Dev blog

process

Posts tagged process.

  1. Rules so far

    Several of the rules Claude reads at the start of a session came after something went wrong. The newest one is about GitHub minutes, and 1,811 of this month's 2,000 are already gone.

  2. Trimming the loop

    Fresh agents grade the work against criteria written first. Today I cut that loop from three verifiers to two, after one package ran three rounds and still failed.

  3. Chasing 8.5

    Two days of agent reviews took the average score from 6.36 to a best of 7.98. It still hasn't reached the 8.5 I asked for.