Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Build a final, machine-readable argument snapshot from current H3 sections and refresh section fingerprints before merge. **Trigger**: argument snapshot, section ledger, consistency contract, final section manifest, 论证快照, 一致性契约. **Use when**: section-level mutations are complete and the merge needs a current argument ledger plus content hashes.
.claude/skills/willoscar-argument-selfloop/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 765% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 375% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 263% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 306% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -44% | 0% |
Create the final C5 snapshot consumed by merge and later audit.
Despite the historical Skill name, the current implementation is not an autonomous rewrite loop. It reads the final H3 files, assigns bounded structural move labels, writes a compact consistency contract, refreshes section hashes, and fails when required prose is absent or too thin.
textsection writing -> style and numeric hygiene -> logic polish -> paragraph compaction -> argument snapshot -> optional transitions -> merge
Running the snapshot after every section mutator prevents the ledger and manifest from describing an earlier draft.
sections/outline/outline.ymlqueries.mdoutput/PARAGRAPH_CURATION_REPORT.mdoutput/ARGUMENT_SELFLOOP_TODO.mdoutput/SECTION_ARGUMENT_SUMMARIES.jsonloutput/ARGUMENT_SKELETON.mdsections/sections_manifest.jsonlSECTION_ARGUMENT_SUMMARIES.jsonl contains one record per expected H3 and one record per paragraph. Current move labels are deterministic signals drawn from:
textsetup, thesis, contrast, evidence, evaluation, limitation, synthesis, takeaway
They make section shape inspectable; they do not prove that an argument is scientifically valid. ARGUMENT_SKELETON.md contains a ## Consistency Contract and a chapter-level map. It is a writer-facing audit artifact, never reader-facing paper content.
ARGUMENT_SKELETON.md contains ## Consistency Contract;sections_manifest.jsonl carries current bytes and sha256 values.On FAIL, inspect ARGUMENT_SELFLOOP_TODO.md and route the named section back to its writer or evidence owner. This Skill does not apply the repair itself.
bashuv run python .codex/skills/argument-selfloop/scripts/run.py \ --workspace workspaces/<name>
Optional runner fields are --unit-id, --inputs, --outputs, and --checkpoint.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-20 | pass→pass | 11,590 | 3,574 | -69% | 1 | 1 | 0% | 1,689 | 1,061 | -37% | 0 | 0 | — |
case-03 | fail→fail | 2,337 | 4,543 | +94% | 1 | 1 | 0% | 228 | 738 | +224% | 0 | 0 | — |
case-01 | fail→fail | 9,385 | 7,304 | -22% | 1 | 1 | 0% | 1,868 | 1,255 | -33% | 0 | 0 | — |
case-02 | fail→fail | 4,587 | 5,292 | +15% | 1 | 1 | 0% | 223 | 858 | +285% | 0 | 0 | — |
case-04 | fail→pass | 5,175 | 12,569 | +143% | 1 | 1 | 0% | 301 | 2,603 | +765% | 0 | 0 | — |
case-05 | fail→pass | 5,602 | 9,083 | +62% | 1 | 1 | 0% | 308 | 1,464 | +375% | 0 | 0 | — |
case-06 | fail→fail | 4,731 | 6,424 | +36% | 1 | 1 | 0% | 227 | 996 | +339% | 0 | 0 | — |
case-07 | fail→pass | 6,832 | 3,928 | -43% | 1 | 1 | 0% | 326 | 1,182 | +263% | 0 | 0 | — |
case-08 | fail→pass | 9,287 | 4,014 | -57% | 1 | 1 | 0% | 181 | 734 | +306% | 0 | 0 | — |
case-09 | fail→fail | 8,594 | 5,639 | -34% | 1 | 1 | 0% | 443 | 780 | +76% | 0 | 0 | — |
case-10 | fail→pass | 10,071 | 2,502 | -75% | 1 | 1 | 0% | 1,718 | 962 | -44% | 0 | 0 | — |
case-11 | fail→pass | 10,835 | 2,132 | -80% | 1 | 1 | 0% | 1,559 | 875 | -44% | 0 | 0 | — |
case-12 | fail→pass | 6,978 | 2,141 | -69% | 1 | 1 | 0% | 1,148 | 895 | -22% | 0 | 0 | — |
case-13 | fail→pass | 13,650 | 2,384 | -83% | 1 | 1 | 0% | 2,018 | 936 | -54% | 0 | 0 | — |
case-19 | pass→pass | 7,171 | 2,375 | -67% | 1 | 1 | 0% | 1,027 | 919 | -11% | 0 | 0 | — |
case-14 | pass→pass | 9,712 | 1,946 | -80% | 1 | 1 | 0% | 1,609 | 813 | -49% | 0 | 0 | — |
case-15 | fail→pass | 9,185 | 1,486 | -84% | 1 | 1 | 0% | 1,439 | 754 | -48% | 0 | 0 | — |
case-16 | fail→pass | 6,776 | 1,751 | -74% | 1 | 1 | 0% | 1,125 | 772 | -31% | 0 | 0 | — |
case-17 | fail→pass | 7,886 | 2,830 | -64% | 1 | 1 | 0% | 1,208 | 969 | -20% | 0 | 0 | — |
case-18 | fail→pass | 8,252 | 2,538 | -69% | 1 | 1 | 0% | 1,212 | 973 | -20% | 0 | 0 | — |
case-21 | fail→pass | 5,568 | 2,772 | -50% | 1 | 1 | 0% | 886 | 1,019 | +15% | 0 | 0 | — |
case-22 | fail→pass | 15,484 | 2,780 | -82% | 1 | 1 | 0% | 1,141 | 923 | -19% | 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 13 counted toward the lift figure. The other 9 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 13 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.