Deep Research

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

Quick Start

1. Clarify the report spec and format contract
2. Build a research plan and query set
3. Collect evidence with parallel agents (multi-source) 3.5. Download source PDFs (arXiv + sci-hub)
4. Triage sources and build an evidence table 4.5. Extract paragraphs from downloaded PDFs
5. Outline and section map
6. Multi-pass full drafting (parallel subagents)
7. UNION merge, enforce format compliance, verify citations
8. Evidence and citation verification
9. Present draft for human review and iterate

Core Workflow

Copy this checklist and track progress:

Deep Research Progress:
- [ ] Step 1: Intake and format contract
- [ ] Step 2: Research plan and query set
- [ ] Step 3: Evidence collection (parallel agents)
- [ ] Step 3.5: PDF download (arXiv + sci-hub)
- [ ] Step 4: Source triage and evidence table
- [ ] Step 4.5: PDF paragraph extraction
- [ ] Step 5: Outline and section map
- [ ] Step 6: Multi-pass full drafting (parallel subagents)
- [ ] Step 7: UNION merge and format compliance
- [ ] Step 8: Evidence and citation verification
- [ ] Step 9: Present draft for human review and iterate

Step 1: Intake and Format Contract

Establish the report requirements before any research:

• Confirm audience, purpose, scope, time range, and geography
• Lock output format: Markdown, DOCX, slides, or user-provided template
• Capture required sections and exact formatting rules
• Confirm citation style (footnotes, inline, numbered, APA, etc.)
• Confirm length targets per section
• Ask for any existing style guide or sample report

Create a concise report spec file:

Report Spec:
- Audience:
- Purpose:
- Scope:
- Time Range:
- Geography:
- Required Sections:
- Section Formatting Rules:
- Citation Style:
- Output Format:
- Length Targets:
- Tone:
- Must-Include Sources:
- Must-Exclude Topics:

If a user provides a template or an example report, treat it as a hard constraint and mirror the structure.

Step 2: Research Plan and Query Set

Define the research strategy before calling tools:

• Break the main question into 3-7 subquestions
• Define key entities, keywords, and synonyms
• Identify primary sources vs secondary sources
• Define disqualifiers (outdated, low quality, opinion-only)
• Assemble a query set per section

Use references/research_plan_checklist.md for guidance.

Step 3: Evidence Collection (Deepresearch Tool)

Use the deepresearch tool to collect evidence and citations.

• Run multiple complete passes if coverage is uncertain
• Vary query phrasing to reduce blind spots
• Preserve raw tool output in files for traceability

Parallel Evidence Collection (Academic Research)

For academic literature searches, split evidence collection across multiple parallel agents by data source:

Execution rules:

• Launch all agents simultaneously with run_in_background=true
• Each agent writes to a separate raw output file (e.g., raw_S2.md, raw_arXiv.md, raw_IEEE.md)
• After all agents complete, merge raw files using UNION (Step 4)
• Never run a single agent for all dimensions when parallel execution is possible

File structure (recommended):

<output_dir>/research/<topic-name>/
  deepresearch_pass1.md
  deepresearch_pass2.md
  deepresearch_pass3.md

If deepresearch is unavailable, rely on user-provided sources only and state limitations explicitly.

Step 3.5: PDF Download (Academic Research)

After evidence collection, download source PDFs for full-text extraction:

Download routes (in priority order):

Execution:

• Extract arXiv IDs and DOIs from evidence collection outputs using Grep
• Launch 2 parallel download agents: one for arXiv (curl), one for DOI (sci-hub)
• Save all PDFs to <output_dir>/research/papers/
• Create download_log.md with success/failure tracking
• Validate downloads: PDF files must be >100KB to be considered valid

File structure:

<output_dir>/research/papers/
  paper1.pdf
  paper2.pdf
  ...
  download_log.md

This step is MANDATORY for academic research tasks. Do not skip to Step 4 without downloading available PDFs.

Step 4: Source Triage and Evidence Table

Normalize and score sources before drafting:

• De-duplicate sources across passes
• Score sources using references/source_quality_rubric.md
• Build an evidence table mapping claims to sources

Evidence table minimum columns:

• Source ID
• Title
• Publisher
• Date
• URL or reference
• Quality tier (A/B/C)
• Notes

Step 4.5: PDF Paragraph Extraction (Academic Research)

Read downloaded PDFs and extract paragraphs matching research keywords:

Execution:

• Split PDFs into batches of ~10 by filename alphabetical order
• Launch parallel reader agents (one per batch) with run_in_background=true
• Each reader uses the Read tool with pages parameter to read PDFs in 20-page chunks
• Search for keyword matches: extract the full surrounding paragraph (at least 3 sentences of context)
• Record: page number, section heading, full paragraph text

Output format per paper:

### [filename] Paper Title
#### Page X — Section Heading
> Extracted paragraph text...

Write extracted paragraphs to:

<output_dir>/research/extracts/
  batch1.md
  batch2.md
  ...

These raw extracts feed into Step 5 (outline) and Step 6 (drafting) as primary evidence.

Step 5: Outline and Section Map

Create an outline that enforces the format contract:

• Use the template in references/research_report_template.md
• Produce a section map with required elements per section
• Confirm ordering and headings match the report spec

Step 6: Multi-Pass Full Drafting (Parallel Subagents)

Avoid single-pass drafting; generate multiple complete reports, then merge.

Preferred Strategy: Parallel Subagents (Complete Draft Each)

Use the Task tool to spawn parallel subagents with isolated context. Each subagent must:

• Load the report spec, outline, and evidence table
• Draft the FULL report (all sections)
• Enforce formatting rules and citation style

Implementation pattern:

Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version1.md
Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version2.md
Task(subagent_type="general-purpose", prompt="Draft complete report ...", run_in_background=false) -> version3.md

Write drafts to files, not conversation context:

<output_dir>/intermediate/<topic-name>/version1.md
<output_dir>/intermediate/<topic-name>/version2.md
<output_dir>/intermediate/<topic-name>/version3.md

Step 7: UNION Merge and Format Compliance

Merge using UNION, never remove content without evidence-based justification:

• Keep all unique findings from all versions
• Consolidate duplicates while preserving the most detailed phrasing
• Ensure every claim in the merged draft has a cited source
• Enforce the exact section order, headings, and formatting
• Re-run formatting rules from references/formatting_rules.md

Step 8: Evidence and Citation Verification

Verify traceability:

• Every numeric claim has at least one source
• Every recommendation references supporting evidence
• No orphan claims without citations
• Dates and time ranges are consistent
• Conflicts are explicitly called out with both sources

Use references/completeness_review_checklist.md.

Step 9: Present Draft for Human Review and Iterate

Present the draft as a reviewable version:

• Emphasize that format compliance and factual accuracy need human review
• Accept edits to format, structure, and scope
• If the user provides another AI output, cross-compare and UNION merge

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

Anti-Patterns

• Single-pass drafting without parallel complete passes
• Splitting passes by section instead of full report drafts
• Ignoring the format contract or user template
• Claims without citations or evidence table mapping
• Mixing conflicting dates without calling out discrepancies
• Copying external AI output without verification
• Deleting intermediate drafts or raw research outputs
• Running a single search agent when multiple data sources are available
• Skipping PDF download and relying only on search snippets/abstracts
• Sequential evidence collection when parallel execution is possible
• Waiting for user prompts between steps instead of pipeline execution