handful of things true of every change regardless of trunk, security and secrets among them, is a separate object and does not move.
How a trunk is taken
A trunk has exactly one responsible person, and that person chose it. Work is pulled, never pushed: nobody is assigned a trunk, nobody is handed the one that was left over, and there is no round of allocation at the end of a meeting.
Taking one is written in the spec’s header: the name, and a few sentences in the person’s own words on why they want this one and what they think they bring to it. Two or three sentences is the whole ceremony. They tell the rest of the team which angle the trunk is about to be attacked from, because the same spec in two different hands produces two different pieces of software. And they are a cheap honesty check: somebody who cannot write three sentences about why they want a trunk has usually not decided that they want it.
The header is not the spec. The body is the what: team property, changed only by team decision. The header is a fact about the present, written by the person it describes, and it needs nobody’s approval. Taking a trunk therefore never waits for a meeting.
- Freedom to take is freedom to put back. Releasing a trunk is announced in the header, never done silently, and it is not a failure. The claim stays in the file’s history, so whoever picks it up next can read why the last person wanted it and what they found.
- Two people wanting the same trunk take it to the meeting. Splitting it, or pairing on it, are both real answers.
- A trunk nobody pulls is information about the spec, not about the team. Unclaimed work is usually badly described, not actually valuable, or larger than it looks.
Where autonomy ends
- The spec is the what. It belongs to the team, and changing it is a team act.
- Everything inside it is the how. Architecture, order of work, naming, tooling, when to stop for the day. Nobody’s business but the owner’s, and not a topic for the daily.
What is built before a trunk is taken is the frame, not the work inside it: the definition of done, the spec, and the frame the owner will code inside, meaning the seams, the types and the tests that whatever is built later must pass. The implementation is the owner’s. Building the frame first is what lets the trunks run in parallel.
When a decision is genuinely ambiguous, ask which side it would surprise someone on. A choice that changes what “done” means is a spec change. A choice that only changes how you got there is yours.
Duplication, on purpose
Two people building the same thing by accident is waste. Two people building the same thing on purpose, and comparing the results, is an experiment, and with an AI in the room it is now cheap enough to actually run. The method asks for both, in this order:
- Map what can be reused, before the work starts. What already exists, what two trunks would each need, what is about to be written twice. Accidental double work is still waste.
- Then decide, out loud, where to duplicate anyway. A component built twice, by two people who each solve it their own way, produces two real implementations instead of one meeting’s opinion about which approach is better. Compare them once they exist, keep the better one, and apply it everywhere.
This is A/B testing applied to development itself, and it is “fail fast” with the cost finally low enough to mean it: the losing version is discarded after it has told you something, instead of being defended because it was expensive to build. What keeps it a method rather than a mess: the duplication is chosen, named, and bounded by the comparison that ends it.
- One attempt
- Two attempts, compared
- Gain from the second attempt
Chance that at least one approach turns out right
Chance that a single approach turns out right
One attempt: 50%. Two attempts, compared: 75%. Gain from the second attempt: 25 points.
Arithmetic, not a measurement. It assumes the two attempts succeed or fail independently; two people solving the same problem their own way makes that more likely, not certain. The gain is largest when nobody knows which way is right: 25 points at a 50 per cent chance, and close to nothing when the answer is already obvious. That is where duplicating on purpose pays most.
The half-finished version is not trash
The second implementation does not have to be finished to be worth something. Its finished parts can be compared with the same parts of the version that shipped, a kind of half A/B test, and the commits show how each developer went about the same problem, how fast and how accurately. The founder asked for this to be told to every developer before they take a trunk, not after:
[…] I do not want them to feel demotivated because their work was not implemented in the end. It hold[s] value to the project, whatever amount of work has been finished. Possibly, the developer that finished after one week actually did implement the feature with many holes and the work from the second slower developer would highlight that.
Joé, founder, 26 September 2026.
Alignment is loaded in advance
With stories, alignment is enforced at the ticket: somebody sizes it, somebody reviews it, and a wrong assumption is caught before much is spent. Hand people trunks instead, and there is no checkpoint at the moment a decision is made. Alignment therefore has to be pre-loaded.
That is the job of a philosophy file: a decision made once, so it does not get made five times, differently, by five autonomous people, none of whom is doing anything wrong. The test of such a file is whether it says no. A file of aspirations, such as “we value quality”, coordinates nobody. A file of refusals coordinates everything. Three examples from our projects’ philosophy files:
- Every trade between privacy and comfort ships private. Comfort is opt-in, with its cost on the label.
- No superlatives unless a benchmark or an audit backs them.
- No dash punctuation in user-facing copy.
The usable measure: can two developers who never speak to each other reach the same call from it? If yes, it is philosophy. If it needs a meeting to interpret, it is a mission statement.
What an AI may write
In our projects an AI is the one participant present at every change, and the only one that never forgets to write something down. So what it may document is decided in advance: the header may be maintained, the body may only be proposed.
Written without asking
Facts about what happened, in the header or in places that only grow.
Status lines, kept true to the code Findings noticed while doing something else, written down unfixed Changelog entries, cross-links and index lines
Never written by an AI
Decisions and declarations belong to people.
The body of a spec: an AI may draft a change and argue for it, never land it A claim on a trunk: generated reasons are a lie Declared hours, or anything else in a member's file A philosophy rule: alignment an AI invented is alignment with nothing
Transcription is not composition. A rule a person dictated may be typed into a file by an assistant. A rule the assistant composed may not be landed, however detailed the description it was built from: it is drafted and handed over. And when an AI is asked how far along something is, it answers from artefacts: commits, the status line, what exists in the folder. When the artefacts do not say, the answer is “I do not know”, never an inference, and never a number that reads as a score.