Deep Research

Deep Research

Create high-fidelity research reports with strict format control, evidence mapping, source governance, and multi-pass synthesis.

What makes this produce provider-grade output

Three things drive deep-research quality, in order of impact. This skill is built around them; do not skip them under time pressure.

1. Compute spent through a loop — dozens of queries and full-page reads, not 5-6 snippet searches. Under-spending is the #1 quality leak.
2. Dynamic re-planning — decide what to search NEXT from what you just READ. The initial plan is a hypothesis, not a contract. See references/replanning_loop.md. This is the difference between this harness and static RAG.
3. Verification — a pass that checks every claim maps to a source that actually supports it. See P6/P7.

Semantic search / embeddings are NOT a quality driver for open-web research — a search engine already ranks the open web. Reranking belongs only on a closed corpus, or as a last-mile token-budget trim on already-scraped pages. See references/retrieval_tiers.md.

Architecture: Lead Agent + Subagents + Replan Gate

Lead Agent (coordinator — minimizes raw search context)
  |
  P0: Probe environment (tier) + set effort budget + source policy
  |
  P1: Initial decomposition into sub-questions (HYPOTHESES, not a frozen plan)
  |
  P2: RESEARCH LOOP  ───────────────────────────────────────┐
  |   Dispatch ──→ Subagent A ──→ search broad → READ pages ─┤
  |            ──→ Subagent B ──→ writes task-b.md ──────────┤ (parallel)
  |            ──→ Subagent C ──→ writes task-c.md ──────────┘
  |        |                                                 |
  |   research-notes/  <─────────────────────────────────────┘
  |        |
  |   REPLAN GATE: lead reads notes → writes knowledge-state
  |        (KNOW / THINK / MISSING / CONTRADICTIONS / DECISION)
  |        |
  |   gaps remain + budget left? ──yes──> issue delta-queries, loop again
  |        |
  |        no (coverage sufficient OR budget exhausted)
  |        v
  P3: Build citation registry with source_type + as_of + authority
  P4: Evidence-mapped outline with counter-claim flags
  P5: Draft from notes (never from raw search results)
  P6: Counter-review (claims, confidence, alternatives, contradictions)
  P7: Verify (every [n] in registry, traceability check)
        → then polish to final report with confidence markers (see Output Requirements)

Context efficiency: Subagents' raw search results stay in their context and are discarded. Lead agent sees only distilled notes (~60-70% context reduction).

Note on phase numbering: the research loop is its own phase (P2); citation registry through verify are P3-P7, then a polish step (folded into Output Requirements). This shifts the old V6 numbering by one because the loop is now explicit. The Enterprise pipeline (E1-E7) is unchanged and slots into P2's loop.

Mode & Effort Selection

Determine three things before starting:

• Enterprise Research Mode: Six-dimension data collection with structured analysis frameworks (SWOT, risk matrix, competitive barrier quantification)
• General Research Mode: Standard P0-P8 research pipeline with source governance

Effort budget (the compute lever — the single biggest quality driver):

Ceilings force a STOP; reaching coverage early also forces a stop. Spend the full budget on deep — but spend it through the replan gate, never as a blind query dump. See references/replanning_loop.md.

Source Governance (V6)

Source Accessibility Classification

CRITICAL RULE: Every source must be classified by accessibility:

⚠️ CIRCULAR VERIFICATION BAN: You must NOT:

• Use user's private data to "discover" what they already know about themselves
• Research user's own company by accessing user's private accounts
• Present user's private knowledge as "research findings"

✅ EXCLUSIVE INFORMATION ADVANTAGE: You SHOULD:

• Use user's Crunchbase Pro to research competitors
• Use user's proprietary databases for market research
• Use user's private APIs for investment analysis
• Leverage any exclusive source user provides for third-party research

Source Type Labels

Every source MUST also be tagged with:

Quality Gates:

• standard / deep effort: ≥30% official sources in final approved set
• quick effort: ≥20% official sources
• Maximum single-source share: ≤25% (standard/deep), ≤30% (quick)
• Minimum unique domains: 5 (standard/deep), 3 (quick)

AS_OF Date Policy

Set AS_OF date explicitly at P0. For all time-sensitive claims:

• Include source publication date with every citation
• Downgrade confidence if source is older than relevant horizon
• Flag stale sources in registry (studies >3 years, news >6 months for fast-moving topics)

P0: Probe Environment & Set Policy

Step 1 — Probe the retrieval tier (no hooks; works in Cowork and Claude Code)

If bash + filesystem are available, run the probe:

python3 scripts/probe_env.py research-notes

It writes research-notes/_env.md with the recommended TIER (T1/T2/T3). You may override it if you know the sandbox blocks egress regardless. If bash is unavailable, you are in T3: use native web_search + web_fetch only, on full pages (never snippet-only). See references/retrieval_tiers.md for what each tier means and the T1 install steps.

