Claude Workflow

Deep-research workflow (fan-out + verify)

Multi-source research: fan out searches, fetch sources, verify claims, synthesize a cited report.

Ada · 31 points · 913 475

#research

## Goal Turn one fuzzy question into a cited, fact-checked brief — by fanning out searches, reading sources, and verifying each claim before it lands. ## Why not just ask A single search-and-summarize pass has three predictable failure modes, and this workflow is built around them: - **One phrasing, one slice of the web.** Search engines reward the phrasing you used. Ask "is X better than Y" and you get the comparison posts; you never see the migration post-mortems that would actually answer you. - **Consensus laundering.** Twenty blogs citing the same original press release look like twenty sources. They are one source with nineteen echoes. - **Silent staleness.** The top result is the best-optimized page, not the most current one. In fast-moving domains the accepted answer is often eighteen months out of date. ## Steps ### 1. Decompose Break the question into 4–8 sub-questions that together cover it. Write down, in advance, what a complete answer must address — that list is what you check coverage against later, and writing it *after* you've read sources means you'll unconsciously scope it to what you found. Include at least one sub-question that could falsify your expected answer. ### 2. Fan out Run parallel searches, one angle per sub-question, and deliberately vary the *mode* of search rather than the words: - **By entity** — the product, company, or person by name - **By mechanism** — how the thing works, without naming it - **By counter-argument** — "problems with", "why we moved off", "post-mortem" - **By time** — restricted to the last six months, and separately to before the thing was popular - **By venue** — the primary docs, the standards body, the filings, the source Each mode is blind to what the others surface. That's the point. ### 3. Read, don't skim Open the promising sources. Extract each claim *with the passage that supports it*, plus URL and publication date. A claim without its supporting quote will be subtly reworded by the time it reaches the brief, and the rewording is where the errors enter. Record the **source type** too: primary (docs, filings, source code, the announcement itself), secondary (reporting that cites primaries), or tertiary (aggregators, listicles, AI-written SEO pages). Tertiary sources support nothing on their own; they're leads. ### 4. Adversarially verify For each load-bearing claim, ask: **who disagrees, and why?** Then go find them. - Corroborate with a second *independent* source — one that isn't citing the first. Check what each cites; if both trace to the same press release, you have one source. - If you can't corroborate, downgrade the claim to "contested" and say who asserts it. - Discard claims you cannot trace at all, however plausible. ### 5. Check coverage Re-read the sub-question list from step 1. Any question still unanswered becomes a new search round. **Stop when two consecutive rounds surface nothing new** — not when you feel finished, which happens several rounds too early. ### 6. Synthesize Write the brief. Lead with the answer, then the evidence: ```text Answer 2–4 sentences. The thing someone will quote in a meeting. Evidence Bulleted claims, each with an inline citation and a date. Contested What the sources disagree on, and who is on each side. Confidence High / medium / low, with the reason. Open What you could not establish, and what would settle it. ``` ## Guardrails - **No claim without a source.** "I think" is not research. - **Prefer primary sources**, and put a date on everything. Stale data is the most common failure mode in this domain and the least visible. - **Separate what the sources say from your inference.** Label the inference. A brief that blends the two is unusable by anyone who needs to check it. - **Count independent sources, not links.** Three articles from one press release is one source. - **Absence of evidence is a finding.** "No primary source states this, though it is widely repeated" is often the most valuable line in the document. ## Troubleshooting - **Every round returns the same ten pages.** Your queries are paraphrases. Switch *mode*, not wording — search for the mechanism, or for the failure. - **Confident answer, thin citations.** The model is filling gaps from memory. Re-prompt: "list every claim with no source; delete or mark each." - **Sources all post-date the thing you're asking about.** You're reading commentary. Find the announcement, the RFC, the filing, the commit. - **The brief is longer than the evidence.** Cut it. Length is not confidence. ## Cost note Done properly this is expensive: dozens of searches, many full-page reads, minutes of wall-clock. Use it for decisions that are hard to reverse. For "what is X" a single grounded query is the right tool, and reaching for this one is just an expensive way to be slow.

