
ALT Built isn't seen working. Count it honestly: Built but never seen working went on one list. It stood at 220. Walk before you build more: An agent drives a real browser through each feature. Done means the walk shows the step: Merged and tested doesn't count on its own. Model and rules, together: A better model on the old rules makes more unverified work, faster.

ALT The planner never writes code. One builder, one git worktree: No two builders ever write to the same checkout. New idea mid-build? New builder.: The running one is left alone to finish. Bug fixes start with a failing test: If you can't make it fail, you haven't found the bug yet. Pick the model by difficulty: Cheap models for mechanical work. The top model only gets the hardest items, after the spec and plan pass review. Progress lives in files: Any session can die and the next one picks up where it stopped.

ALT Fable orchestrates. Opus builds. It treats "green" as a claim to check: Stopped a batch at 5 of 20 when the output was wrong, rather than report "exported". It reads the build server's run anyway: Found two failures after two clean reviews and pulled the change within the hour. Routine work goes to cheaper models, by rule: And a lower auto-compaction limit stops long sessions dragging stale context. It still makes mistakes: A finished report sat unread for two hours because nothing was waiting on it. My bottlenecks are now on the CI side.

ALT Before any code, fill this in. For: The user, by role. Never a name. Job, in their words: Quote them. Don't translate it into spec-speak. We'll know it worked when: Something you could watch happen. Not doing: Write down what's out of scope too. Must pass: The cases the feature has to handle. Must refuse: Can't fill this one in? You don't understand the work yet. A brief asked for a lock. The builder measured first and found the clash could never happen in real use, so we shipped a test instead.



















