Deep research

Deep Research

One workflow, three modes. Pick the mode from the user's intent:

Mode	Use when
trends	"what's happening in X", "emerging patterns", "who are the players"
analyze	"go deep on X", "tradeoffs of X", "SWOT", "compare perspectives"
factcheck	document or list of claims handed in for verification

All three share the same workspace + checkpoint discipline below. Read this section once, then jump to your mode.

Workspace & Checkpointing (all modes)

Workspace: .research/<YYYY-MM-DD>/<mode>_<session>/

00_plan.md         # written before any research
notes.md           # rolling findings, source URLs
checkpoint_NNN.md  # batch results (factcheck) or section drafts (trends/analyze)
report.md          # final output, assembled from disk — never from memory

Rules that apply to every mode:

1. Plan first. Write 00_plan.md before searching anything. Include the question, sub-questions, and the sources you intend to hit.
2. Disk is truth, memory is scratch. Append findings to notes.md as you go with source URLs. If the session dies, you resume from disk.
3. Checkpoint on a budget. Every ~10 research steps OR when context approaches ~150K tokens, write a checkpoint and drop detailed research from working memory — keep only 2–3 line summaries.
4. Assemble the final report from disk. Use cat / Read over checkpoints. Never reconstruct from memory.
5. Resume, don't restart. On error, ls the workspace, read the last checkpoint, continue from the next unprocessed item.

Mode: trends

Goal: identify the current state and trajectory of a space.

1. Plan sub-questions: who are the players, what's new in the last 6–12 months, what are adoption signals, what are the contrarian takes.
2. tavily_search for each sub-question. For technical topics, also resolve-library-id + query-docs (Context7) to ground claims in current docs.
3. Note publication dates aggressively — anything older than 12 months gets flagged as "background, not signal."
4. Final report.md:
   • State of the space (2–3 paragraphs)
   • Key players (table: name, focus, signal)
   • Emerging patterns (bullets, each with a dated source)
   • Contrarian / risk signals
   • Where this is heading (1 paragraph, clearly labeled as opinion)

Prioritize actionable insight over exhaustive coverage.

Mode: analyze

Goal: deep, balanced decomposition of a single topic.

1. Plan: break the topic into core components, stakeholder perspectives, and decision axes.
2. Research each component with tavily_search / tavily_extract. For each, capture both the strongest case for and the strongest case against.
3. Final report.md:
   • Executive summary (3–5 sentences, no hedging)
   • Components (one section each)
   • Perspectives & counterarguments
   • SWOT (or equivalent: tradeoffs, risks, opportunities)
   • Recommendation with reasoning, clearly separated from facts

Depth over breadth. Nuance over coverage.

Mode: factcheck

Goal: extract every factual claim from a document and label each one with evidence.

Setup

• Read the document.
• Extract every factual claim (statements presented as objective truth — stats, dates, quotes, attributions, causal claims). Write all of them, numbered, to 00_claims_extracted.md before any verification.
• Skip opinions, predictions, value judgments. Mark them in a separate opinions.md if useful for the final summary.
• Categorize: statistical / historical / scientific / attributional. Order by significance.

Verification (batches of 8–10)

For each batch:

• tavily_search for current/breaking claims, tavily_research for synthesis-heavy claims, tavily_extract for primary-source verification.
• Aim for ≥2 independent sources per claim. Note publication dates and source authority.
• Watch for context manipulation (accurate quote, misleading framing) and outdated facts presented as current.
• Write checkpoint_NNN.md with full findings for the batch, then drop the details from memory — keep only claim# | verdict | confidence lines.

Checkpoint format:

# Checkpoint NNN — Claims X–Y

## Claim #X: <verbatim claim>
**Status**: VERIFIED | FALSE | MISLEADING | UNVERIFIABLE | OUTDATED
**Evidence**: <urls>
**Analysis**: <what the sources say vs. the claim>
**Confidence**: HIGH | MEDIUM | LOW

Final report

Read all checkpoints from disk and assemble report.md:

## FACT-CHECK REPORT: <doc title>

### Summary
<N claims checked: V verified, F false, M misleading, U unverifiable, O outdated>

### Detailed Findings
<one block per claim, same format as checkpoint>

### Overall Assessment
<one paragraph on the document's factual reliability>

Be impartial. Distinguish FALSE (contradicted by evidence) from UNVERIFIABLE (insufficient evidence either way). Cite sources for every verdict.

Cross-mode rules

• Never fabricate sources. If you can't find evidence, say so.
• Never include raw page dumps in the final report — summarize and link.
• Flag conflicting evidence rather than picking a winner silently.
• Mask any PII / credentials encountered during searches.