Step 2 — Confirm core capabilities

Step 3 — Set policy variables

• AS_OF: Today's date (YYYY-MM-DD) — mandatory for timed topics
• EFFORT: quick / standard / deep (sets the loop budget — see Mode & Effort table)
• TIER: T1 / T2 / T3 (from the probe)
• SOURCE_TYPE_POLICY: Enforce official/academic/secondary/journalism/community/other labels
• COUNTER_REVIEW_PLAN: What opposing interpretation to test

Report: [P0 complete] Tier: {T1/T2/T3}. Subagent: {yes/no}. Effort: {quick/standard/deep}. AS_OF: {YYYY-MM-DD}.

When researching a specific company/enterprise, follow this specialized workflow that ensures six-dimension coverage, quantified analysis frameworks, and three-level quality control.

Enterprise Workflow Overview

Enterprise Research Progress:
- [ ] E1: Intake — confirm company entity, research depth, format contract
- [ ] E2: Six-dimension data collection (parallel where possible)
  - [ ] D1: Company fundamentals (entity, founding, funding, ownership)
  - [ ] D2: Business & products (segments, products, revenue structure)
  - [ ] D3: Competitive position (industry rank, competitors, barriers)
  - [ ] D4: Financial & operations (3-year financials, efficiency metrics)
  - [ ] D5: Recent developments (6-month events, strategic signals)
  - [ ] D6: Internal/proprietary sources (or note limitation)
- [ ] E3: Structured analysis frameworks
  - [ ] SWOT analysis (evidence-backed, 4 quadrants × 3-5 entries)
  - [ ] Competitive barrier quantification (7 dimensions, weighted score)
  - [ ] Risk matrix (8 categories, probability × impact)
  - [ ] Comprehensive scorecard (6 dimensions, weighted total)
- [ ] E4: L1/L2/L3 quality checks at each stage transition
- [ ] E5: Draft report using 7-chapter enterprise template
- [ ] E6: Multi-pass drafting + UNION merge (same as general Step 6-7)
- [ ] E7: Present draft for human review and iterate

P1: Initial Decomposition (Hypotheses)

Decompose the research question into initial sub-questions, sized by EFFORT (quick 2-3, standard 4-6, deep 6-10). These are hypotheses about where the answer lives, not a frozen plan. P2's replan gate will revise, drop, and add sub-questions based on what gets read. Do not over-invest in a perfect plan here — getting into the field and reading is what reveals the real structure.

Each initial sub-question includes:

• Expert Role: Specialist persona (e.g., "Policy Historian", "Ecosystem Mapper")
• Objective: One-sentence investigation goal
• Seed Queries: 2-3 starting search queries (the loop will generate more)
• Output: Path to research notes file
• Parallel Group: Group A (independent) or Group B (depends on Group A)

Note: every sub-question is DEEP by default — subagents READ full pages, they do not stop at snippets. "SCAN-only" is not a depth choice; it is a degraded fallback when no fetch/scrape is available at all.

Decomposition Rules

1. Each sub-question covers one coherent area a specialist would own
2. Group A sub-questions must be independent and source-diverse
3. Max 3 dispatched in parallel per round (concurrency limit)
4. Flag time-sensitive claims and citation-aging risk up front
5. Start wide, then narrow — first round explores the landscape; later rounds drill into the specific gaps the gate surfaces

Enterprise Research Integration

When in Enterprise Research Mode, the initial sub-questions map to six dimensions:

• Company fundamentals / Business & products / Competitive position / Financial & operations / Recent developments / Internal-proprietary (or documented limitation)

Report: [P1 complete] {N} initial sub-questions in {M} groups. Entering research loop.

Enterprise Research Mode (Specialized Pipeline)

When researching a specific company/enterprise, follow this specialized workflow that ensures six-dimension coverage, quantified analysis frameworks, and three-level quality control.

E1: Intake

Same as P0/P1 above, plus:

• Confirm the exact legal entity being researched (parent vs subsidiary)
• Select research depth: Quick scan (3-5 pages) / Standard (10-20 pages) / Deep (20-40 pages)
• Identify any specific comparison targets (benchmark companies)

P2: Research Loop (Dispatch → Read → Replan)

Subagents execute using references/subagent_prompt.md and output to references/research_notes_format.md.

