Research

Research — Hybrid Router + Fallback

The runtime orchestrator for the research domain. Architecture C: deterministic classification → specialist delegation OR own plan-decompose-search-synthesize-cite workflow.

Portability

Requires WebSearch + WebFetch for the fallback workflow; specialist skills (pulse, grants, litreview, syllabus, patent, dossier, deepread) must be present for delegation to work. Node.js with docx package required if Q2 = document mode. Works in Claude Code CLI natively. In Claude.ai with web tools + Code Execution, the workflow is supported.

Distinct From engineering/autoresearch-agent

These two skills share the word "research" but serve completely different use cases:

• research/research/ (this skill) — research-query router + fallback workflow ("Research X")
• engineering/autoresearch-agent/ — Karpathy's autonomous file-optimization experiment loop ("Make this code faster")

No overlap. They coexist.

Hybrid Architecture (C)

Every invocation produces one of three outcomes:

1. Delegation — Classified as specialist-domain. Routes there. User sees the specialist's output.
2. Fallback execution — Classified as general research. Runs own plan → search → synthesize workflow.
3. Clarification request — Classification ambiguous OR a single bare-noun signal matched. Asks one forcing question (with a recommended answer) to disambiguate, then routes.

The skill never silently runs its fallback when a specialist would have done better. Routing transparency is what makes the hybrid architecture trustworthy.

Specialist Registry

Escalation → deep-research: when a wrong answer is expensive (strategy, comparing N options, hypothesis validation, mapping a field) and rigor matters more than speed, escalate to the deep-research skill instead of the fast fallback workflow — it runs a triangulated, multi-round, adversarial investigation and persists an auditable, reusable research folder. This router is the fast path; deep-research is the heavyweight one.

Agent Integrity Rules

This skill obeys the research-pack convention:

• Execution discipline (fallback only): Sequential searches. 1 q/sec rate limit. Confirm response received before next call.
• Source discipline: Cite only sources returned by this session's tool calls. Training knowledge labeled [Background — not from search] and excluded from counts.
• Three-count tracking (fallback only): Queries sent / sources received / sources cited.
• Retry policy: On failure → wait 3s → retry once → log. After 3 consecutive failures: stop, alert user.
• Routing discipline: Never delegate silently. Always state the decision + accept override.

Phase 1: Grill-Me Intake (2–4 Questions)

Intake is intentionally minimal — the goal is to route fast, not to interrogate. One question per turn.

Q1 (always) — Research question

What's the research question? State it in 1–2 sentences. Specific is better than broad — "AI for healthcare" gets you a vague survey; "How are health systems integrating LLM-based clinical decision support?" gets you a useful answer.

Refuse mush. If user says "research AI", push back once: "What about AI specifically — adoption, safety, capability, funding, regulation, comparison? Pick an angle."

Q2 (always) — Output preference

What output do you want? Pick one:

1. Quick chat briefing (5-min read, markdown in chat)
2. Standalone document (.docx with citations, shareable)

Forcing choice. Document mode triggers deeper search budgets and full audit logs.

Q3 (asked only when classification returns ask or fallback with no signals) — Domain disambiguation

Quick clarification — pick the closest match (recommended: {N} — your question matched a {specialist} signal):

1. Academic literature (papers, peer-reviewed)
2. Industry / trends (what's the buzz, news, sentiment)
3. Specific entity (a company, person, organization)
4. Technology / patents (prior art, IP landscape)
5. Grant funding (NIH, foundations)
6. Course material (syllabus or curriculum)
7. None of the above — run general research

When the classifier returned ask (single bare-noun signal), pre-mark the recommended option. Skip if classification produced a silent route (≥2 signals OR one strong multi-word phrase).

Q4 (asked only if Q3 was needed AND user picked "none of the above") — General-research scope

For general research, what's your time horizon — quick scan (5 searches) or thorough (15 searches)?

Skip if a specialist took over.

Stop condition: After Q4 (or earlier if dependency skips applied), commit and start Phase 2. Most invocations exit intake after Q1 + Q2.

Phase 2: Deterministic Classification

This is deterministic, not LLM-reasoned — for speed, debuggability, and consistency.