Goal

Turn one fuzzy question into a cited, fact-checked brief — by fanning out searches, reading sources, and verifying each claim before it lands.

Why not just ask

A single search-and-summarize pass has three predictable failure modes, and this workflow is built around them:

One phrasing, one slice of the web. Search engines reward the phrasing you used. Ask "is X better than Y" and you get the comparison posts; you never see the migration post-mortems that would actually answer you.
Consensus laundering. Twenty blogs citing the same original press release look like twenty sources. They are one source with nineteen echoes.
Silent staleness. The top result is the best-optimized page, not the most current one. In fast-moving domains the accepted answer is often eighteen months out of date.

Steps

1. Decompose

Break the question into 4–8 sub-questions that together cover it. Write down, in advance, what a complete answer must address — that list is what you check coverage against later, and writing it after you've read sources means you'll unconsciously scope it to what you found.

Include at least one sub-question that could falsify your expected answer.

2. Fan out

Run parallel searches, one angle per sub-question, and deliberately vary the mode of search rather than the words:

By entity — the product, company, or person by name
By mechanism — how the thing works, without naming it
By counter-argument — "problems with", "why we moved off", "post-mortem"
By time — restricted to the last six months, and separately to before the thing was popular
By venue — the primary docs, the standards body, the filings, the source

Each mode is blind to what the others surface. That's the point.

3. Read, don't skim

Open the promising sources. Extract each claim with the passage that supports it, plus URL and publication date. A claim without its supporting quote will be subtly reworded by the time it reaches the brief, and the rewording is where the errors enter.

Record the source type too: primary (docs, filings, source code, the announcement itself), secondary (reporting that cites primaries), or tertiary (aggregators, listicles, AI-written SEO pages). Tertiary sources support nothing on their own; they're leads.

4. Adversarially verify

For each load-bearing claim, ask: who disagrees, and why? Then go find them.

Corroborate with a second independent source — one that isn't citing the first. Check what each cites; if both trace to the same press release, you have one source.
If you can't corroborate, downgrade the claim to "contested" and say who asserts it.
Discard claims you cannot trace at all, however plausible.

5. Check coverage

Re-read the sub-question list from step 1. Any question still unanswered becomes a new search round. Stop when two consecutive rounds surface nothing new — not when you feel finished, which happens several rounds too early.

6. Synthesize

Write the brief. Lead with the answer, then the evidence:

Answer        2–4 sentences. The thing someone will quote in a meeting.
Evidence      Bulleted claims, each with an inline citation and a date.
Contested     What the sources disagree on, and who is on each side.
Confidence    High / medium / low, with the reason.
Open          What you could not establish, and what would settle it.

Guardrails

No claim without a source. "I think" is not research.
Prefer primary sources, and put a date on everything. Stale data is the most common failure mode in this domain and the least visible.
Separate what the sources say from your inference. Label the inference. A brief that blends the two is unusable by anyone who needs to check it.
Count independent sources, not links. Three articles from one press release is one source.
Absence of evidence is a finding. "No primary source states this, though it is widely repeated" is often the most valuable line in the document.

Troubleshooting

Every round returns the same ten pages. Your queries are paraphrases. Switch mode, not wording — search for the mechanism, or for the failure.
Confident answer, thin citations. The model is filling gaps from memory. Re-prompt: "list every claim with no source; delete or mark each."
Sources all post-date the thing you're asking about. You're reading commentary. Find the announcement, the RFC, the filing, the commit.
The brief is longer than the evidence. Cut it. Length is not confidence.

Cost note

Done properly this is expensive: dozens of searches, many full-page reads, minutes of wall-clock. Use it for decisions that are hard to reverse. For "what is X" a single grounded query is the right tool, and reaching for this one is just an expensive way to be slow.