Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Front door for agent automation: choose
.claude/skills/boshu2-automation-shape-routing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 95% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 160% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 76% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 182% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 108% | 0% |
Choose the smallest execution shape that preserves the required evidence and control. This skill routes; it does not build or start a substrate.
Ordered routing works because each rung is strictly cheaper to operate than the next: if the smallest shape truly preserves the evidence and control the task needs, every larger shape can only add coordination cost, never correctness.
Named failure mode — substrate romance: routing to persistent workers because the topology is interesting, not because any deciding axis demands it.
Anti-pattern: starting the chosen substrate as part of routing "to save a step". Corrective: return the one-line verdict and let the owner start under its own authority.
a judgment step, while launching NTM, Agent Mail, or Gas City changes runtime state and requires separate operator authority.
coordination add recovery and ownership costs that one-shot work cannot repay.
topology cannot make overlapping production writes safe.
fanout only when independent perspectives are the product. Do not create a reusable artifact for a one-off task.
skill-builder.operationalize to a gate.workflow-builder onlywhere that runtime is explicitly selected and available.
agent-native. NTM is the pane adapter; Agent Mail coordinates only explicitly selected live actors.
using-gc only when the operator explicitly selects Gas City. GC is not an automatic fallback or an ao runtime enum.
| Axis | Lightweight choice | Escalated choice | |---|---|---| | lifetime | current turn | persistent/attachable worker | | topology | one writer or bounded fanout | durable role graph | | control | no mid-run steering | observe/nudge/replace | | output | one artifact | reusable skill/workflow/gate | | store | caller-owned packet set | operator-selected GC quest store | | contention | one writer | partition, then Agent Mail reservation |
Parallelism buys independence, not guaranteed speed. Refuse persistent orchestration for one-shot work, colliding write scopes, or a sequential chain that has no exploitable concurrency.
Return exactly one of:
inline or bounded-fanoutskill-builderworkflow-builderagent-native with a named reason persistent panes helpusing-gc with explicit operator choiceoperationalize:gateName the deciding axis and owner, then stop. The caller may invoke that owner separately; do not copy or start the delegated workflow from this router.
owner=<owning-skill> using one of the shapes listed under Handoff.
inline and bounded-fanout use current-agent;skill-builder, workflow-builder, agent-native, and using-gc use the same value for owner; operationalize:gate uses operationalize.
$verdict as exactly one linewith the declared shape/owner mapping:
bash printf '%s\n' "$verdict" | awk ' NR > 1 { extra = 1 } { valid = ($0 ~ /^shape=(inline|bounded-fanout); axis=[^;]+; owner=current-agent$/ || $0 ~ /^shape=skill-builder; axis=[^;]+; owner=skill-builder$/ || $0 ~ /^shape=workflow-builder; axis=[^;]+; owner=workflow-builder$/ || $0 ~ /^shape=agent-native; axis=[^;]+; owner=agent-native$/ || $0 ~ /^shape=using-gc; axis=[^;]+; owner=using-gc$/ || $0 ~ /^shape=operationalize:gate; axis=[^;]+; owner=operationalize$/) } END { exit !(NR == 1 && !extra && valid) } '
it; inline and bounded-fanout describe the recommended execution shape but do not authorize this routing skill to begin execution.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→pass | 7,978 | 6,715 | -16% | 1 | 1 | 0% | 1,317 | 2,565 | +95% | 0 | 0 | — |
case-01 | fail→pass | 3,713 | 3,331 | -10% | 1 | 1 | 0% | 702 | 1,825 | +160% | 0 | 0 | — |
case-03 | fail→pass | 7,462 | 4,892 | -34% | 1 | 1 | 0% | 1,280 | 2,252 | +76% | 0 | 0 | — |
case-04 | fail→pass | 4,503 | 3,559 | -21% | 1 | 1 | 0% | 631 | 1,779 | +182% | 0 | 0 | — |
case-05 | fail→pass | 4,565 | 3,423 | -25% | 1 | 1 | 0% | 869 | 1,804 | +108% | 0 | 0 | — |
case-06 | fail→pass | 7,619 | 3,830 | -50% | 1 | 1 | 0% | 1,266 | 1,859 | +47% | 0 | 0 | — |
case-07 | fail→pass | 8,264 | 3,618 | -56% | 1 | 1 | 0% | 1,397 | 1,900 | +36% | 0 | 0 | — |
case-08 | fail→pass | 7,417 | 6,461 | -13% | 1 | 1 | 0% | 1,294 | 1,929 | +49% | 0 | 0 | — |
case-09 | fail→pass | 7,483 | 3,152 | -58% | 1 | 1 | 0% | 1,330 | 1,745 | +31% | 0 | 0 | — |
case-10 | fail→pass | 7,939 | 4,750 | -40% | 1 | 1 | 0% | 1,242 | 2,090 | +68% | 0 | 0 | — |
case-11 | fail→pass | 5,738 | 2,962 | -48% | 1 | 1 | 0% | 939 | 1,775 | +89% | 0 | 0 | — |
case-12 | fail→pass | 8,208 | 4,077 | -50% | 1 | 1 | 0% | 1,405 | 2,057 | +46% | 0 | 0 | — |
case-13 | fail→pass | 5,995 | 4,635 | -23% | 1 | 1 | 0% | 1,073 | 2,052 | +91% | 0 | 0 | — |
case-14 | fail→pass | 12,815 | 3,012 | -76% | 1 | 1 | 0% | 2,242 | 1,738 | -22% | 0 | 0 | — |
case-15 | fail→pass | 6,744 | 4,689 | -30% | 1 | 1 | 0% | 1,162 | 1,924 | +66% | 0 | 0 | — |
case-16 | fail→pass | 9,855 | 2,604 | -74% | 1 | 1 | 0% | 1,768 | 1,662 | -6% | 0 | 0 | — |
case-17 | fail→pass | 5,410 | 3,521 | -35% | 1 | 1 | 0% | 934 | 1,868 | +100% | 0 | 0 | — |
case-18 | pass→pass | 9,451 | 2,988 | -68% | 1 | 1 | 0% | 1,717 | 1,777 | +3% | 0 | 0 | — |
case-19 | fail→pass | 7,713 | 4,552 | -41% | 1 | 1 | 0% | 1,235 | 2,075 | +68% | 0 | 0 | — |
case-20 | pass→fail | 3,347 | 8,872 | +165% | 1 | 1 | 0% | 712 | 2,777 | +290% | 0 | 0 | — |
case-21 | pass→fail | 7,014 | 9,089 | +30% | 1 | 1 | 0% | 1,377 | 2,955 | +115% | 0 | 0 | — |
case-22 | pass→pass | 9,361 | 10,309 | +10% | 1 | 1 | 0% | 1,837 | 3,267 | +78% | 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 +73 percentage points is the difference between those two pass rates over the 22 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.