Research Scoping

Research Scoping

A single-context triage pass that maps a question before you decide to spend real tokens on it. It runs the multi-perspective move from Stanford's STORM method compressed into one prompt: discover the perspectives the question demands, surface where they fight, sketch how a full answer would be structured, then say whether to handle it inline or escalate. It retrieves nothing and cites nothing. The output is a map and a plan, framed as questions to pursue, not answers.

The value is economic and analytical at once. A full COMPLEX /ann run costs hundreds of thousands of tokens; this pass costs a fraction of that in one context with no subagent fan-out. So it either saves you the big run, or sharpens the roster so the big run is better aimed.

Lane — when this skill, not another

• research-scoping (this): map a question's perspectives and tensions in one cheap pass, then triage. No retrieval, no citations, no subagents, no deliverable.
• /grill-mel: interview Ane one question at a time to sharpen a design she already holds. research-scoping runs the opposite direction, handing Ane a map without interviewing her.
• /ann: full orchestration that produces the deliverable. research-scoping is the front door you walk through before deciding to open it.
• /researcher, /evidence-synthesis: retrieve and cite real sources. research-scoping never does; it hands off to these when the question needs evidence.
• /decision-debate: token-heavy multi-agent debate on one contested decision. research-scoping is single-context and spawns nothing.

If the question clearly warrants the deliverable, skip scoping and go straight to /ann or the right builder.

The honest-by-construction rule

This skill must never print a citation or assert a fact about IPPF, an MA, a partner, a contact, a figure, or a date. Everything it surfaces is a hypothesis to verify, by design. A fast, un-sourced map that looks authoritative is the failure mode the whole MEL system exists to prevent, so the guardrail is structural, not stylistic:

• No Author (year) citations. If a framework is relevant, name the analytic move ("test rival explanations") and add it to the claims-to-verify list for the handoff. Do not cite it.
• No asserted facts. Where a fact would matter, write ⚠️ to verify: [what] instead of a value.
• Close every output with the banner below, so the reader cannot mistake the map for findings.

Method — one context, no fan-out

Run every step in a single pass. Light context reads (the conversation, a named project file) are fine to understand the question. Do not run a literature search or spawn specialists; that is the handoff's job, and reaching for it means you are in the wrong skill.

1. Restate the question in one line, so the scope is explicit and shared.
2. Discover perspectives. Derive the 4 to 6 angles the question genuinely demands, not a generic checklist. This is the same move Vi runs at SELECT, but standalone and up front. Name each angle plainly.
3. Sharpest questions. For each perspective, the one or two questions it would press hardest.
4. Tension map. Where these perspectives most likely disagree. This is the payload: the conflict, not a smoothed summary. One line per live tension. If you cannot find three real tensions, the question is probably SIMPLE.
5. Shape of a full answer. A skeleton outline of how the real deliverable would be structured if commissioned.
6. Triage verdict. SIMPLE (handle inline now) or COMPLEX (escalate). For COMPLEX, name a seed specialist roster and the claims that would need verifying. For SIMPLE, say what the inline answer would cover.

Output structure

Use this template every time:

Scoping map: [question]

1. Question — one-line restatement.
2. Perspectives — the 4 to 6 angles, named.
3. What each presses on — the sharpest question per perspective.
4. Where they conflict — the tension map, one line per live disagreement. This is the core.
5. Shape of a full answer — skeleton outline.
6. Verdict — SIMPLE or COMPLEX, the recommended next step, and for COMPLEX a seed roster plus a ⚠️ to verify: list.

End with:

Scoping map, not findings. No claim here is sourced. Treat each line as a hypothesis to verify, not an answer. To produce the real thing, hand this map to /ann, /researcher, or /evidence-synthesis.

Standing rules

• Stay cheap. One context, no subagents, no retrieval. The moment you reach for a literature search, hand off.
• Recommend, do not survey. The verdict carries a pick (SIMPLE or COMPLEX) and a named next step, not a menu.
• Voice. Plain English (Flesch-Kincaid grade 9 to 10), gloss any MEL term on first use, no em-dashes in body prose. Pass the translatability test: would a Romanian, Tunisian, Ethiopian, or Vietnamese English-speaker understand it on first read?
• Three perspectives minimum. The discovery move forces this. Fewer than three real angles is itself a signal the question is SIMPLE.

Close

Deliver the map, then offer the handoff:

"Map done. Verdict: [SIMPLE / COMPLEX]. Want me to take this to /ann, /researcher, or /evidence-synthesis, or handle it inline if SIMPLE?"

Do not start producing the deliverable from inside this skill. The map is the product; the handoff produces the answer.