A faster groom without losing quality

Idea§

The owner wants to try different variants of the groom playbook, with one goal: find a groom that runs faster without losing quality (owner, 2026-10-15). At the time a plan of 12-29 SP spent 6-12 minutes in groom on top of 17-43 minutes of develop, and small plans paid the same fixed groom cost as large ones. Variants to compare included skipping groom for small mechanical plans (straight to develop with a short brief), a lighter graph, and depth scaled to the task, which went on as idea/0009.

A measurement (reference/0006) puts the playbook groom at 2.4× the wall time, 5.2× the main-thread API calls and 3.0× the output tokens of the old /groom skill. The measurement does not say which steps the extra cost came from.

How to check§

• Experiment (owner, 2026-10-16; protocol in booping-benchmarks groom/): a corpus of past requests from two other projects (2 small, 2 medium, 2 complex), each in an isolated worktree at the commit before its plan with a vault stripped of plans, retros, notes and reviews; variants V0 the current groom, V1 a bare "write a plan", V2 bare + plan template, V3 template + stages, V4 V3 + cross-review (codex, fable, three OpenRouter models via pi-developer); metrics wall time, tokens and calls from booping session-stats, plan quality by a blind rubric judge against the historical plan, and for the best two a develop run scored by bench-score; each groom prompt paragraph probed against an empty-context agent to cut what the model already knows. Results land in a reference record, the chosen shape in an ADR.
• Added 2026-10-16 (owner): D1 — the whole development as one claude -p run with no groom or develop playbooks, compared on time and outcome with the historical pipeline; every variant runs in two isolation modes — a clean vault, and a vault with the project's plans, retros and reviews dated before the task (realistic: past plans may guide the model); historical reference plans came from older booping versions, so fresh references are made with the current booping at the end and the judge checks against both. Corpus: 6 done tasks, mostly web/client work (booping-benchmarks groom/corpus/).
• Owner's targets: small tasks (10-15 SP) reach develop with little or no grooming; complex ones keep research quality and also get faster; tokens about the same or lower.
• Run each variant on the same set of requests on the same commit, and compare plan quality (the groom evals' rubric, the plan's later develop outcome: rework, failed attempts) and cost (booping session-stats: wall time, tokens, API calls); a variant wins when its cost is clearly lower and quality holds. booping-benchmarks holds the harness for such runs.

Verdict§

2026-10-17: the benchmark (booping-benchmarks groom/, reference/0007) picked one drafting pass on the plan template, research then draft, with no self-review step (adr/0030): about 8 minutes a plan against 13–23 for the old groom, at equal or better judged quality. A script cuts the drafted plan into milestone files (adr/0031), and groom runs no separate review step (adr/0035). The light path for small, already-decided requests and depth scaled to the task are not done yet; they went on as idea/0009.