Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Extract per-subsection “anchor facts” (NO PROSE) from evidence packs so the writer is forced to include concrete numbers/benchmarks/limitations instead of generic summaries. **Trigger**: anchor sheet, anchor facts, numeric anchors, evidence hooks, 写作锚点, 数字锚点, 证据钩子. **Use when**: `outline/evidence_drafts.jsonl` exists and you want stronger, evidence-anchored writing in `sections/*.md`.
.claude/skills/willoscar-anchor-sheet/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 331% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -26% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -36% | 0% |
Purpose: make “what to actually say” explicit:
This prevents the writer from producing paragraph-shaped but content-poor prose.
outline/evidence_drafts.jsonlcitations/ref.biboutline/anchor_sheet.jsonloutline/anchor_sheet.jsonl)JSONL (one object per H3 subsection).
Required fields:
sub_id, titleanchors (list; each anchor has hook_type, text, citations, and optional paper_id/evidence_id/pointer)outline/evidence_drafts.jsonl.citations/ref.bib.outline/anchor_sheet.jsonl.TODO/…/(placeholder)).Anchors are intended to prevent “long but empty” prose. Treat them as must-use hooks, not optional ideas.
Recommended minimums per H3 (A150++):
Note:
uv run python .codex/skills/anchor-sheet/scripts/run.py --helpuv run python .codex/skills/anchor-sheet/scripts/run.py --workspace <workspace>--workspace <dir>--unit-id <U###>--inputs <semicolon-separated>--outputs <semicolon-separated>--checkpoint <C#>uv run python .codex/skills/anchor-sheet/scripts/run.py --workspace <workspace>uv run python .codex/skills/anchor-sheet/scripts/run.py --workspace <workspace> --inputs "outline/evidence_drafts.jsonl;citations/ref.bib" --outputs "outline/anchor_sheet.jsonl"When you are satisfied with anchor facts (and they are actually subsection-specific), create:
outline/anchor_sheet.refined.okThis is an explicit "I reviewed/refined this" signal:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 11,525 | 5,315 | -54% | 1 | 1 | 0% | 2,323 | 1,002 | -57% | 0 | 0 | — |
case-02 | fail→fail | 9,557 | 4,580 | -52% | 1 | 1 | 0% | 309 | 971 | +214% | 0 | 0 | — |
case-03 | fail→fail | 33,932 | 4,575 | -87% | 1 | 1 | 0% | 6,217 | 1,003 | -84% | 0 | 0 | — |
case-04 | fail→fail | 3,550 | 19,124 | +439% | 1 | 1 | 0% | 524 | 4,296 | +720% | 0 | 0 | — |
case-05 | fail→pass | 2,960 | 6,757 | +128% | 1 | 1 | 0% | 445 | 1,920 | +331% | 0 | 0 | — |
case-06 | pass→pass | 17,892 | 14,130 | -21% | 1 | 1 | 0% | 3,862 | 3,629 | -6% | 0 | 0 | — |
case-07 | fail→pass | 6,585 | 3,275 | -50% | 1 | 1 | 0% | 1,062 | 1,302 | +23% | 0 | 0 | — |
case-08 | fail→pass | 8,132 | 2,257 | -72% | 1 | 1 | 0% | 1,302 | 1,155 | -11% | 0 | 0 | — |
case-09 | fail→pass | 12,221 | 4,103 | -66% | 1 | 1 | 0% | 1,887 | 1,392 | -26% | 0 | 0 | — |
case-10 | fail→pass | 8,889 | 1,304 | -85% | 1 | 1 | 0% | 1,462 | 935 | -36% | 0 | 0 | — |
case-11 | pass→pass | 12,889 | 1,953 | -85% | 1 | 1 | 0% | 1,835 | 1,014 | -45% | 0 | 0 | — |
case-12 | pass→pass | 9,454 | 1,927 | -80% | 1 | 1 | 0% | 1,416 | 1,049 | -26% | 0 | 0 | — |
case-13 | fail→pass | 12,081 | 1,707 | -86% | 1 | 1 | 0% | 1,839 | 990 | -46% | 0 | 0 | — |
case-14 | pass→pass | 8,872 | 2,210 | -75% | 1 | 1 | 0% | 1,234 | 1,115 | -10% | 0 | 0 | — |
case-15 | fail→pass | 9,830 | 4,216 | -57% | 1 | 1 | 0% | 1,378 | 1,393 | +1% | 0 | 0 | — |
case-16 | fail→pass | 8,917 | 2,009 | -77% | 1 | 1 | 0% | 1,496 | 1,044 | -30% | 0 | 0 | — |
case-17 | pass→pass | 5,922 | 1,387 | -77% | 1 | 1 | 0% | 825 | 907 | +10% | 0 | 0 | — |
case-18 | pass→pass | 11,223 | 1,639 | -85% | 1 | 1 | 0% | 1,552 | 982 | -37% | 0 | 0 | — |
case-19 | fail→pass | 17,999 | 1,396 | -92% | 1 | 1 | 0% | 1,184 | 930 | -21% | 0 | 0 | — |
case-20 | fail→pass | 13,071 | 1,624 | -88% | 1 | 1 | 0% | 1,897 | 963 | -49% | 0 | 0 | — |
case-21 | pass→pass | 28,595 | 6,428 | -78% | 1 | 1 | 0% | 2,390 | 1,880 | -21% | 0 | 0 | — |
case-22 | fail→pass | 6,632 | 1,538 | -77% | 1 | 1 | 0% | 942 | 971 | +3% | 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.