This is the engine. It runs in ROUNDS. Each round dispatches subagents that search broad and READ full pages, then a REPLAN GATE decides whether to loop again. Full protocol: references/replanning_loop.md.

Retrieval calls by tier (from P0 probe, in research-notes/_env.md):

• T1: subagents call python3 scripts/retrieve.py search "<q>" then scrape "<url>" for full pages. If a call returns NEEDS_NATIVE, fall back to native tools for that call.
• T2: native web_search to discover, then retrieve.py scrape (trafilatura cleanup) or native web_fetch for full pages.
• T3: native web_search + web_fetch only. Full pages, never snippet-only.

Each round

1. Dispatch up to 3 subagents in parallel (per EFFORT's subagents/round)
2. Each subagent: searches broad → READS full pages → tags source types → writes task-{id}.md
3. Wait for the round to complete
4. REPLAN GATE — lead reads the round's notes and appends a knowledge-state block to research-notes/_knowledge_state.md:
   • KNOW (established, ≥2 sources or 1 official) / THINK (weak, single-source) / MISSING (gaps + the specific next query for each) / CONTRADICTIONS (sources disagree + how to adjudicate) / DECISION (continue with listed delta-queries | stop: coverage sufficient | stop: budget exhausted)
5. If DECISION = continue → dispatch a new round targeting ONLY the MISSING items (delta-queries). Else exit the loop.

At least one gate is mandatory (quick), two (standard), three+ (deep). Running initial sub-questions to completion and skipping straight to P3 is the V6.1 failure mode — forbidden.

Subagent Output Requirements

Each task-{id}.md must contain:

• Sources section: URLs from actual results with Source-Type, As Of, Authority (1-10), Accessibility
• Findings section: ≤10 one-sentence facts with source numbers
• Full-Read Notes: 2-3+ sources READ IN FULL with key data/insights (not snippet paraphrase)
• Gaps section: what was searched but NOT found, alternative interpretations

Without Subagents (Degraded Mode)

Lead agent runs each sub-question sequentially, acting as each specialist, discarding raw results after writing notes. The replan gate still runs — sequential does not mean single-pass. Loop until coverage or budget.

Stop conditions (exit the loop when EITHER holds)

• Coverage sufficient: every original sub-question has a KNOW-level answer, no MISSING item blocks the conclusion, contradictions resolved or flagged.
• Budget exhausted: hit the EFFORT query/round ceiling. On budget-exit, carry residual MISSING items into the report as stated limitations. Never hide a gap by stopping silently.

Enterprise Research: Six-Dimension Collection

Follow references/enterprise_research_methodology.md for per-dimension collection, source-priority matrix, and cross-validation. The six dimensions are the initial sub-questions; the replan gate still applies (e.g. a gap in D4 financials triggers a targeted delta-round).

Key principles: evidence-driven (every conclusion traces to a citable source); multi-source validation (key data needs ≥2 independent sources); restrained judgment (mark speculation); structured presentation (tables, lists).

Run L1 quality check after each dimension (see enterprise_quality_checklist.md).

Status per round: [P2 round-{N}] {N} sources, {M} findings, decision: {continue/stop}. Status exit: [P2 complete] {R} rounds, {N} total sources. Exit: {coverage/budget}. Building registry.

E3: Structured Analysis Frameworks

Apply frameworks from references/enterprise_analysis_frameworks.md in order:

1. SWOT analysis — each entry with evidence + source + impact assessment
2. Competitive barrier quantification — 7 dimensions with weighted scoring → A+/A/B+/B/C+/C rating
3. Risk matrix — 8 mandatory categories, probability × impact → Red/Yellow/Green
4. Comprehensive scorecard — 6-dimension weighted total → X/10

Run L2 quality check after analysis is complete.

E4: Quality Control

Three-level checks from references/enterprise_quality_checklist.md:

• L1 (Data): Source count, attribution, cross-validation, timeliness
• L2 (Analysis): SWOT completeness, risk coverage, barrier scoring, conclusion support
• L3 (Document): Structure compliance, format consistency, readability, appendices

E5: Draft Using Enterprise Template

Use the 7-chapter enterprise report template from enterprise_quality_checklist.md:

1. Company Overview
2. Business & Product Structure
3. Market & Competitive Position
4. Financial & Operations Analysis
5. Risks & Concerns
6. Recent Developments
7. Comprehensive Assessment & Conclusion

Plus appendices: Data Source Index, Glossary, Disclaimer.

