Research Methods

Research Methods (multi-perspective, fast, self-critiqued)

Apply the judgment of a strong research analyst who knows that the value isn't in the first answer — it's in asking the same question from five angles, finding where the answers fight, and grading your own confidence before you act. One prompt gives you the majority view; this gives you the blind spots.

How to use this skill

1. Read research-methods-guide.md in this directory — the 4-phase method, how to choose perspectives, the verify gate, and the limits (when to stop trusting the model and pull real sources). Apply it.
2. For the four copy-paste prompts (and a domain-swapped perspective set), read examples.md.
3. Run the phases in order and keep perspectives independent — generate each lens before it sees the others, or they collapse into one view. End with the peer-review phase every time; it's the step that separates a briefing from a confident guess.

The essentials (full detail in research-methods-guide.md)

• One prompt returns the majority framing. The breakthrough (Stanford STORM, NAACL 2024) is multi-perspective question asking — independent expert lenses catch what single-prompt research never sees.
• Phase 1 — Multi-perspective scan. Simulate 4–6 genuinely different experts (default: practitioner, academic, skeptic, economist/incentives, historian). For each: core position, strongest evidence, and the one thing only they would tell you.
• Phase 2 — Contradiction map. Where do lenses clash (that's where understanding lives)? What do all agree on (likely true)? What did none address (the field's blind spot — often the most valuable finding)?
• Phase 3 — Synthesis. A briefing: one-paragraph nuanced summary, findings ranked by reliability (with which lenses support/challenge each), the non-obvious connection, and a specific action.
• Phase 4 — Peer-review gate (never skip). Make the model grade its own briefing: confidence scores, weakest link, bias check (did one voice dominate?), missing 6th perspective, overall grade + fixes.
• Perspectives are a tool, not magic — pick lenses that actually differ for your topic. Swap the defaults per domain (for an infra decision: SRE, security, FinOps/cost, vendor, outage-historian).
• The verify gate matters most because the model is confidently wrong. STORM's documented weakness is no self-critique → source bias and fact misassociation. Phase 4 mitigates; it does not eliminate.
• This organizes reasoning; it does not replace primary sources. For load-bearing or high-stakes claims, verify against real sources / tool-backed retrieval → the deep-research harness, web search, [[verification-and-debugging]].
• Independence is the whole trick. If each perspective can see the others before answering, you've rebuilt the single-prompt majority view with extra steps.

Related skills

• [[ai-research-science]] — research-scientist content depth on ML/AI topics; this skill is the method for researching any topic, ML or not.
• [[ml-evaluation-evals]] — LLM-as-judge done right (position/verbosity/self-preference bias): the discipline behind the Phase-4 self-critique gate.
• [[verification-and-debugging]] — verify load-bearing claims to root cause instead of trusting a fluent synthesis.
• [[spec-driven-development]] — sharpen the research question/scope before you spend prompts; a vague topic yields a vague briefing.
• [[staff-plus-engineering]] — turning a briefing into a decision doc / recommendation with a clear ask.