Deep research brief

Deep Research Brief

Produces a brief a decision-maker can act on without re-checking your work: the question broken into sub-questions, every claim triangulated across at least 3 independent sources, every claim labeled with confidence, and a numbered citation list. No claim without a source; no source without a URL.

When to Use

• The user needs facts, market data, or comparisons before deciding something.
• A downstream content skill (LinkedIn pack, newsletter) needs a sourced foundation.
• The user says "research", "deep dive", "what's the evidence for", "is it true that".
• Not for: writing the final content piece (hand the brief to linkedin-content-pack or newsletter-assembler), or single-fact lookups a quick search answers in one step.

Quick Reference

Action	Command / Call
Search a sub-question	web toolset search, 2–3 query phrasings per sub-question
Fetch a source	web_extract on the URL; save title, author, date
Check source independence	Compare domains AND original reporting origin
Count sources per claim	python3 -c "..." tally from your notes (step 5)
Label confidence	High = 3+ independent agree; Medium = 2, or 3 with 1 dissent; Low = 1, or sources conflict

Procedure

1. Precheck — confirm the web toolset is available (run one trivial search). If not, stop and tell the user this skill needs web access. Confirm the decision the brief serves: ask "What decision does this inform?" if not stated.
2. Decompose — split the main question into 3–6 sub-questions. Each must be answerable with evidence, not opinion. Write them down first; they become the brief's section headers.
3. Search — for each sub-question, run 2–3 differently-phrased searches. Collect candidate URLs. Prefer primary sources (official docs, filings, original studies, first-party announcements) over aggregators.
4. Extract & log — fetch each promising source. For every claim you keep, log: claim text, source URL, publish date, and whether the source is original reporting or citing someone else. Two outlets quoting the same press release count as ONE independent source.
5. Triangulate — every claim that will appear in the brief needs 3+ independent sources for a High label. Verify your tally with stdlib Python (write your claim→sources map as JSON to a temp file first): python3 -c "import json;d=json.load(open('/tmp/claims.json'));[print(k,len(set(v))) for k,v in d.items()]" Expected: each claim prints its independent-source count. Anything under 3 gets Medium or Low.
6. Label — attach a confidence label to every claim: High / Medium / Low per the Quick Reference rules. If sources actively conflict, say so and show both sides — never average a disagreement into a fake consensus.
7. Deliver — assemble the Output Template. Lead with the answer, not the journey. Citations are numbered and every in-text claim carries its bracketed numbers.

Output Template

# Research Brief: [Question]
**Decision this informs:** [one line]
**Bottom line:** [2–3 sentence answer with confidence label]

## Findings
### [Sub-question 1]
- [Claim] — **High** [1][2][3]
- [Claim] — **Medium** [4][5]
(repeat per sub-question)

## What we couldn't verify
- [Open question + what evidence would settle it]

## Sources
1. [Title — Publisher, Date](URL)
2. ...

Pitfalls

• Echo-chamber sourcing — five URLs, one press release. Recovery: trace each source to its origin; collapse duplicates to one, then search specifically for dissenting or independent coverage.
• Stale data presented as current — a 2023 stat answering a 2026 question. Recovery: log publish dates in step 4; flag anything older than 18 months in-text ("as of [date]") and search for a newer figure before labeling High.
• Scope creep — brief balloons past the decision it serves. Recovery: re-read the decision statement from step 1; cut any finding that doesn't change the decision into a one-line "Related but out of scope" note.
• Confidence inflation — labeling High because the claim "feels obvious". Recovery: the label is mechanical — rerun the step 5 count; if the tally says 2, it's Medium, full stop.

Verification

• Every claim in Findings has ≥1 bracketed citation and a confidence label
• Every High-labeled claim traces to 3+ independent origins (step 5 tally re-run)
• Bottom line answers the stated decision in ≤3 sentences
• All citation URLs were actually fetched this session (no invented links)
• "What we couldn't verify" section exists, even if it says "Nothing material"