Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Run STORM phase 1 — multi-perspective research. This skill should be used when the user asks to "research a topic for an article", "find sources on X from multiple perspectives", "do storm research on X", or invokes /storm:research. Discovers personas, runs simulated Q&A per persona in parallel, and produces an information table + deduplicated sources.
.claude/skills/fradser-storm-research/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | -6% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-17 | ✗→✓ | ▲ Improved | -34% | 0% |
| case-18 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-20 | ✗→✓ | ▲ Improved | -37% | 0% |
Phase 1 of the STORM pipeline. Discovers research personas, runs a simulated multi-turn conversation per persona (each grounded in retrieval), and produces the information table that downstream phases consume.
Load storm-engine via the Skill tool. Its "Persona Discovery", "Simulated Conversation", "Retrieval", and "Citation Hygiene" sections govern this phase.
This phase is complete iff research/sources.json exists and has ≥1 entry. If --force is not set and the artifact exists, skip and exit early.
research/ subdir exists.--max-perspective personas (default 3), each with a distinct question category, plus one "Basic fact writer". Write research/personas.json.ToolSearch, look for exa-mcp-server search tools (code-search, research-paper-search, company-search, personal-site-search, financial-report-search, x-search). Record which are available in run-config.json as retriever. If none and not --docs, fall back to WebSearch/WebFetch.storm-researcher subagent per persona in a single message (parallel). Each subagent:max_turns.question_to_query, retrieves, answers with source attribution.max_turns reached.{question, queries, snippets, answer, cited_sources}.research/conversations.jsonl (one JSON object per line). Deduplicate cited sources by URL into research/sources.json, assigning sequential ids. Strip any inline [n] from snippets before storing (citation hygiene).sources.json has ≥1 entry and every persona produced ≥1 turn. If a persona produced nothing, note it but do not fail the whole phase.run-config.json: phases.research = "completed", retriever, source count.If the user reports rate-limit errors during this phase, reduce concurrency by running personas in smaller batches (e.g. 2 at a time) rather than lowering max_turns — mirroring upstream's max_thread_num guidance.
Report: number of personas, total conversation turns, number of deduplicated sources, and the path to research/sources.json.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 14,399 | 8,356 | -42% | 1 | 1 | 0% | 863 | 929 | +8% | 0 | 0 | — |
case-02 | fail→fail | 106,383 | 5,312 | -95% | 1 | 1 | 0% | 547 | 902 | +65% | 0 | 0 | — |
case-03 | fail→fail | 49,974 | 4,810 | -90% | 1 | 1 | 0% | 8,252 | 852 | -90% | 0 | 0 | — |
case-04 | pass→fail | 21,791 | 6,418 | -71% | 1 | 1 | 0% | 3,749 | 923 | -75% | 0 | 0 | — |
case-05 | pass→fail | 51,206 | 9,900 | -81% | 1 | 1 | 0% | 8,223 | 1,229 | -85% | 0 | 0 | — |
case-06 | fail→fail | 3,012 | 33,627 | +1016% | 1 | 1 | 0% | 230 | 5,902 | +2466% | 0 | 0 | — |
case-07 | fail→fail | 3,984 | 14,653 | +268% | 1 | 1 | 0% | 662 | 2,345 | +254% | 0 | 0 | — |
case-08 | fail→fail | 49,007 | 6,180 | -87% | 1 | 1 | 0% | 8,216 | 1,085 | -87% | 0 | 0 | — |
case-09 | pass→fail | 12,252 | 5,207 | -58% | 1 | 1 | 0% | 1,726 | 1,093 | -37% | 0 | 0 | — |
case-10 | pass→fail | 25,060 | 3,939 | -84% | 1 | 1 | 0% | 3,837 | 911 | -76% | 0 | 0 | — |
case-11 | fail→fail | 12,306 | 5,849 | -52% | 1 | 1 | 0% | 1,888 | 1,117 | -41% | 0 | 0 | — |
case-12 | fail→pass | 14,699 | 9,359 | -36% | 1 | 1 | 0% | 2,208 | 2,067 | -6% | 0 | 0 | — |
case-13 | fail→fail | 2,972 | 5,720 | +92% | 1 | 1 | 0% | 430 | 862 | +100% | 0 | 0 | — |
case-14 | fail→fail | 11,636 | 3,702 | -68% | 1 | 1 | 0% | 2,014 | 1,347 | -33% | 0 | 0 | — |
case-15 | fail→pass | 14,891 | 9,360 | -37% | 1 | 1 | 0% | 2,446 | 2,363 | -3% | 0 | 0 | — |
case-16 | pass→pass | 13,141 | 6,517 | -50% | 1 | 1 | 0% | 2,261 | 1,690 | -25% | 0 | 0 | — |
case-17 | fail→pass | 14,639 | 6,161 | -58% | 1 | 1 | 0% | 2,862 | 1,880 | -34% | 0 | 0 | — |
case-18 | fail→pass | 10,836 | 3,178 | -71% | 1 | 1 | 0% | 2,063 | 1,273 | -38% | 0 | 0 | — |
case-19 | pass→pass | 9,292 | 2,938 | -68% | 1 | 1 | 0% | 1,566 | 1,211 | -23% | 0 | 0 | — |
case-20 | fail→pass | 20,753 | 7,068 | -66% | 1 | 1 | 0% | 3,303 | 2,074 | -37% | 0 | 0 | — |
case-21 | pass→pass | 10,991 | 3,303 | -70% | 1 | 1 | 0% | 1,842 | 1,269 | -31% | 0 | 0 | — |
case-22 | fail→pass | 18,543 | 7,328 | -60% | 1 | 1 | 0% | 2,678 | 2,178 | -19% | 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 12 counted toward the lift figure. The other 10 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 +9 percentage points is the difference between those two pass rates over the 12 comparable cases. 5 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.