E3-E7: Enterprise Analysis, Drafting, and Review

• E3: Structured Analysis — Apply frameworks from references/enterprise_analysis_frameworks.md
• E4: Quality Control — Run L1/L2/L3 checks per references/enterprise_quality_checklist.md
• E5: Draft — Use 7-chapter enterprise template
• E6-E7: Multi-Pass Drafting and Review — Same as P4-P7 below

P3: Citation Registry + Source Governance

Lead agent reads all task notes and builds unified registry.

Registry Process

1. Read every task file's ## Sources section
2. Merge all sources, deduplicate by URL
3. Assign sequential [n] numbers by first appearance
4. Tag: source_type, as_of date, authority score (1-10), task id
5. Apply quality gates:
   • standard / deep: ≥12 approved sources, ≥5 unique domains, ≥30% official
   • quick: ≥6 approved sources, ≥3 unique domains, ≥20% official
   • Max single-source share: ≤25% (standard/deep), ≤30% (quick)
6. Drop sources below threshold and list them explicitly

Registry Output Format

CITATION REGISTRY

Approved:
[1] Author/Org — Title | URL | Source-Type: official | Accessibility: public | Date: 2026-03-01 | Auth: 8 | task-a
[2] ...

Dropped:
x Source | URL | Source-Type: community | Accessibility: privileged | Auth: 3 | Reason: PRIVILEGED SOURCE - NOT ALLOWED

Stats: {approved}/{total}, {N} domains, official_share {xx}%
Privileged sources rejected: {N}

Critical rule: These [n] are FINAL. P5 may only cite from Approved list. Dropped sources never reappear.

Circular verification handling: When researching the user's own company/assets, if you discover data in user's private accounts (e.g., user's domain registrar showing they own domains), you MUST:

1. Reject it from the registry (user already knows this)
2. Note it as "CIRCULAR - USER ALREADY KNOWS" in Dropped
3. Search for equivalent PUBLIC sources (e.g., public WHOIS, news articles)
4. Report from external investigator perspective only

Exclusive source handling: When user EXPLICITLY PROVIDES their paid subscriptions or private APIs for third-party research (e.g., "Use my Crunchbase Pro to research competitors"), you SHOULD:

1. Accept it as "exclusive-user-provided" accessibility
2. Use it as competitive advantage
3. Cite it properly in registry
4. If no public equivalent exists, mark as [unverified] or omit the claim

Report: [P3 complete] {approved}/{total} sources. {N} domains. Official share: {xx}%. Privileged rejected: {N}.

Handling Information Black Box

When researching entities with no public footprint (like the "字节跳动子公司" example):

What an external researcher would find:

• WHOIS: Privacy protected → No owner info
• Web search: No news, no press releases
• Social media: No company pages
• Business registries: No public API or requires local access
• Result: Complete information black box

Correct response:

Findings: NO PUBLIC INFORMATION AVAILABLE

Sources checked:
- WHOIS (public): Privacy protected [failed]
- Company registry (public): Access denied/No API [failed]
- News media: No coverage [failed]
- Corporate website: Placeholder only [minimal]

Verdict: UNABLE TO VERIFY COMPANY EXISTENCE from external perspective
Sources found: 0 (or minimal, e.g., only WHOIS showing domain exists)
Confidence: N/A - Insufficient evidence

DO NOT:

• ❌ Use user's own credentials to "fill in the gaps"
• ❌ Assume the company exists based on domain registration alone
• ❌ Fill missing data with speculation
• ❌ Claim to have "verified" information you accessed through privileged means

DO:

• ✅ Clearly state what an external researcher can/cannot verify
• ✅ Document all failed search attempts
• ✅ Mark claims as [unverified] or omit entirely
• ✅ Downgrade effort to quick or stop if insufficient public sources
• ✅ Recommend direct contact for due diligence

P4: Evidence-Mapped Outline

Lead agent reads notes + registry to build outline.

1. Identify cross-task patterns
2. Design sections topic-first, not task-order-first
3. Map each section to specific findings with source numbers
4. Flag sections needing counter-review
5. Mark recency-sensitive claims with AS_OF checks

Outline format:

## N. {Section Title}
Sources: [1][3][7] from tasks a, b
Claims: {claim from task-a finding 3}, {claim from task-b finding 1}
Counter-claim candidates: {alternative explanations}
Recency checks: {source dates + AS_OF}
Gaps: {limited official evidence}

P5: Draft from Notes

Write section by section using references/report_template_v6.md.

