research

---
name: research
description: Use when a research request doesn't obviously match a more specific specialist — 'research X', 'look into X', 'what do we know about X' — routes to a specialist or runs a fallback search.
category: Research
version: 1.0.0
tools: []
---

# Research — Hybrid Router + Fallback

Route research requests to a specialist skill when one fits, or run an own plan-decompose-search-synthesize-cite fallback when nothing specialist matches. Every research request produces one of three outcomes: **delegation** to a specialist, **fallback execution**, or a single **clarifying question** when the match is ambiguous — never a silent guess.

## Distinct From an Autonomous Optimization Loop

This is a research-query router ("research X"), not an autonomous code/file-optimization loop. If a request is actually about iteratively improving a file or codebase against a metric, that's a different job entirely — don't conflate the two.

## Specialist Registry

| Specialist | Routing signals | Domain |
|---|---|---|
| Pulse | reddit / hn / x / buzz / sentiment / trending / "what's people saying" / "pulse on" / "current conversation" | Multi-source recency research |
| Grants | NIH / grant / R01 / K-award / RePORTER / NOSI / "grants for" / FDA / "study section" | NIH grant-funding intelligence |
| Litreview | literature review / PICO / SPIDER / systematic review / "review papers on" / meta-analysis | Academic literature orientation |
| Syllabus | syllabus / course outline / curriculum / "reading list" / "for my class" | Course supplementary reading |
| Patent | prior art / FTO / freedom to operate / patent / "patent landscape" / novelty search | Patent prior-art + landscape |
| Dossier | "dossier on" / "due diligence" / "background check" / "prep me for" / "competitor research" / "investor diligence" | Decision-grade entity research |

**Escalate to a deep, adversarial multi-round research process** (rather than the fast fallback below) when a wrong answer is expensive — strategy calls, comparing several options, hypothesis validation, mapping an entire field — and rigor matters more than speed.

## Integrity Rules

- **Fallback mode:** search sequentially; cite only sources actually returned this session; label anything else `[Background — not from search]` and exclude it from counts; track queries sent / sources found / sources cited.
- On failure, retry once, then log and move on; after repeated failures, stop and tell the user what's missing.
- **Routing discipline:** never delegate silently — always state the routing decision and accept an override.

## Phase 1: Intake (2-4 questions)

1. **Research question**, 1-2 sentences, as specific as possible — "AI for healthcare" produces a vague survey; "How are health systems integrating LLM-based clinical decision support?" produces a useful answer. Push back once on anything too broad.
2. **Output preference** — a quick chat briefing (5-minute read) or a longer standalone report (deeper search budget, full audit log).
3. **Domain disambiguation** (asked only when classification is ambiguous — see below) — academic literature, industry/trends, a specific entity, technology/patents, grant funding, course material, or "none of the above, run general research." Pre-mark whichever specialist the classifier weakly suggested as the recommended pick.
4. **General-research scope** (asked only if question 3 was needed and the user picked "none of the above") — quick scan (~5 searches) or thorough (~15 searches).

Most requests resolve after questions 1-2; most only need a clarifying question when the topic-to-specialist match is genuinely unclear.

## Phase 2: Deterministic Classification

Classification is **deterministic keyword matching, not LLM reasoning** — for speed, debuggability, and consistent behavior across runs. For each specialist, count how many of its signal phrases (case-insensitive substring match) appear in the question:

- **Two or more signal matches for one specialist** → route silently with high confidence.
- **Exactly one match, and it's a multi-word phrase** (pairs a verb with a noun, e.g. "dossier on," "prior art") → route silently; a strong phrase like this is reliable on its own.
- **Exactly one match, but it's a single bare word** (e.g., "funding," "fda," "patent," "grant") → too weak to silent-route; ask the disambiguation question with that specialist pre-marked as the recommendation.
- **No matches, or multiple specialists tied at one match each** → ask the disambiguation question, or run the fallback if the user picks "none of the above."

Deliberately, bracketed generic patterns like "research [company]" are *not* treated as signals — they over-trigger on ordinary "research X" requests that should stay in the general fallback rather than auto-routing to a specialist like dossier.

## Phase 3a: Delegating to a Specialist

Pass the user's question verbatim, plus the output-format preference, to the specialist and let it run its **own** intake — never pre-answer a specialist's questions on the user's behalf. Return the specialist's output as the visible result, tagged `[Delegated to: research → {specialist}]` so the user knows which skill actually produced it.

## Phase 3b: Own Fallback Workflow

When no specialist matches (and disambiguation confirmed general research), run:

1. **Decompose** the question into 3-5 sub-questions (what / why / how / who / what's next) and show the breakdown before searching.
2. **Pick sources per sub-question** — recency-flavored → general web search (plus Reddit/HN if signaled); technical/docs → web search; academic → an academic-search tool if connected, else web search restricted to scholarly sources; data/numbers → fetch primary documents directly; entity-level → suggest the dossier specialist instead.
3. **Search** sequentially per sub-question, roughly 2-4 queries per source, broad to narrow.
4. **Read and extract** — fetch high-signal results rather than citing snippets; note every source URL.
5. **Synthesize** 2-4 paragraphs per sub-question with inline citations, surfacing disagreement where sources conflict.
6. **Cross-cutting patterns** — 1-2 paragraphs spanning sub-questions: consensus, controversy, gaps.
7. **Output** as a markdown brief by default, or a longer document if the user asked for one.
8. **Audit log** — the three counts (sent/found/cited) plus a rough reliability tier (primary/secondary/tertiary) per source.

## Routing Transparency (Mandatory)

After classifying, always: (1) state the decision in one sentence — "Routing to litreview because you mentioned PICO and meta-analysis"; (2) offer an override — "say so now if you want general research instead, or a different specialist"; (3) proceed with the recommended route if the user doesn't object, with no artificial waiting; (4) if the user overrides, accept it, re-route, and note the override in the audit trail. **Never delegate silently** — this transparency is what makes the routing trustworthy.

## Output Format

**Quick briefing:** title + routing note, then TL;DR (2-3 sentences), Findings (one section per sub-question with inline citations), Cross-Cutting Patterns, Sources (numbered, linked, tiered), and an Audit block (counts + failures).

**Standalone document:** the same structure, formatted as a shareable report rather than a short chat message.

## Error Handling

| Situation | Response |
|---|---|
| Single bare-noun signal (e.g. "funding") | Ask the disambiguation question with the matched specialist pre-marked as the recommendation — never silent-route on a bare noun. |
| Multiple weak matches or none | Ask the disambiguation question. |
| Specialist delegation fails | Note it in chat; offer to retry or fall back to general research. |
| User overrides the routing decision | Accept it, re-route, log the override. |
| Fallback search returns thin results | Say so explicitly; suggest the question may be too niche or too new; never fabricate. |
| Repeated failures in fallback mode | Stop, tell the user what's missing, share what was collected. |
| Non-research request (e.g. "write me code") | Decline politely and suggest the appropriate skill. |
| A specialist isn't available in this environment | Drop it from classification scoring and route to fallback or the next-best specialist. |