Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when a plan has parallel tasks and the harness supports dispatching subagents.
.claude/skills/hashgraph-online-subagents/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 43% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 47% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -2% | 0% |
If topology=multi-module (triage announcement or coordinator spec/plan declaration), read ../references/multi-module.md.
Fresh context per task. No session history.
Subagents have file-read tools. Don't inline what they can read.
TASK: <one sentence>
ROLE: <implementer | spec-reviewer | quality-reviewer>
MODEL: <fast | balanced | strongest>
REPOSITORY: <registered repo id and path>
MODULE: <module id and root>
FILES: <paths>
SPEC: <docs/staging/specs/...#anchor>
ACCEPTANCE: <test or cmd>
EXTRA: <only what's NOT in referenced files>ROLE is required. See role charters below for what each role owns and does not own.
MODEL is optional. Omitted → uses harness default. Resolution: .praxis/model-tiers.yaml in project root or user home maps the three tiers (fast, balanced, strongest) to concrete model IDs. No config file → all subagents use harness default.
Pasting >10 lines from the spec? Stop - let the subagent read it.
Expand milestone → atomic steps at dispatch time, not in the plan.
One task has one repository and module write scope. Agents may read shared coordinator contracts, but must not edit another module or repository. Never dispatch two writers to overlapping module roots. Integration tasks are dispatched only after every required module reports local acceptance.
implementer → on DONE: spec-reviewer (matches spec?) → quality-reviewer (review) → mark complete, continue. The coordinator marks - [x], never the subagent.
DONE_WITH_CONCERNS - address if correctness; note if observation.NEEDS_CONTEXT - supply the fact, re-dispatch.BLOCKED - diagnose (missing context / stronger model / too big / plan wrong). Never silently retry.MODEL: fast | balanced | strongest — optional, omitted → harness default.
Resolved via .praxis/model-tiers.yaml in project root or user home. Template at Praxis repo root: model-tiers.example.yaml.
The coordinator decides the tier at dispatch time based on task complexity — it is not hardcoded per role. The role charters below list defaults; override when:
| Scenario | Tier | Examples | |----------|------|----------| | Mechanical edits, simple checks | fast | rename a constant, fix a typo, pattern-match review | | Standard implementation, single-file feature | balanced | add a function, modify one module, write routine tests | | Multi-file refactor, concurrency/state-machine, safety review | strongest | implement a protocol, review crash-recovery, cross-module integration |
balanced. Override to fast for mechanical edits, strongest for multi-file protocol/state-machine implementation.fast. Override to balanced when spec has ≥3 cross-referencing rules or state enums.review skill against the diff — convention compliance, simplification opportunities, idiomatic patterns, security, scope.strongest. Override to balanced for single-file changes where deep reasoning is unnecessary.- [x].All done → review whole diff → ship.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 10,271 | 3,280 | -68% | 1 | 1 | 0% | 1,756 | 1,534 | -13% | 0 | 0 | — |
case-01 | fail→pass | 13,711 | 10,769 | -21% | 1 | 1 | 0% | 1,534 | 2,187 | +43% | 0 | 0 | — |
case-02 | fail→pass | 8,523 | 8,675 | +2% | 1 | 1 | 0% | 1,609 | 2,365 | +47% | 0 | 0 | — |
case-03 | fail→pass | 8,711 | 12,406 | +42% | 1 | 1 | 0% | 1,542 | 2,324 | +51% | 0 | 0 | — |
case-04 | fail→pass | 15,017 | 9,974 | -34% | 1 | 1 | 0% | 1,728 | 1,899 | +10% | 0 | 0 | — |
case-05 | pass→pass | 22,596 | 9,977 | -56% | 1 | 1 | 0% | 1,842 | 1,969 | +7% | 0 | 0 | — |
case-07 | fail→pass | 9,056 | 7,794 | -14% | 1 | 1 | 0% | 1,528 | 1,504 | -2% | 0 | 0 | — |
case-08 | pass→pass | 18,102 | 5,550 | -69% | 1 | 1 | 0% | 2,180 | 1,977 | -9% | 0 | 0 | — |
case-09 | pass→pass | 9,385 | 3,121 | -67% | 1 | 1 | 0% | 1,702 | 1,571 | -8% | 0 | 0 | — |
case-10 | fail→pass | 12,316 | 9,760 | -21% | 1 | 1 | 0% | 1,939 | 1,865 | -4% | 0 | 0 | — |
case-11 | fail→pass | 15,172 | 3,747 | -75% | 1 | 1 | 0% | 1,560 | 1,733 | +11% | 0 | 0 | — |
case-12 | fail→pass | 13,453 | 8,567 | -36% | 1 | 1 | 0% | 1,481 | 1,511 | +2% | 0 | 0 | — |
case-13 | fail→pass | 12,890 | 7,703 | -40% | 1 | 1 | 0% | 1,325 | 1,403 | +6% | 0 | 0 | — |
case-14 | pass→pass | 3,727 | 2,093 | -44% | 1 | 1 | 0% | 615 | 1,335 | +117% | 0 | 0 | — |
case-15 | fail→pass | 10,860 | 2,987 | -72% | 1 | 1 | 0% | 1,671 | 1,463 | -12% | 0 | 0 | — |
case-16 | fail→pass | 14,748 | 7,428 | -50% | 1 | 1 | 0% | 1,579 | 1,396 | -12% | 0 | 0 | — |
case-17 | fail→pass | 17,570 | 2,114 | -88% | 1 | 1 | 0% | 1,865 | 1,299 | -30% | 0 | 0 | — |
case-18 | fail→pass | 15,267 | 2,305 | -85% | 1 | 1 | 0% | 1,709 | 1,403 | -18% | 0 | 0 | — |
case-19 | fail→pass | 10,496 | 8,500 | -19% | 1 | 1 | 0% | 1,663 | 1,557 | -6% | 0 | 0 | — |
case-20 | pass→fail | 21,802 | 8,628 | -60% | 1 | 1 | 0% | 3,246 | 1,241 | -62% | 0 | 0 | — |
case-21 | pass→pass | 2,851 | 7,176 | +152% | 1 | 1 | 0% | 479 | 1,391 | +190% | 0 | 0 | — |
case-22 | pass→pass | 16,795 | 13,395 | -20% | 1 | 1 | 0% | 2,667 | 3,091 | +16% | 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 21 counted toward the lift figure. The other 1 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 +59 percentage points is the difference between those two pass rates over the 21 comparable cases. 1 case got worse with the skill loaded, and it is 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.