Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →SIA (Self-Improvement Agent) harness exemplar — Meta/Target/Feedback loops, public/private splits, fixture replay, fail-closed validation.
.claude/skills/docxology-template-sia/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 6% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -52% | 0% |
Project-scoped skill for the in-repo exemplar at projects/templates/template_sia/. Load this when working inside the project.
template_sia exemplar — running scripts, editing source,or regenerating outputs.
no-mocks testing) still hold after changes.
bash# From the repository root uv run pytest projects/templates/template_sia/tests --cov=projects/templates/template_sia/src --cov-fail-under=90 uv run python scripts/pipeline/stage_02_analysis.py --project templates/template_sia uv run python scripts/pipeline/stage_03_render.py --project templates/template_sia uv run python scripts/pipeline/stage_04_validate.py --project templates/template_sia uv run python scripts/pipeline/stage_05_copy.py --project templates/template_sia
src/ or sharedinfrastructure/, not in scripts/. src/template_sia/loop/loop.py owns the project adapter; scripts/run_sia_loop.py only parses flags and presents result paths.
computation.
output/ — regenerate fromsource and config.
as working directory unless the child AGENTS.md states otherwise.
AGENTS.mdREADME.mdTODO.mdprojects/AGENTS.md| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 4,542 | 3,475 | -23% | 1 | 1 | 0% | 776 | 647 | -17% | 0 | 0 | — |
case-02 | fail→fail | 2,792 | 3,376 | +21% | 1 | 1 | 0% | 312 | 867 | +178% | 0 | 0 | — |
case-03 | fail→fail | 3,296 | 6,022 | +83% | 1 | 1 | 0% | 305 | 799 | +162% | 0 | 0 | — |
case-04 | fail→pass | 12,230 | 3,741 | -69% | 1 | 1 | 0% | 2,152 | 1,089 | -49% | 0 | 0 | — |
case-05 | pass→pass | 9,242 | 5,084 | -45% | 1 | 1 | 0% | 1,699 | 1,330 | -22% | 0 | 0 | — |
case-06 | fail→pass | 7,195 | 6,697 | -7% | 1 | 1 | 0% | 1,218 | 1,293 | +6% | 0 | 0 | — |
case-07 | pass→pass | 8,276 | 2,324 | -72% | 1 | 1 | 0% | 1,555 | 879 | -43% | 0 | 0 | — |
case-08 | fail→pass | 9,644 | 2,201 | -77% | 1 | 1 | 0% | 1,551 | 855 | -45% | 0 | 0 | — |
case-09 | fail→pass | 10,191 | 5,423 | -47% | 1 | 1 | 0% | 1,640 | 783 | -52% | 0 | 0 | — |
case-10 | fail→pass | 9,086 | 1,455 | -84% | 1 | 1 | 0% | 1,503 | 646 | -57% | 0 | 0 | — |
case-11 | fail→pass | 14,915 | 4,459 | -70% | 1 | 1 | 0% | 2,582 | 1,173 | -55% | 0 | 0 | — |
case-12 | fail→pass | 14,011 | 9,280 | -34% | 1 | 1 | 0% | 2,651 | 1,941 | -27% | 0 | 0 | — |
case-13 | fail→pass | 14,103 | 7,496 | -47% | 1 | 1 | 0% | 2,503 | 1,857 | -26% | 0 | 0 | — |
case-14 | fail→pass | 9,685 | 2,131 | -78% | 1 | 1 | 0% | 1,754 | 812 | -54% | 0 | 0 | — |
case-15 | fail→pass | 7,545 | 1,731 | -77% | 1 | 1 | 0% | 1,210 | 749 | -38% | 0 | 0 | — |
case-16 | pass→pass | 25,965 | 4,177 | -84% | 1 | 1 | 0% | 1,184 | 1,176 | -1% | 0 | 0 | — |
case-17 | fail→pass | 10,932 | 5,544 | -49% | 1 | 1 | 0% | 1,845 | 1,303 | -29% | 0 | 0 | — |
case-18 | pass→pass | 6,418 | 2,608 | -59% | 1 | 1 | 0% | 1,122 | 850 | -24% | 0 | 0 | — |
case-19 | fail→pass | 3,056 | 1,513 | -50% | 1 | 1 | 0% | 503 | 674 | +34% | 0 | 0 | — |
case-20 | pass→pass | 8,141 | 4,751 | -42% | 1 | 1 | 0% | 1,318 | 1,309 | -1% | 0 | 0 | — |
case-21 | pass→pass | 11,226 | 8,097 | -28% | 1 | 1 | 0% | 1,921 | 1,987 | +3% | 0 | 0 | — |
case-22 | pass→pass | 9,155 | 5,665 | -38% | 1 | 1 | 0% | 1,538 | 1,362 | -11% | 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 21 counted toward the lift figure. The other 1 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 +59 percentage points is the difference between those two pass rates over the 21 comparable cases.
The publisher has shipped newer versions since this run, so these numbers describe v1, not the version currently listed.
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.