deep-research

A harness for research you can actually trust. Most AI "research" is one search and a confident summary. This skill forces breadth, then verification, then citation, so the output is defensible instead of plausible.

Use it when the answer needs to be correct, multi-source, and quotable: market and competitor scans, technical comparisons, due diligence, "what is the current state of X," or any question where being wrong is expensive.

The loop

1. Scope first. If the question is underspecified, ask 2 to 3 sharp clarifying questions before searching (budget, region, time window, use case). A narrow question returns a useful answer; a vague one returns mush.
2. Fan out, do not tunnel. Generate 4 to 8 distinct search angles, not rephrasings of one query. Cover the topic by container (who/what/where), by counter-claim (search for the opposite of your hypothesis), by time (recent vs established), and by primary source (filings, docs, repos, datasets, not just blogs).
3. Fetch and score. Open the actual sources. Score each on recency, authority (primary > reputable secondary > anonymous), and independence (two outlets citing the same press release are one source, not two).
4. Extract claims, not vibes. Pull the specific factual claims that matter. Tag each with its source.
5. Adversarially verify. For every load-bearing claim, look for an independent corroborating source, and actively search for one that contradicts it. A claim with one source is a lead, not a fact. Mark single-source claims as such.
6. Synthesize with citations. Write the report. Every non-obvious claim carries an inline source link. Separate what is well-supported from what is contested or thin.

Verification rules (the part that makes it trustworthy)

• No assertion without a source. If you cannot link it, label it as inference or estimate, not fact.
• Two independent sources for anything material. Note when you only found one.
• Prefer primary. A company's own filing or a project's own repo beats a summary of it.
• Distrust round numbers and viral stats. Trace them to origin; aggregators routinely inflate or mis-attribute figures.
• Date everything. "Current" decays. State the as-of date for time-sensitive claims.
• Surface disagreement. When good sources conflict, present both and say which is better supported and why. Do not average them into a fake consensus.

Scaling depth

For a heavy question, run sub-questions in parallel: split the topic into independent threads, research each thread on its own (fan out, fetch, verify), then merge and dedupe findings before writing. A final pass asks one question: what is still missing, unverified, or single-sourced? That gap list becomes the next round or an explicit "open questions" section.

Output format

# <Question>

## Bottom line
2 to 4 sentences. The answer, stated plainly, with the confidence level.

## Findings
- <claim> [source](url) <confidence: high | medium | low>
- <claim> [source](url) <confidence: ...>
  - contested: <the opposing view> [source](url)

## What is thin or contested
- <single-source or disputed claims, called out honestly>

## Sources
Numbered list of every source used, with what each contributed.

## Open questions
What a follow-up round should chase.

Anti-patterns

• Searching once and summarizing. That is a guess with footnotes.
• Treating the first page of results as the truth.
• Citing a source you did not open.
• Hiding uncertainty to sound authoritative. Confidence levels are a feature, not a weakness.
• Padding the report. Length is not rigor; verified claims are.