Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Runs a team of 50 LLM personas through a 30-round structured debate to predict market direction. Uses the Canvas Agent Swarm Service — the same shared infrastructure other skills use for smaller agent teams. Output is an influence-weighted consensus direction + trade recommendation with a transparent record of every argument, research query, and cross-examination.
.claude/skills/spyderweb47-predict-analysis/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-01 | ✗→✓ | ▲ Improved | 17% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -35% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -34% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -20% | 0% |
> Previously known as swarm_intelligence. The skill id > swarm_intelligence is retained as an alias for backward > compatibility. See docs/PREDICT_ANALYSIS.md for the full technical > walkthrough.
This skill uses the largest team of any skill — 50 agents in total — orchestrated via the shared Agent Swarm Service (core/engine/agent_swarm.py).
| Role(s) | Count | What they do | |---|---|---| | Asset classifier | 1 | Identifies the asset + its price drivers | | Context analyser | 1 | Extracts regime + key levels from bars | | Intelligence gatherer | 1 | Web-searches news / analysis / regulation / indicators | | Personas (bull/bear/neutral/observer) | 50 | Debate the asset for 30 rounds | | Cross-examiner | 1 | Probes divergent personas with targeted questions | | Reporter | 1 | Synthesises final research note |
All coordination — parallelism, timeouts, retries, event recording — is handled by the Agent Swarm Service, not this skill.
When the Canvas has multiple chart windows, the focused chart is the primary asset (drives the full pipeline); siblings are summarised into the intel briefing as portfolio context. Personas reference them naturally in their arguments.
See docs/PREDICT_ANALYSIS.md § 5 for the processor-level normalisation (focused → index 0, missing-dataset warnings, etc.).
| Tool | When | Purpose | |---|---|---| | simulation.set_debate | On completion | Push full debate payload to the store | | bottom_panel.activate_tab | On completion | Switch to DAG Graph tab | | notify.toast | On completion | Toast with consensus summary |
| Tab | Shows | |---|---| | DAG Graph | React Flow pipeline visualisation | | Personalities | 50 persona cards; click → full profile + research trail + live /interview chat | | Debate Thread | Flat list of all messages with sentiment bars + tool chips + agreement references | | Run Stats | Consensus + briefing + market context + data feeds + cross-exams + convergence chart + PDF export + Run Warnings banner |
See docs/PREDICT_ANALYSIS.md § 13. Summary: no streaming (user waits for full 10-30 min run), no persona caching (every run regenerates), global DDG rate limiter serialises web searches, no cross-session memory.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-07 | fail→pass | 11,583 | 3,365 | -71% | 1 | 1 | 0% | 2,144 | 1,557 | -27% | 0 | 0 | — |
case-01 | fail→pass | 19,206 | 17,028 | -11% | 1 | 1 | 0% | 3,511 | 4,121 | +17% | 0 | 0 | — |
case-02 | fail→fail | 19,000 | 14,419 | -24% | 1 | 1 | 0% | 3,388 | 3,468 | +2% | 0 | 0 | — |
case-03 | fail→fail | 17,258 | 19,191 | +11% | 1 | 1 | 0% | 3,223 | 4,229 | +31% | 0 | 0 | — |
case-04 | pass→pass | 8,473 | 7,692 | -9% | 1 | 1 | 0% | 2,136 | 2,757 | +29% | 0 | 0 | — |
case-05 | pass→pass | 13,979 | 10,510 | -25% | 1 | 1 | 0% | 3,441 | 3,395 | -1% | 0 | 0 | — |
case-06 | pass→pass | 10,055 | 10,373 | +3% | 1 | 1 | 0% | 2,437 | 3,367 | +38% | 0 | 0 | — |
case-08 | fail→pass | 8,682 | 1,432 | -84% | 1 | 1 | 0% | 1,667 | 1,082 | -35% | 0 | 0 | — |
case-09 | fail→pass | 7,793 | 1,122 | -86% | 1 | 1 | 0% | 1,547 | 1,024 | -34% | 0 | 0 | — |
case-10 | fail→pass | 10,080 | 3,231 | -68% | 1 | 1 | 0% | 1,855 | 1,488 | -20% | 0 | 0 | — |
case-11 | fail→pass | 6,271 | 1,357 | -78% | 1 | 1 | 0% | 1,259 | 1,099 | -13% | 0 | 0 | — |
case-12 | fail→pass | 7,557 | 1,387 | -82% | 1 | 1 | 0% | 1,471 | 1,105 | -25% | 0 | 0 | — |
case-13 | fail→pass | 8,787 | 1,265 | -86% | 1 | 1 | 0% | 1,716 | 1,008 | -41% | 0 | 0 | — |
case-14 | fail→pass | 6,383 | 2,408 | -62% | 1 | 1 | 0% | 1,220 | 1,219 | -0% | 0 | 0 | — |
case-15 | fail→pass | 4,006 | 1,165 | -71% | 1 | 1 | 0% | 772 | 1,039 | +35% | 0 | 0 | — |
case-16 | fail→pass | 11,229 | 1,324 | -88% | 1 | 1 | 0% | 2,193 | 1,100 | -50% | 0 | 0 | — |
case-17 | fail→pass | 7,956 | 2,443 | -69% | 1 | 1 | 0% | 1,633 | 1,344 | -18% | 0 | 0 | — |
case-18 | fail→pass | 9,655 | 2,148 | -78% | 1 | 1 | 0% | 1,875 | 1,252 | -33% | 0 | 0 | — |
case-19 | fail→pass | 12,676 | 5,693 | -55% | 1 | 1 | 0% | 2,107 | 1,867 | -11% | 0 | 0 | — |
case-20 | fail→pass | 8,789 | 1,616 | -82% | 1 | 1 | 0% | 1,653 | 1,131 | -32% | 0 | 0 | — |
case-21 | fail→pass | 5,969 | 1,208 | -80% | 1 | 1 | 0% | 1,119 | 1,056 | -6% | 0 | 0 | — |
case-22 | fail→pass | 9,277 | 1,893 | -80% | 1 | 1 | 0% | 1,836 | 1,127 | -39% | 0 | 0 | — |
case-23 | fail→pass | 13,569 | 3,810 | -72% | 1 | 1 | 0% | 1,862 | 1,523 | -18% | 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. 23 cases were attempted. The headline lift of +78 percentage points is the difference between those two pass rates over the 23 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.