Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Load only the standards relevant to a
.claude/skills/hashgraph-online-standards/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 38% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 41% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-21 | ✗→✓ | ▲ Improved | -26% | 0% |
Load the smallest set of standards justified by the caller's files, language, and risks. Do not preload the entire reference corpus.
common-standards.md plus only the matching language or checklistreferences.
not-checked scope.
This skill provides context and findings. It does not edit, validate, retry, approve, commit, release, deliver, or decide continuation.
When the caller is about to WRITE code (not only review it), surface the matching language rules INLINE in the working context — a behind-the-link reference does not change behavior; an inline imperative does. The Go core:
fmt.Errorf("doing X: %w", err)— never return a bare inner error.
[]struct cases + t.Run percase), asserting exact expected values including the error cases.
For other languages, pull the matching reference below and inline its top rules the same way.
> Measured 2026-08-04, probe standards-go-conventions (gpt-5.6-luna, N=2, > directional): control produced the %w-wrapped + table-driven shape in 1/2 > runs; with these rules inline, 2/2. Inline-imperative beats reference-link — > the graphify probe measured a linked doc instruction obeyed 0/2. Ledger: > evals/skill-probes/LEDGER.md.
When the supplied change rewrites existing files in bulk — formatters, codemods, migration scripts, generators pointed at hand-written sources — check it against three standards and report each as a finding when absent:
command or script whose inputs, outputs, and dry-run mode can be inspected. Edits scattered across ad-hoc one-liners and manual touch-ups are the diffuse mutation failure mode: no single point can be audited, re-run, or blamed. Finding: name every mutation path outside the chokepoint.
originals are preserved with content hashes recorded (a committed baseline counts), so "the rewrite changed only what it claims" is checkable byte-for-byte, not asserted. Finding: a bulk rewrite with no verifiable before-state.
does not touch, and the diff respects it. A formatter run that also renames, a codemod that also refactors, is the scope-creep rewrite failure mode. Finding: any file class in the diff outside the change's own stated scope.
Stop condition for this check: all three standards have an explicit pass or finding; a bulk-rewrite review that reports style nits but skips these is incomplete.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | pass→pass | 4,073 | 12,624 | +210% | 1 | 1 | 0% | 575 | 2,981 | +418% | 0 | 0 | — |
case-13 | fail→fail | 12,601 | 19,082 | +51% | 1 | 1 | 0% | 1,130 | 2,943 | +160% | 0 | 0 | — |
case-01 | pass→pass | 32,081 | 29,939 | -7% | 1 | 1 | 0% | 6,807 | 6,530 | -4% | 0 | 0 | — |
case-02 | fail→pass | 24,453 | 17,401 | -29% | 1 | 1 | 0% | 2,676 | 3,686 | +38% | 0 | 0 | — |
case-03 | fail→pass | 7,675 | 8,988 | +17% | 1 | 1 | 0% | 1,110 | 1,570 | +41% | 0 | 0 | — |
case-04 | pass→pass | 4,215 | 15,313 | +263% | 1 | 1 | 0% | 657 | 2,563 | +290% | 0 | 0 | — |
case-05 | pass→pass | 11,367 | 2,948 | -74% | 1 | 1 | 0% | 1,008 | 1,387 | +38% | 0 | 0 | — |
case-06 | fail→pass | 23,960 | 29,312 | +22% | 1 | 1 | 0% | 3,922 | 5,131 | +31% | 0 | 0 | — |
case-07 | fail→fail | 22,043 | 15,443 | -30% | 1 | 1 | 0% | 3,310 | 4,350 | +31% | 0 | 0 | — |
case-08 | pass→fail | 35,684 | 38,030 | +7% | 1 | 1 | 0% | 6,831 | 9,165 | +34% | 0 | 0 | — |
case-09 | pass→pass | 23,634 | 15,106 | -36% | 1 | 1 | 0% | 2,765 | 2,638 | -5% | 0 | 0 | — |
case-10 | pass→pass | 21,052 | 10,449 | -50% | 1 | 1 | 0% | 2,807 | 2,516 | -10% | 0 | 0 | — |
case-11 | fail→pass | 18,862 | 15,218 | -19% | 1 | 1 | 0% | 2,787 | 2,483 | -11% | 0 | 0 | — |
case-14 | pass→pass | 17,855 | 15,009 | -16% | 1 | 1 | 0% | 2,019 | 2,440 | +21% | 0 | 0 | — |
case-15 | pass→pass | 23,892 | 13,462 | -44% | 1 | 1 | 0% | 3,401 | 3,629 | +7% | 0 | 0 | — |
case-16 | pass→pass | 25,771 | 9,256 | -64% | 1 | 1 | 0% | 3,869 | 2,276 | -41% | 0 | 0 | — |
case-17 | pass→fail | 23,684 | 20,257 | -14% | 1 | 1 | 0% | 3,371 | 3,832 | +14% | 0 | 0 | — |
case-18 | fail→fail | 19,529 | 21,083 | +8% | 1 | 1 | 0% | 3,288 | 3,854 | +17% | 0 | 0 | — |
case-19 | pass→pass | 14,584 | 16,255 | +11% | 1 | 1 | 0% | 2,544 | 2,763 | +9% | 0 | 0 | — |
case-20 | pass→pass | 18,370 | 3,950 | -78% | 1 | 1 | 0% | 1,900 | 1,596 | -16% | 0 | 0 | — |
case-21 | fail→pass | 32,851 | 16,579 | -50% | 1 | 1 | 0% | 3,887 | 2,892 | -26% | 0 | 0 | — |
case-22 | pass→pass | 28,138 | 16,802 | -40% | 1 | 1 | 0% | 3,798 | 4,407 | +16% | 0 | 0 | — |
case-23 | pass→pass | 15,907 | 3,868 | -76% | 1 | 1 | 0% | 1,827 | 1,565 | -14% | 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. 23 cases were attempted. The headline lift of +13 percentage points is the difference between those two pass rates over the 23 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.