Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Comprehension gate before porting, copying, or adapting an existing implementation. Proves understanding of the reference code before any new code is written. Use when the task is "port X", "copy how Y does it", "adapt Z", or reimplementing existing behavior.
.claude/skills/hiendinhngoc-verify-ref/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -52% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 159% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -2% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -40% | 0% |
Prove you understand the reference implementation BEFORE porting it. Misunderstood references produce confidently wrong ports.
<HARD-GATE>Write no implementation code until the user confirms the comprehension proof below.</HARD-GATE>
ambiguous, ask which implementation is the reference — don't guess.
its direct callers, and direct dependencies that affect observable behavior. Follow deeper calls only while they change outputs, side effects, errors, or ordering. List unresolved or external dependencies instead of recursively expanding without a stop condition.
and what each protects against
idempotency, transactional boundaries)
assumptions, side effects the signature doesn't show. For UI code, also name the framework semantics the reference leans on: lifecycle hooks that fire once vs. per-render, view identity, and task/cleanup-on-teardown behavior — these are where ports break silently.
Proportionality: for a small same-repo adaptation (roughly ≤30 lines, no concurrency, no money/auth/persistence), a short proof of just Data flow and What would break if I got this wrong? suffices — but step 2's full read of the reference and its callers is never skipped; that read is where latent bugs surface.
name the 2-3 misreadings most likely to cause a subtly wrong port and say why your reading is correct (cite file:line evidence).
confirmation may implementation begin. A plan the user already approved that names this exact port also counts as confirmation — say the proof is being accepted under that approval and proceed. Otherwise, if no interactive user is available, stop after the proof and identify the exact uncertainty that needs review; do not silently treat the proof as approval.
## Comprehension proof: <reference name> (<file:line range>)
### Data flow
### Edge cases handled
### Invariants
### Hidden dependencies
### What would break if I got this wrong?| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 2,185 | 3,742 | +71% | 1 | 1 | 0% | 442 | 1,369 | +210% | 0 | 0 | — |
case-02 | fail→pass | 10,127 | 2,417 | -76% | 1 | 1 | 0% | 2,066 | 983 | -52% | 0 | 0 | — |
case-03 | fail→pass | 6,751 | 8,415 | +25% | 1 | 1 | 0% | 867 | 2,242 | +159% | 0 | 0 | — |
case-04 | fail→pass | 13,570 | 18,075 | +33% | 1 | 1 | 0% | 3,226 | 4,045 | +25% | 0 | 0 | — |
case-05 | fail→pass | 15,099 | 12,577 | -17% | 1 | 1 | 0% | 2,922 | 2,851 | -2% | 0 | 0 | — |
case-06 | fail→fail | 2,726 | 4,829 | +77% | 1 | 1 | 0% | 468 | 1,641 | +251% | 0 | 0 | — |
case-07 | fail→pass | 18,618 | 11,346 | -39% | 1 | 1 | 0% | 4,262 | 2,555 | -40% | 0 | 0 | — |
case-08 | fail→fail | 4,739 | 2,937 | -38% | 1 | 1 | 0% | 838 | 1,182 | +41% | 0 | 0 | — |
case-09 | pass→fail | 20,115 | 4,876 | -76% | 1 | 1 | 0% | 4,566 | 1,537 | -66% | 0 | 0 | — |
case-10 | fail→fail | 2,105 | 4,604 | +119% | 1 | 1 | 0% | 423 | 1,475 | +249% | 0 | 0 | — |
case-21 | pass→fail | 9,628 | 7,561 | -21% | 1 | 1 | 0% | 1,386 | 1,726 | +25% | 0 | 0 | — |
case-11 | fail→fail | 3,971 | 6,172 | +55% | 1 | 1 | 0% | 274 | 1,900 | +593% | 0 | 0 | — |
case-12 | fail→fail | 9,234 | 6,135 | -34% | 1 | 1 | 0% | 1,715 | 1,790 | +4% | 0 | 0 | — |
case-13 | fail→pass | 9,314 | 13,217 | +42% | 1 | 1 | 0% | 2,224 | 3,360 | +51% | 0 | 0 | — |
case-14 | fail→fail | 6,132 | 4,327 | -29% | 1 | 1 | 0% | 1,213 | 1,421 | +17% | 0 | 0 | — |
case-15 | fail→fail | 11,176 | 4,144 | -63% | 1 | 1 | 0% | 2,259 | 1,365 | -40% | 0 | 0 | — |
case-16 | pass→fail | 15,830 | 3,650 | -77% | 1 | 1 | 0% | 2,911 | 1,315 | -55% | 0 | 0 | — |
case-17 | fail→fail | 2,019 | 4,383 | +117% | 1 | 1 | 0% | 349 | 1,446 | +314% | 0 | 0 | — |
case-18 | pass→fail | 18,451 | 5,049 | -73% | 1 | 1 | 0% | 3,705 | 1,423 | -62% | 0 | 0 | — |
case-19 | fail→fail | 3,954 | 4,027 | +2% | 1 | 1 | 0% | 696 | 1,384 | +99% | 0 | 0 | — |
case-20 | pass→pass | 8,537 | 9,599 | +12% | 1 | 1 | 0% | 2,062 | 2,971 | +44% | 0 | 0 | — |
case-22 | fail→fail | 8,880 | 3,685 | -59% | 1 | 1 | 0% | 1,804 | 795 | -56% | 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 20 counted toward the lift figure. The other 2 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 +9 percentage points is the difference between those two pass rates over the 20 comparable cases. 4 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.