STORM-style multi-perspective research

Research before concluding. Follow the core STORM pattern: discover perspectives, let each perspective ask source-grounded questions iteratively, then synthesize the collected evidence into an outline and final report. Do not reduce the method to generic stakeholder role-play or a list of opinions.

Read references/storm-method.md before applying the workflow. Read references/perspective-discovery.md when selecting perspectives. Read references/evidence-rubric.md whenever sources or consequential conclusions are involved. Read references/output-templates.md before writing the final answer.

Scope

Use this skill as an evidence-research process. Center it on an unresolved question about external reality: facts, source origin, stakeholder experience, competing explanations, effects, or important unknowns. Its deliverable is an evidence map or synthesis that can inform later learning, decisions, writing, or design.

If later work exposes a fresh evidence question, research that bounded subproblem and return the findings to the surrounding task. Stop at evidence-backed understanding and validation needs; leave product specification and implementation outside this research process.

Announce use

In the first progress update, state:

• that the STORM-style skill is being used;
• the real research question and intended use;
• the initial perspective and evidence-discovery plan;
• whether the work is a quick map, full evidence research, or validation design.

At completion, include a compact STORM execution summary so the user can see that the skill actually ran.

Set the purpose and depth

Choose the primary purpose:

• 理解或溯源: understand an idea, field, document, history, or debate.
• 诊断或需求调研: explain an outcome or discover real tasks, pain, behavior, and causes.
• 评估或决策: compare a feature, tool, method, intervention, or realistic options.
• 复盘或风险研究: reconstruct an outcome or identify credible failure paths and controls.

Choose proportionate depth:

• 快速地图: use supplied material and limited retrieval; label unverified areas.
• 完整调研: run multiple perspective tracks with current, relevant sources. Use by default for “深入研究”.
• 验证设计: add interviews, behavioral data, experiments, pilots, or other tests when documentary research cannot resolve the key uncertainty.

Never present a quick map, simulated perspective, or model memory as completed field research.

Workflow

1. Frame the real research question

Write one sentence:

> Research X in order to understand / diagnose / decide Y, within constraints Z.

Identify the intended use, affected people, time horizon, success meaning, and cost of being wrong. Infer low-risk missing details and label them. Ask only when a missing choice would materially change the research.

2. Build an orientation map, not an answer

Survey the user-provided material and a small set of relevant, high-quality adjacent sources or related topics. Use their structures, recurring themes, named actors, methods, and disagreements to reveal possible coverage areas.

Treat this as a seed map for discovering perspectives and questions. Do not lock the final outline or conclusion yet. If retrieval is unavailable or disallowed, use supplied material and state the resulting coverage limit.

3. Discover topic-specific perspectives

Choose 3–7 perspective tracks that would ask materially different questions. Include a basic-coverage track so essential facts are not lost. Derive the remaining tracks from the topic and orientation evidence, not a universal persona list.

For each track, record:

• its focus;
• the evidence or experience it uniquely seeks;
• questions it is likely to ask that other tracks would miss;
• its likely blind spot or incentive.

A perspective may be a real stakeholder, practitioner, affected group, critic, historical school, scientific method, economic lens, safety lens, or other evidence-bearing viewpoint. Do not create five differently named tracks that share the same questions.

4. Run independent question tracks

Keep each perspective's inquiry separate until synthesis so early consensus does not erase diversity.

For each track:

1. Ask one material question at a time.
2. Break a complicated question into searchable subqueries when needed.
3. Retrieve and filter relevant sources.
4. Answer only to the strength of the retrieved evidence; cite it and mark missing support.
5. Let the answer generate the next follow-up question.
6. Avoid repeating questions already answered by that track or another source.
7. Stop when new rounds add no material evidence, the question leaves scope, or the available sources cannot responsibly answer it.

When a useful answer reveals a new mechanism, exception, affected group, contradiction, or source, ask at least one follow-up rather than returning immediately to a preset checklist.

The “writer” and “expert” in this process are simulated research roles. The expert is only a source-grounded answer function, not a real interviewee or independent witness. Never report these conversations as user research.

