happier-instruction-eval

# Happier Instruction Evaluation Evaluate whether instruction wording changes agent behavior without telling the agents they are being evaluated. This workflow is advisory: it produces evidence for a human decision and never edits the canonical instruction owner by itself. ## 1. Establish the decision Name: - the instruction owner and variants being compared, including the current baseline; - the concrete behavior the change is meant to improve or failure it is meant to prevent; - one organic task or a small risk-selected set of tasks that can expose that difference; - a rubric of observable outcomes fixed before any run; - the user decision the evidence will inform. Do not evaluate prose elegance in isolation. A useful task forces the instruction to affect routing, investigation, ownership, implementation shape, validation, stopping, or reporting. Skip evaluation when a source inspection or deterministic check can decide the question directly. ## 2. Control the comparison Hold constant everything except the instruction variant when practical: - use the same task prompt, repository basis, allowed tools, permissions, time/effort budget, and available evidence; - give each run only the ordinary context an agent would receive for that task; - label variants and output locations neutrally so neither runner nor judge sees “baseline,” “preferred,” model identity, or another run's existence; - isolate writes in separate temporary directories or authorized worktrees; never sw...