Captures lessons from the skeleton eval: workflow boundary scoping,
human turn filtering, token deduplication, context normalization,
and five common pitfalls to avoid.
Strip task sharding, rigid approval menu, and VC backfill. Add
intent_gap-first classification cascade to review. Use slug-based
spec naming in plan step.
Second QD2 skeleton test run (add plan-review step to task-01) with
raw JSONL log and analysis. Add CAPTURE-RUN.md as a repeatable prompt
for post-run log capture and summarization.
18 tasks: skeleton test/eval, then per-step tighten/test/eval
cycles, plus end-to-end eval. Each task file is a self-contained
intent expression that can be fed to QD2.
Alex Verkhovsky