5. Maintain an evidence ledger

For each material claim, record:

• claim and plain-language meaning;
• source producer and date;
• what the source actually supports;
• source class and independence;
• applicable conditions and limits;
• counterevidence or conflict;
• evidence label from the rubric.

Prefer direct evidence and primary or authoritative sources. Treat first-party material as authoritative for what its owner states, not independent proof of real-world outcomes. Trace repeated claims back to their common origin before calling them corroborated.

When tracing ideas, distinguish verified origin, later adaptation, conceptual similarity, and unproven attribution.

6. Check coverage and contradictions

Compare the independent tracks only after their questioning rounds. Identify:

• unique findings from each track;
• meaningful agreement supported by genuinely independent evidence;
• contradictions in facts, definitions, metrics, samples, incentives, values, time horizons, or causal explanations;
• missing people, methods, geographies, time periods, or failure cases;
• possible Internet or source-selection bias;
• facts that are merely adjacent rather than causally connected.

For each important contradiction, state the smallest new evidence or test that could resolve it. Add a targeted follow-up round only when it could materially change the result.

7. Build the outline after research

Create two lightweight structures:

• a baseline skeleton from the user's question and essential background;
• an evidence-derived structure from the perspective conversations and ledger.

Merge them into the final outline. Preserve important findings that the baseline missed, remove unsupported sections, and organize around the user's purpose rather than around whichever source was longest.

For decisions, structure the synthesis around options, criteria, effects, trade-offs, and uncertainty. For learning, structure it around meaning, origins, mechanisms, disagreements, boundaries, and applications.

8. Write with citations and a human-language layer

Write section by section using the evidence relevant to that section. Place citations near the claims they support. Do not fabricate connective claims between unrelated facts.

Always lead the final answer with a short 人话版. For every material technical term, abstract claim, or recommendation, be able to give:

• 专业说法: the necessary original term or claim;
• 人话意思: what it means in ordinary language;
• 为什么重要: what changes if it is true;
• 具体例子: what it looks like in real life;
• 边界: where the simplified explanation stops being accurate.

For abstract or jargon-heavy topics, include one short story, fable, or everyday analogy and map it back to reality. Explain the analogy's limit. Simplify the expression, not the underlying constraints, disagreement, or uncertainty.

9. End according to the user's purpose

• For learning: provide the clearest mental model, origins, boundaries, competing interpretations, and practical relevance.
• For diagnosis or demand research: rank competing causes or needs and identify evidence that would distinguish them.
• For evaluation or decision: compare realistic options, conditions, downside, dissent, reversibility, and opportunity cost.
• For review or risk work: show contributing conditions or credible failure chains, controls, owners, warning signals, and unresolved exposure.

Do not automatically prescribe a pilot. When validation is useful, recommend the smallest action that can reduce the largest important uncertainty. Label proposed sample sizes, durations, budgets, weights, and thresholds as provisional unless supported by a cited benchmark or user policy.

Quality gates

Before finalizing, verify:

• Perspectives were discovered for this topic rather than copied from a generic role list.
• Perspective tracks generated materially different questions.
• At least one follow-up was driven by newly retrieved information when the evidence created a useful lead.
• Answers are grounded in retrieved or supplied sources, not the simulated expert's authority.
• Simulated conversations are not described as interviews or independent confirmation.
• The outline was revised after research rather than fixed before it.
• Source origin, later adaptation, and conceptual similarity are not conflated.
• Sources repeating one origin are not counted as independent convergence.
• Internet bias and over-association of unrelated facts were considered.
• Facts, supported judgments, disputed explanations, hypotheses, and unknowns are separated.
• The first screen is understandable without specialist knowledge.
• The human-language explanation preserves important conditions and limits.
• The ending matches the user's purpose rather than forcing the same report shape every time.

If a gate cannot be satisfied, state what is missing and provide the strongest bounded result available.

Execution summary

Finish with:

STORM 执行摘要

• 研究目的与深度
• 发现并采用的视角
• 关键追问如何演进
• 已用证据及独立性
• 模拟、未验证和未覆盖内容
• 主要矛盾或剩余未知
• 建议的下一步（如需要）