Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Record a place where the code forced a deviation from the agreed plan. Use mid-implementation whenever reality contradicts the plan, or when the user says "log this deviation". The log feeds future planning and the merge quiz.
.claude/skills/hiendinhngoc-log-deviation/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 122% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 33% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -22% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 164% | 0% |
The plan said one thing; the code demanded another. Record it while it's fresh — these entries are the highest-value input to the next plan.
docs/deviations/YYYY-MM-DD-<task-slug>.md at the repo root(create docs/deviations/ if needed). <task-slug> = short kebab-case name of the current task. Slugs are minted freely, so two sessions will name the same task differently: before creating a new file, list what already exists in docs/deviations/ and reuse the slug of any file that is plausibly this task — appending to it even if it's from an earlier day/session — rather than minting a near-duplicate.
the path.
markdown## <YYYY-MM-DD HH:MM> — <one-line summary> - **Plan said:** <what the plan/spec assumed> - **Code forced:** <what was actually done instead> - **Why:** <the real-world constraint that made the plan wrong> - **Ripple:** <other plan steps this invalidates, or "none">
The heading carries the full date: the file spans days by design, and HH:MM alone goes ambiguous on the second session.
Not every deviation is caught live. When work reveals that plans or docs already disagree with the shipped code (a stale roadmap item, a doc promising a design the code abandoned), log it with the same fields plus one more:
markdown- **Found while:** <the current activity that surfaced it>
Date the heading with the discovery time; if the deviation's original date is known, name it in Code forced.
never just "it didn't work".
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,496 | 4,159 | -51% | 1 | 1 | 0% | 1,668 | 768 | -54% | 0 | 0 | — |
case-02 | fail→pass | 6,231 | 7,434 | +19% | 1 | 1 | 0% | 927 | 2,054 | +122% | 0 | 0 | — |
case-03 | fail→pass | 7,290 | 6,150 | -16% | 1 | 1 | 0% | 1,340 | 1,779 | +33% | 0 | 0 | — |
case-04 | pass→pass | 6,797 | 2,618 | -61% | 1 | 1 | 0% | 1,374 | 998 | -27% | 0 | 0 | — |
case-09 | pass→fail | 16,756 | 6,799 | -59% | 1 | 1 | 0% | 2,739 | 1,036 | -62% | 0 | 0 | — |
case-05 | fail→pass | 8,920 | 3,958 | -56% | 1 | 1 | 0% | 1,782 | 1,420 | -20% | 0 | 0 | — |
case-06 | fail→fail | 7,206 | 3,628 | -50% | 1 | 1 | 0% | 1,325 | 1,171 | -12% | 0 | 0 | — |
case-07 | fail→pass | 9,836 | 3,969 | -60% | 1 | 1 | 0% | 1,765 | 1,373 | -22% | 0 | 0 | — |
case-08 | fail→fail | 8,463 | 4,648 | -45% | 1 | 1 | 0% | 1,671 | 1,253 | -25% | 0 | 0 | — |
case-10 | fail→fail | 5,889 | 4,437 | -25% | 1 | 1 | 0% | 1,176 | 878 | -25% | 0 | 0 | — |
case-11 | pass→fail | 3,851 | 4,132 | +7% | 1 | 1 | 0% | 784 | 850 | +8% | 0 | 0 | — |
case-12 | fail→fail | 2,442 | 4,704 | +93% | 1 | 1 | 0% | 406 | 791 | +95% | 0 | 0 | — |
case-13 | fail→fail | 4,619 | 5,702 | +23% | 1 | 1 | 0% | 949 | 1,714 | +81% | 0 | 0 | — |
case-14 | fail→fail | 5,190 | 4,693 | -10% | 1 | 1 | 0% | 862 | 1,167 | +35% | 0 | 0 | — |
case-15 | pass→pass | 6,106 | 3,482 | -43% | 1 | 1 | 0% | 1,164 | 1,210 | +4% | 0 | 0 | — |
case-16 | pass→pass | 4,498 | 2,878 | -36% | 1 | 1 | 0% | 917 | 1,024 | +12% | 0 | 0 | — |
case-17 | pass→pass | 9,103 | 4,476 | -51% | 1 | 1 | 0% | 1,727 | 1,432 | -17% | 0 | 0 | — |
case-18 | fail→pass | 4,745 | 6,643 | +40% | 1 | 1 | 0% | 692 | 1,830 | +164% | 0 | 0 | — |
case-19 | pass→pass | 9,739 | 11,694 | +20% | 1 | 1 | 0% | 1,817 | 1,541 | -15% | 0 | 0 | — |
case-20 | pass→pass | 13,718 | 14,348 | +5% | 1 | 1 | 0% | 2,382 | 2,932 | +23% | 0 | 0 | — |
case-21 | fail→fail | 3,595 | 4,206 | +17% | 1 | 1 | 0% | 731 | 1,309 | +79% | 0 | 0 | — |
case-22 | pass→pass | 4,516 | 9,309 | +106% | 1 | 1 | 0% | 892 | 2,279 | +155% | 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 17 counted toward the lift figure. The other 5 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 +14 percentage points is the difference between those two pass rates over the 17 comparable cases. 3 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.