Discussion about this post

User's avatar
Halfcycle's avatar

"Making intent survive generation" is the right framing, and a green PR hiding architectural drift is the case that justifies the whole harness. For teams that can't build their own constitution and review agents, npx halfcycle packages the same idea inside Claude Code: decisions recorded before the build, and a reviewer that didn't write the plan. Did the drift show up more at boundaries between modules, or inside them?

Halfcycle's avatar

"Nothing learned in prompt v1 carries into prompt v2" has a bigger version: nothing learned in this project carries into the next one. The agent folding three authorities into one aggregate is not a BrewUp problem. It will happen again in the next codebase with a payment step, and that team will write its own domain carrier from scratch after the same green PR. Have Alessandro and Alberto found a way to carry those lessons across repositories, or do the carriers stay per project?

Ready for more?