Optimise and ship the questions drafter

A Python harness around claude -p that takes the enquiries drafter from a legacy prompt to a versioned Haiku configuration. Synthetic packs with frozen splits, property graders, an example ablation, three logged iterations, a reflective optimiser costed by role, skill-trigger and tool-selection evals, a simulated online test with an SRM check, a tiered regression gate between two configurations, and the release note.

Back to the project

The harness has one rule that everything else hangs on: every model call goes through one function (harness/runner.py), which runs claude -p on your subscription and logs the prompt hash, model, effort, role, tokens, cost and output. That single log is what makes the milestones comparable: an iteration, an optimiser round and a regression run are all rows you can group by prompt hash or role. The second rule is about honesty: the test split refuses to run unless you type run test, no prompt version runs until it has a changelog entry, and the optimiser and any proposing session only ever see dev.

Back to the project

Optimise and ship the questions drafter

Optimise and ship the questions drafter

Sign in to keep reading the answer.

Sign in to keep reading the answer.