Act II was three weeks of December and January in which I stopped writing prompts and started building structure. The recap, before the story gets a name.
Info
- The bottleneck moved — December’s models made capability cheap; the constraint moved to everything around the model. A forty-eight-thousand-line naive weekend, then three systems in three weeks.
- Logging everything, inside git — effort capture in git notes, a vocabulary for agent work, and the governance rule I’d tattoo somewhere visible: the LLM cannot approve its own plan.
- One schema, every surface — the boring January commit that mattered most: humans and agents consuming the same typed contract. Constraints you enforce beat constraints you beg for.
- The 36-hour failure — prompts-dist died in a weekend and left behind the evals insight (human confidence, not pass rates) and tools I still use daily.
The one-line version of Act II: when the model stopped being the constraint, contracts, records, and evals became the work. Not one line of it was a prompt.