Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Canonical long-task pack for daemon-managed work with deliberate approval checkpoints, status summaries, rollback notes, and mobile-safe governance-aware updates.
.claude/skills/mkurman-approval-checkpoint-long-task/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-23 | ✗→✓ | ▲ Improved | -33% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -4% | 0% |
I want long-running work to pause at explicit checkpoints with clear status, next step, and rollback notes, so I can approve or deny risky transitions from in-app or mobile chat without losing context.
task_kind: task or goalcheckpoint_titles: ordered checkpoint labelsrollback_notes: per-checkpoint rollback instructionssummary_cadence: when to emit status summariesThis pack is usually manual or task/goal-backed rather than purely cron-driven; if scheduled, it should only materialize the governed task template, not auto-bypass approvals.
Use the Approval-Checkpoint Long Task pack for this migration. Define checkpoints for backup, schema update, validation, and cutover, with rollback notes and mobile-safe approval summaries.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-23 | fail→pass | 12,329 | 5,407 | -56% | 1 | 1 | 0% | 1,887 | 1,257 | -33% | 0 | 0 | — |
case-06 | fail→pass | 10,275 | 7,572 | -26% | 1 | 1 | 0% | 1,521 | 1,556 | +2% | 0 | 0 | — |
case-01 | fail→pass | 20,042 | 14,455 | -28% | 1 | 1 | 0% | 3,680 | 2,983 | -19% | 0 | 0 | — |
case-02 | fail→pass | 22,746 | 12,675 | -44% | 1 | 1 | 0% | 3,533 | 2,590 | -27% | 0 | 0 | — |
case-03 | pass→pass | 13,246 | 7,016 | -47% | 1 | 1 | 0% | 1,967 | 1,518 | -23% | 0 | 0 | — |
case-04 | pass→pass | 13,318 | 5,827 | -56% | 1 | 1 | 0% | 2,085 | 1,368 | -34% | 0 | 0 | — |
case-07 | fail→pass | 21,551 | 19,561 | -9% | 1 | 1 | 0% | 4,029 | 3,860 | -4% | 0 | 0 | — |
case-05 | fail→pass | 11,625 | 4,824 | -59% | 1 | 1 | 0% | 1,714 | 1,148 | -33% | 0 | 0 | — |
case-08 | pass→pass | 11,785 | 7,079 | -40% | 1 | 1 | 0% | 1,684 | 1,520 | -10% | 0 | 0 | — |
case-09 | pass→pass | 9,929 | 5,017 | -49% | 1 | 1 | 0% | 1,443 | 1,135 | -21% | 0 | 0 | — |
case-10 | fail→pass | 48,414 | 11,286 | -77% | 1 | 1 | 0% | 3,342 | 2,366 | -29% | 0 | 0 | — |
case-11 | fail→pass | 11,492 | 2,958 | -74% | 1 | 1 | 0% | 1,853 | 901 | -51% | 0 | 0 | — |
case-12 | pass→pass | 12,648 | 4,827 | -62% | 1 | 1 | 0% | 1,973 | 1,150 | -42% | 0 | 0 | — |
case-13 | fail→fail | 11,628 | 1,508 | -87% | 1 | 1 | 0% | 1,772 | 663 | -63% | 0 | 0 | — |
case-14 | fail→fail | 7,432 | 3,680 | -50% | 1 | 1 | 0% | 1,068 | 1,053 | -1% | 0 | 0 | — |
case-15 | pass→pass | 16,714 | 3,587 | -79% | 1 | 1 | 0% | 2,311 | 984 | -57% | 0 | 0 | — |
case-16 | fail→pass | 9,419 | 5,203 | -45% | 1 | 1 | 0% | 1,271 | 1,327 | +4% | 0 | 0 | — |
case-22 | pass→pass | 20,363 | 22,236 | +9% | 1 | 1 | 0% | 3,377 | 4,284 | +27% | 0 | 0 | — |
case-17 | fail→pass | 15,572 | 9,222 | -41% | 1 | 1 | 0% | 2,455 | 2,071 | -16% | 0 | 0 | — |
case-18 | fail→fail | 7,939 | 4,189 | -47% | 1 | 1 | 0% | 1,214 | 1,091 | -10% | 0 | 0 | — |
case-19 | pass→pass | 12,735 | 6,690 | -47% | 1 | 1 | 0% | 2,086 | 1,478 | -29% | 0 | 0 | — |
case-20 | pass→pass | 17,164 | 12,202 | -29% | 1 | 1 | 0% | 3,328 | 2,859 | -14% | 0 | 0 | — |
case-21 | pass→pass | 7,074 | 7,507 | +6% | 1 | 1 | 0% | 1,362 | 1,803 | +32% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +43 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.