Deep Research

---
name: deep-research
description: Runs a thorough multi-source research process - decompose the question, gather independent sources, cross-check, synthesize - and delivers a structured, cited brief with confidence levels and gaps. Use when someone asks "research this topic in depth", "give me everything on X", "what does the evidence say about Y", or needs a broad question answered from secondary sources with citations. Do NOT use for designing a study that collects new data - use primary-research instead; for distilling an already-gathered pile of 20+ sources into themes, use research-synthesis; for verifying one discrete claim, use fact-checker.
metadata:
  title: "Deep Research"
---

# Deep Research

Go wide, then deep: triangulate across independent sources and report what was actually found, including the gaps. The costly failure this skill prevents is the confident brief built on an echo - a dozen articles that all trace to one blog post or one press release, presented as consensus. "Not found" is a legitimate finding; a smooth answer papering over thin evidence is not.

## Inputs to collect

1. The research question, restated until the user confirms it. Ambiguity here multiplies through every later step.
2. What decision or output the research feeds. A brief informing a purchase differs from one informing a publication.
3. Scope boundaries: geography, industry, population, and what is explicitly out of scope.
4. Depth and deadline. Default: 3-6 sub-questions, at least two independent sources per sub-question.
5. Known sources or prior work to build on. Treat supplied sources as leads to appraise, not settled evidence.

Label any assumed scope as an assumption.

## Operating procedure

Order matters: sub-questions must exist before gathering, or the research becomes an unstructured pile shaped by whatever surfaced first.

1. **Frame.** Restate the question and define what a good answer looks like - the form (a number, a ranked list, a causal story) and the bar (how sure it needs to be).
2. **Map.** Decompose into 3-6 sub-questions and research each one, rather than the topic as a blob. If a sub-question splits again, it was two sub-questions.
3. **Gather.** Collect diverse, independent sources per sub-question; prefer primary sources - original data, papers, filings, direct statements - over summaries. Two sources that cite the same origin are one source. Record recency and potential bias for each as you go, not retroactively.
4. **Cross-check.** For each sub-question, mark where sources agree, disagree, or run thin. Disagreement is signal: hypothesize why (different definitions, populations, time periods, incentives) instead of averaging it away.
5. **Synthesize.** Answer the original question with the weight of evidence. Separate evidence from inference and label the inferences. Assign each finding a confidence level: High (multiple independent primary sources agree), Medium (mixed or secondary-heavy support), Low (single source or inference).
6. **Audit before delivery.** Re-check each citation is real and locatable, and that every High-confidence finding actually has the sourcing it claims.

## Worked example (findings-table format)

Question: "Is four-day-workweek adoption growing among knowledge-work employers, and what happens to output?"

```
Sub-question            Finding                                   Confidence  Sources
Adoption trend          Growing from a small base; most           Medium      Coordinated-trial reports (primary);
                        employers remain on 5 days                            national labor-survey data (primary)
Output effect           Trials report maintained revenue and                  Trial-organization data (primary);
                        lower attrition - but trial firms         Medium      independent academic follow-up
                        self-selected, so effects may not
                        generalize
Who it fails for        Thin evidence for client-facing and       Low         Single practitioner survey; rest is
                        shift-based work                                      anecdote - GAP: no controlled study found
```

Note the third row: the honest output is a labeled gap, not a stretched conclusion.

## Deliverable

Produce a research brief containing: a summary answering the question in a few sentences up top; findings organized by sub-question, each with citations and a confidence tag; a confidence-and-gaps section stating what is well-supported, what is uncertain, and what is unknown; and a source list with enough detail to locate each item plus a one-line note on its independence and bias.

## Do NOT

- Do not invent or embellish a citation - one fabricated source invalidates the entire brief. Never cite something not actually consulted.
- Do not count correlated sources as corroboration; trace what each actually relies on before calling anything consensus.
- Do not average disagreeing sources into a mushy middle. Present the disagreement and the best hypothesis for why it exists.
- Do not let the gathering order set the narrative - the first three sources found are the most findable ones, not the most representative.
- Do not blur evidence and inference. A reader must be able to tell which sentences report sources and which report reasoning.

## Quality bar

Before shipping: every sub-question has at least two genuinely independent sources or an explicit gap note; every finding carries a confidence tag consistent with its sourcing; disagreements are surfaced with a hypothesis, not smoothed over; the summary is answerable-from and consistent with the findings; every citation is locatable from the detail given.

## Escalation

Route adjacent jobs: designing a study to collect new data is primary-research; compressing an existing corpus of sources into themes and tensions is research-synthesis; a single checkable claim is fact-checker; academic-literature-specific reviews are literature-review or systematic-review. Findings that will drive legal, medical, or financial decisions need professional review before action.