SIGNALS = {
  pulse:    ["reddit", "hn", "hacker news", "x.com", "twitter", "buzz",
             "sentiment", "trending", "what are people saying",
             "what's happening", "the conversation around",
             "pulse on", "take the pulse", "current conversation"],
  grants:   ["nih", "grant", "grants for", "r01", "r21", "k-award", "reporter",
             "nosi", "funding", "fda", "study section", "principal investigator"],
  litreview:["literature review", "lit review", "litreview", "pico", "spider",
             "systematic review", "review papers on", "research papers on",
             "papers about", "meta-analysis"],
  syllabus: ["syllabus", "course outline", "curriculum", "reading list",
             "for my class", "for my students", "course material"],
  patent:   ["prior art", "fto", "freedom to operate", "patent",
             "patent landscape", "invention", "novelty search",
             "patent search", "ip landscape"],
  dossier:  ["dossier on", "due diligence", "background check",
             "prep me for", "competitor research", "investor diligence",
             "interview prep", "research my competitor", "background on"],
  deepread: ["deep read", "deeply read", "read this book", "read this pdf",
             "read this document", "extract the claims", "extract claims from",
             "knowledge map", "feynman", "argument map"]
}

# Signals are case-insensitive literal phrases (multi-word substring match).
# Bracketed placeholders (e.g., "research [company]") are intentionally NOT
# signals — they over-trigger on generic "research X" queries that should
# fall back to general research, not auto-route to dossier.
# STRONG signal = multi-word phrase (contains a space): pairs verb with noun
# ("dossier on", "prior art") and routes reliably.
# BARE-NOUN signal = single word ("funding", "fda", "patent", "grant"):
# too weak to silent-route on alone — it must trigger Q3 with a
# recommended answer instead.

For each specialist S:
  score[S] = count of SIGNALS[S] phrases matched in question (case-insensitive substring)

if max(score) >= 2:
  route_to = argmax(score)                  # high confidence — silent route
elif max(score) == 1 and only one specialist has score 1:
  if the matched phrase is multi-word (contains a space):
    route_to = that specialist              # strong phrase — silent route
  else:
    route_to = "ask"                        # bare noun — ask Q3, recommend that specialist
else:
  route_to = "fallback"                     # ambiguous or no match — ask Q3 / run fallback

Implementation: scripts/classifier.py --question "..." returns the routing decision + matched signals + per-specialist scores + (for ask) the recommended specialist. Use it; don't re-implement. The SIGNALS map and rules above are kept phrase-for-phrase in sync with the script — drift = bug.

Phase 3a: Specialist Delegation (≥2 signals OR one strong multi-word phrase)

When delegating:

1. Pass the user's question verbatim plus the output preference (Q2)
2. Let the specialist run its own grill-me intake — do NOT pre-answer specialist questions
3. Return specialist output as the user-visible result
4. Tag the result with [Delegated to: research → {specialist}] in the chat output so the user knows what skill produced it
5. Tag the audit log via scripts/routing_transparency_logger.py --action record_delegation

Phase 3b: Own Fallback Workflow

If routing produced no specialist match (and Q3 confirmed general research), run the 8-step fallback:

1. Decompose — break the question into 3–5 sub-questions (what / why / how / who / what's next). Show the decomposition before searching. scripts/fallback_decomposer.py --question "..." gives a deterministic starting point.
2. Source selection — per sub-question: recency → WebSearch+WebFetch (+Reddit/HN on signal); technical/docs → WebSearch+WebFetch; academic → Consensus MCP if connected, else WebSearch with scholar.google.com site filter; data/numbers → WebFetch primary documents; entity-level → offer dossier re-route.
3. Search — sequential per sub-question, 1 q/sec, 2–4 queries per source, broad-to-narrow.
4. Read + extract — WebFetch high-signal results; note every source URL.
5. Synthesize — 2–4 paragraphs per sub-question with inline citations; surface disagreement when sources disagree.
6. Cross-cutting patterns — 1–2 paragraphs across sub-questions: consensus, controversy, gaps.
7. Output — markdown brief by default; DOCX if user picked document mode.
8. Audit log — three counts (sent / received / cited) + per-source reliability tier (primary / secondary / tertiary).