Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Review and record one pending human checkpoint in a research Workspace; use when a HUMAN Unit is blocked on an `Approve C*` decision, and never treat silence or artifact existence as approval.
.claude/skills/willoscar-human-checkpoint/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -13% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -10% | 0% |
A checkpoint is consent, not a formatting step. It binds a named human Decision to the exact Artifacts and constraints reviewed before execution may continue.
DECISIONS.md.UNITS.csv and STATUS.md for the active checkpoint.DECISIONS.md.Inspect STATUS.md, UNITS.csv, and the active runner message. Select the first blocked HUMAN Unit and its C* identifier. If multiple checkpoints appear active or the checklist is missing, stop and repair the projection before approving anything.
Completion criterion: exactly one pending checkpoint and its owning HUMAN Unit are identified.
Read the locked Pipeline's checkpoint contract and inspect every named Artifact. Record requested constraints or scope changes in the checkpoint block before approval; do not silently modify reader-facing content as part of sign-off.
Completion criterion: the reviewer can name the Artifacts inspected and any constraints attached to the Decision.
Use the Pipeline adapter so the Markdown checkbox and machine Decision ledger remain synchronized:
bashuv run python scripts/pipeline.py approve \ --workspace workspaces/<name> \ --checkpoint <C*>
Do not infer approval from chat silence, a completed Artifact, or an existing but unchecked checklist item.
Completion criterion: DECISIONS.md contains [x] Approve C* and the Run ledger records checkpoint.approved for the same checkpoint.
Resume through the Pipeline adapter. The Harness may complete the HUMAN Unit and expose the next eligible Unit; this Skill does not execute downstream semantic work itself.
Completion criterion: the checkpoint is no longer the active blocker, or one new specific blocker is visible in durable Workspace state.
pipelines/*.pipeline.md owns checkpoint purpose and requiredreview Artifacts.
DECISIONS.md is the human-readable Decision surface..harness/decisions.jsonl is the machine-readable history; update itthrough the adapter rather than by hand.
checkpoint-brief to recreate a post-route checkpoint review block.scripts/run.py is a runner-compatibility helper that only toggles theMarkdown checkbox. Prefer scripts/pipeline.py approve, which also records the machine Decision.
with checkpoint-brief before approval. Use pipeline-router only for the initial C0 route.
stale approval and request a new Decision.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-04 | fail→pass | 10,552 | 4,472 | -58% | 1 | 1 | 0% | 1,641 | 1,422 | -13% | 0 | 0 | — |
case-01 | fail→fail | 10,302 | 7,899 | -23% | 1 | 1 | 0% | 1,589 | 1,285 | -19% | 0 | 0 | — |
case-02 | fail→fail | 8,043 | 5,788 | -28% | 1 | 1 | 0% | 696 | 882 | +27% | 0 | 0 | — |
case-03 | fail→fail | 6,535 | 4,905 | -25% | 1 | 1 | 0% | 208 | 949 | +356% | 0 | 0 | — |
case-05 | fail→pass | 7,223 | 2,907 | -60% | 1 | 1 | 0% | 1,114 | 1,140 | +2% | 0 | 0 | — |
case-06 | pass→pass | 7,180 | 4,560 | -36% | 1 | 1 | 0% | 1,080 | 1,369 | +27% | 0 | 0 | — |
case-07 | fail→pass | 9,460 | 3,396 | -64% | 1 | 1 | 0% | 1,403 | 1,246 | -11% | 0 | 0 | — |
case-08 | fail→pass | 7,887 | 2,779 | -65% | 1 | 1 | 0% | 1,161 | 1,108 | -5% | 0 | 0 | — |
case-09 | fail→pass | 9,248 | 3,984 | -57% | 1 | 1 | 0% | 1,366 | 1,235 | -10% | 0 | 0 | — |
case-10 | pass→pass | 10,538 | 5,034 | -52% | 1 | 1 | 0% | 1,530 | 1,495 | -2% | 0 | 0 | — |
case-11 | fail→pass | 12,128 | 3,351 | -72% | 1 | 1 | 0% | 1,851 | 1,223 | -34% | 0 | 0 | — |
case-12 | fail→pass | 7,634 | 2,951 | -61% | 1 | 1 | 0% | 1,244 | 1,134 | -9% | 0 | 0 | — |
case-13 | fail→pass | 16,430 | 2,981 | -82% | 1 | 1 | 0% | 2,529 | 1,102 | -56% | 0 | 0 | — |
case-14 | fail→pass | 5,238 | 3,720 | -29% | 1 | 1 | 0% | 727 | 1,293 | +78% | 0 | 0 | — |
case-15 | fail→pass | 11,131 | 5,643 | -49% | 1 | 1 | 0% | 1,554 | 1,671 | +8% | 0 | 0 | — |
case-16 | fail→fail | 15,176 | 3,808 | -75% | 1 | 1 | 0% | 2,241 | 1,355 | -40% | 0 | 0 | — |
case-17 | fail→pass | 9,682 | 2,675 | -72% | 1 | 1 | 0% | 1,468 | 1,067 | -27% | 0 | 0 | — |
case-18 | pass→pass | 4,420 | 2,242 | -49% | 1 | 1 | 0% | 639 | 1,005 | +57% | 0 | 0 | — |
case-19 | fail→pass | 27,182 | 2,256 | -92% | 1 | 1 | 0% | 1,177 | 935 | -21% | 0 | 0 | — |
case-20 | fail→pass | 12,727 | 5,894 | -54% | 1 | 1 | 0% | 2,101 | 1,661 | -21% | 0 | 0 | — |
case-21 | fail→fail | 19,294 | 37,408 | +94% | 1 | 1 | 0% | 2,555 | 2,860 | +12% | 0 | 0 | — |
case-22 | fail→pass | 9,362 | 8,546 | -9% | 1 | 1 | 0% | 1,467 | 1,998 | +36% | 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. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +64 percentage points is the difference between those two pass rates over the 18 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.