Rules:

• Every factual claim needs citation [n]
• Numbers/percentages must have source
• Add confidence marker per section: High/Medium/Low with rationale
• Add counter-claim sentence when evidence conflicts
• No new sources may be introduced
• Use [unverified] for unsupported statements

Anti-hallucination:

• Lead agent never invents URLs — only from subagent notes
• Lead agent never fabricates data — mark [unverified] if number not in notes

Status: [P5 in progress] {N}/{M} sections, ~{words} words.

P6: Counter-Review (Mandatory)

For each major conclusion, perform opposite-view checks:

1. Could the conclusion be wrong?
2. Which high-impact claims depend on a single source?
3. Which claims lack official/academic support?
4. Are stale sources used for time-sensitive claims?
5. Find ≥3 issues (re-examine if 0 found)

Using Counter-Review Team (Recommended)

For comprehensive parallel review, use the Counter-Review Team:

# 1. Prepare inputs
counter-review-inputs/
  ├── draft_report.md
  ├── citation_registry.md
  ├── task-notes/
  └── p0_config.md

# 2. Dispatch to 4 specialist agents in parallel
SendMessage to: claim-validator
SendMessage to: source-diversity-checker
SendMessage to: recency-validator
SendMessage to: contradiction-finder

# 3. Wait for all specialists to complete

# 4. Send to coordinator for synthesis
SendMessage to: counter-review-coordinator
  inputs: [4 specialist reports]

# 5. Receive final P6 Counter-Review Report

See references/counter_review_team_guide.md for detailed usage.

Manual Counter-Review (Fallback)

If Counter-Review Team is unavailable, perform manual checks:

• Verify every high-confidence claim has ≥2 sources
• Check official/academic backing for key claims
• Verify AS_OF dates on time-sensitive claims
• Document opposing interpretations

Output

Include in final report:

## 核心争议 / Key Controversies
- **争议 1:** [主张 A 与反向证据 B 对比] [n][m]
- **争议 2:** ...

Report: [P6 complete] {N} issues found: {critical} critical, {high} high, {medium} medium.

P7: Verify

Cross-check before finalization:

1. Registry cross-check: List every [n] in report vs approved registry
2. Spot-check 5+ claims: Trace to task notes
3. Remove/fix non-traceable claims
4. Validate no dropped source resurrected
5. Check source concentration for key claims

Report: [P7 complete] {N} spot-checks, {M} violations fixed.

Output Requirements

• Match the requested language and tone
• Preserve technical terms in English
• Respect the report spec and formatting rules
• Include a references section or bibliography

Reference Files

Core Pipeline References (V7)

General Research References

Enterprise Research References (load when in Enterprise Research Mode)

Anti-Patterns

• Frozen plan — running P1's initial sub-questions to completion and going straight to P3 with no replan gate. The #1 thing that makes output static-RAG-grade. At least one gate is mandatory.
• Under-spending compute — 5-6 snippet searches and done. Compute through the loop is the top quality driver; spend the EFFORT budget.
• Snippet satisficing — claiming a fact as established from a search snippet without a full-page read.
• Replanning without reading — writing the knowledge-state from result titles instead of subagent notes.
• Contrarian padding — inventing MISSING items to justify another round. Question answered → stop.
• Bolting on semantic search for discovery — embeddings do not fix open-web research; the search engine already ranks. Reranking only on closed corpus or as a token-budget trim on scraped pages.
• Single-pass drafting; splitting passes by section instead of full report drafts
• Ignoring the format contract; claims without citations
• Mixing conflicting dates without calling out discrepancies
• Copying external AI output without verification; deleting intermediate drafts
• Lead agent reading raw search results — only read subagent notes
• Inventing URLs — only use URLs from actual search results
• Resurrecting dropped sources — dropped in P3 never reappear
• Missing AS_OF for time-sensitive claims — always include source date
• Skipping counter-review — mandatory P6 must find ≥3 issues
• CIRCULAR VERIFICATION — never use user's private data to "discover" what they already know about themselves
• IGNORING EXCLUSIVE SOURCES — when user provides Crunchbase Pro etc. for competitor research, USE IT
• Self-graded quality labels — do not claim "outperforms X" in the output. Prove quality with the evaluation rubric, not a slogan.

Next Step: Verify and Deliver

After completing research, suggest verification and output:

Research report complete: [N] sources cited, [M] claims made.

Options:
A) Verify facts — run /fact-checker on the report (Recommended)
B) Create slides — run /daymade-docs:ppt-creator from the findings
C) Export as PDF — run /daymade-docs:pdf-creator for formal delivery
D) No thanks — the report is ready as-is