Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Project completed stage Artifacts into a concise human-review block in `DECISIONS.md`; use before a Workflow checkpoint, not for Workflow routing or approval.
.claude/skills/willoscar-checkpoint-brief/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 206% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 118% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 895% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -44% | 0% |
Prepare the review surface between machine-produced Artifacts and a human checkpoint. This Skill summarizes; it does not choose a Workflow, approve a checkpoint, or modify semantic Artifacts.
PIPELINE.lock.mdDECISIONS.mdDECISIONS.mdnew research claims.
prior Decision history.
human-checkpoint owns the approval Decision.DECISIONS.md names the current checkpoint and the Artifacts reviewed.bashuv run python .codex/skills/checkpoint-brief/scripts/run.py \ --workspace workspaces/<name> \ --checkpoint C2 \ --inputs 'outline/taxonomy.yml;outline/outline.yml'
The Pipeline adapter supplies the current Unit's declared inputs. Manual invocations must do the same; an empty input list is reported as unreviewable rather than inferred from unrelated Workspace files.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-03 | fail→pass | 5,232 | 4,292 | -18% | 1 | 1 | 0% | 311 | 951 | +206% | 0 | 0 | — |
case-04 | pass→pass | 3,238 | 3,841 | +19% | 1 | 1 | 0% | 498 | 982 | +97% | 0 | 0 | — |
case-01 | fail→fail | 8,553 | 5,963 | -30% | 1 | 1 | 0% | 1,430 | 640 | -55% | 0 | 0 | — |
case-02 | fail→fail | 6,625 | 5,535 | -16% | 1 | 1 | 0% | 1,096 | 642 | -41% | 0 | 0 | — |
case-05 | fail→fail | 12,219 | 15,536 | +27% | 1 | 1 | 0% | 1,897 | 2,770 | +46% | 0 | 0 | — |
case-06 | fail→fail | 8,671 | 6,301 | -27% | 1 | 1 | 0% | 1,277 | 734 | -43% | 0 | 0 | — |
case-07 | fail→pass | 12,847 | 7,115 | -45% | 1 | 1 | 0% | 1,927 | 1,559 | -19% | 0 | 0 | — |
case-08 | fail→pass | 6,199 | 11,114 | +79% | 1 | 1 | 0% | 1,005 | 2,194 | +118% | 0 | 0 | — |
case-09 | fail→pass | 4,215 | 7,709 | +83% | 1 | 1 | 0% | 151 | 1,502 | +895% | 0 | 0 | — |
case-10 | fail→pass | 8,272 | 2,554 | -69% | 1 | 1 | 0% | 1,416 | 798 | -44% | 0 | 0 | — |
case-11 | fail→pass | 7,058 | 2,613 | -63% | 1 | 1 | 0% | 1,124 | 766 | -32% | 0 | 0 | — |
case-12 | pass→pass | 6,996 | 3,358 | -52% | 1 | 1 | 0% | 1,000 | 910 | -9% | 0 | 0 | — |
case-13 | fail→pass | 10,231 | 5,251 | -49% | 1 | 1 | 0% | 1,624 | 1,158 | -29% | 0 | 0 | — |
case-22 | pass→pass | 6,331 | 2,413 | -62% | 1 | 1 | 0% | 959 | 692 | -28% | 0 | 0 | — |
case-14 | fail→pass | 10,593 | 3,296 | -69% | 1 | 1 | 0% | 1,551 | 879 | -43% | 0 | 0 | — |
case-15 | fail→pass | 11,489 | 2,961 | -74% | 1 | 1 | 0% | 1,724 | 778 | -55% | 0 | 0 | — |
case-16 | fail→pass | 6,767 | 2,291 | -66% | 1 | 1 | 0% | 889 | 623 | -30% | 0 | 0 | — |
case-17 | fail→pass | 10,262 | 4,137 | -60% | 1 | 1 | 0% | 1,508 | 1,016 | -33% | 0 | 0 | — |
case-18 | fail→fail | 9,583 | 3,348 | -65% | 1 | 1 | 0% | 1,253 | 909 | -27% | 0 | 0 | — |
case-19 | pass→pass | 10,335 | 6,930 | -33% | 1 | 1 | 0% | 1,668 | 1,516 | -9% | 0 | 0 | — |
case-20 | pass→pass | 6,016 | 5,561 | -8% | 1 | 1 | 0% | 1,046 | 1,341 | +28% | 0 | 0 | — |
case-21 | pass→pass | 6,647 | 3,386 | -49% | 1 | 1 | 0% | 947 | 870 | -8% | 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 +50 percentage points is the difference between those two pass rates over the 18 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
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.