Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Lifecycle orchestrator for the unknowns plugin. Detects the current phase of work (pre-implementation, mid-implementation, pre-merge) and routes to the right technique skill. Use when the user says "/unknowns", "know my unknowns", or is unsure which unknowns skill applies.
.claude/skills/hiendinhngoc-unknowns/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 77% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -13% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -29% | 0% |
Pure router. Detect the phase, invoke the right skill via the Skill tool. If this agent has no skill-invocation tool, read that skill's SKILL.md and follow it directly. No technique logic lives here.
Trust explicit conversation intent first: requests to review/merge, reports of a plan deviation, porting language, or prototype language route directly. Use git only when intent is not clear.
When git evidence is needed, resolve the comparison base in this order: the current branch's configured upstream; refs/remotes/origin/HEAD; an existing local main; then master. Verify every candidate before using it.
Then check, in order:
Incidental dirt is not evidence: modifications limited to generated or IDE-managed files (*.pbxproj, lockfiles, .DS_Store, build outputs) do not indicate a task in flight — treat such a worktree as clean.
ahead of the resolved base and the worktree has no tracked modifications → invoke unknowns:merge-quiz.
the task at hand, or the conversation shows an agreed plan being executed → if a deviation was just discussed, invoke unknowns:log-deviation; otherwise report that no concrete deviation is available to log and present the three pre-implementation techniques without invoking one speculatively.
question: what kind of unknown are they facing?
unknowns:blindspotunknowns:verify-refunknowns:mockno intent or target can be inferred.
unknowns: namespace (Claude Code plugin install).If the skills were installed flat (Codex, OpenCode: cp -r skills/*), they go by their bare directory names — blindspot, verify-ref, mock, log-deviation, merge-quiz.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,212 | 3,531 | -57% | 1 | 1 | 0% | 1,604 | 1,204 | -25% | 0 | 0 | — |
case-02 | fail→fail | 8,694 | 9,736 | +12% | 1 | 1 | 0% | 1,603 | 1,856 | +16% | 0 | 0 | — |
case-03 | fail→fail | 12,933 | 4,499 | -65% | 1 | 1 | 0% | 2,046 | 1,369 | -33% | 0 | 0 | — |
case-04 | fail→fail | 8,752 | 4,460 | -49% | 1 | 1 | 0% | 1,543 | 971 | -37% | 0 | 0 | — |
case-05 | fail→pass | 15,771 | 5,843 | -63% | 1 | 1 | 0% | 2,636 | 1,623 | -38% | 0 | 0 | — |
case-06 | pass→pass | 9,812 | 2,935 | -70% | 1 | 1 | 0% | 1,803 | 1,052 | -42% | 0 | 0 | — |
case-07 | fail→pass | 3,665 | 3,213 | -12% | 1 | 1 | 0% | 655 | 1,158 | +77% | 0 | 0 | — |
case-08 | fail→pass | 7,307 | 3,042 | -58% | 1 | 1 | 0% | 1,382 | 1,199 | -13% | 0 | 0 | — |
case-09 | fail→pass | 7,441 | 2,447 | -67% | 1 | 1 | 0% | 1,662 | 1,009 | -39% | 0 | 0 | — |
case-10 | fail→pass | 10,863 | 4,555 | -58% | 1 | 1 | 0% | 2,017 | 1,441 | -29% | 0 | 0 | — |
case-11 | fail→pass | 3,617 | 2,170 | -40% | 1 | 1 | 0% | 650 | 932 | +43% | 0 | 0 | — |
case-12 | fail→pass | 7,120 | 1,821 | -74% | 1 | 1 | 0% | 1,439 | 932 | -35% | 0 | 0 | — |
case-13 | pass→pass | 4,561 | 6,878 | +51% | 1 | 1 | 0% | 883 | 2,029 | +130% | 0 | 0 | — |
case-14 | pass→pass | 13,926 | 1,687 | -88% | 1 | 1 | 0% | 2,728 | 861 | -68% | 0 | 0 | — |
case-15 | fail→pass | 10,072 | 3,982 | -60% | 1 | 1 | 0% | 1,985 | 1,380 | -30% | 0 | 0 | — |
case-16 | fail→fail | 8,756 | 5,975 | -32% | 1 | 1 | 0% | 1,699 | 1,710 | +1% | 0 | 0 | — |
case-17 | fail→pass | 5,914 | 2,147 | -64% | 1 | 1 | 0% | 1,177 | 960 | -18% | 0 | 0 | — |
case-18 | fail→pass | 5,964 | 4,807 | -19% | 1 | 1 | 0% | 1,028 | 1,486 | +45% | 0 | 0 | — |
case-19 | fail→pass | 10,122 | 8,010 | -21% | 1 | 1 | 0% | 1,764 | 1,326 | -25% | 0 | 0 | — |
case-20 | pass→pass | 5,622 | 3,224 | -43% | 1 | 1 | 0% | 1,225 | 1,211 | -1% | 0 | 0 | — |
case-21 | pass→pass | 2,994 | 18,675 | +524% | 1 | 1 | 0% | 538 | 981 | +82% | 0 | 0 | — |
case-22 | pass→pass | 5,196 | 5,713 | +10% | 1 | 1 | 0% | 977 | 1,684 | +72% | 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. The headline lift of +50 percentage points is the difference between those two pass rates over the 22 comparable cases.
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.