← ClaudeAtlas

ops-looplisted

Run the measure-intervene-note-diff weekly loop for tuning a Claude Code rig's cost/effectiveness. Use when the user asks "is my rig getting better", "did that config change help", "weekly review", or wants to know whether some change actually moved the numbers.
xantorres/rigops · ★ 0 · AI & Automation · score 72
Install: claude install-skill xantorres/rigops
# Ops loop The rigops ledger only tells you something useful if it's read in a fixed loop, not glanced at ad hoc. Four steps, one week apart, in this order: ## 1. Measure `rigops ledger write` (or `/rigops:week`, which runs the whole loop) appends this week's row: EIT/turn, context percentiles, cache-hit %, output/turn, Agent-delegation rate, cheap-model EIT share, fixed context tax. This is the baseline the next diff compares against - write it even on a week with no planned changes, or the loop has gaps. ## 2. Intervene Make **one** deliberate change: a CLAUDE.md trim, a new hook, a subagent delegation habit, a model-routing rule. One, not three. Multiple simultaneous interventions in the same week make the next diff unreadable - if EIT/turn drops 15%, there's no way to say which of the three changes did it, and no way to know if one of them actually made things worse while another compensated. ## 3. Note The moment an intervention lands (not at week's end from memory), run `rigops ledger note "<what changed>"`. Interventions are timestamped and dated; `ledger diff` pulls in everything noted between two ledger rows. A note written after the fact, backdated by guess, is worse than no note. ## 4. Diff Next week, after step 1's fresh `ledger write`, run `rigops ledger diff` (or `rigops ledger diff --since <date>` for a wider window than the last two rows). It prints the delta on every numeric column plus the interventions noted in that window side by side. Read this *