Research to report: a human sign-off workflow for Claude

Short answer: let Claude gather and draft, then make the human-owned steps explicit. Set the data scope before the run, ask for a claim table (each statement tied to a source and a supporting passage), have a person open every source, rewrite from the verified claims only, and keep a named approval step before the report leaves your hands. This is a researcher recommendation assembled from documentation. It has not been run, and the sources reviewed do not document a Research feature inside Claude Duet.

Status: documentation review, not hands-on tested. Sample prompts are illustrative, not observed output. Labels: Vendor documentation, Practitioner description (a creator’s own words, unverified) and Researcher recommendation.

What the documentation establishes

Vendor documentation: the Research feature. The Help Center says research runs “multiple searches that build on each other” and decides what to look into next. It is available on paid plans (Pro, Max, Team or Enterprise) on the web, Claude Desktop or Claude Mobile, and “you must have web search turned on”. Two lines matter for governance: “During research, Claude can use tools from your connectors automatically, without asking for approval each time”, and when research starts in Claude Mobile it “only searches the web and doesn’t use your connectors”, with no “error or other sign” that connectors were skipped.

Vendor documentation: what Anthropic checked in its own system. Anthropic’s engineering write-up (published 2025-06-13, describing an internal system, not a promise about any product) scored outputs on factual accuracy, citation accuracy (“do the cited sources match the claims?”), completeness and source quality. It also reports that “human testers noticed that our early agents consistently chose SEO-optimized content farms over authoritative but less highly-ranked sources”, and that “human evaluation catches what automation misses”.

Vendor documentation: citations in the API. The citations feature guarantees “valid pointers to the provided documents”. A valid pointer is not proof the passage supports the claim. That distinction is a researcher inference, but Anthropic’s own rubric above scores citation accuracy as a separate dimension from the existence of a citation.

Vendor documentation: Duet. The announcement says Duet reads, drafts and prepares work on its own, and “before it sends a message, makes a payment, signs a document, or deletes anything, it shows you exactly what it’s about to do and waits for your approval”. It also warns that “it still might misread a thread or miscalibrate what’s urgent”.

Practitioner description. A legal-education channel’s video description states “The model can suggest. The lawyer signs.”, says “Sourcing reduces fabrication risk. It does not eliminate it.”, and includes a step that “confirms the proposition the case is being cited for is the proposition the case actually stands for”. That is legal citation practice, read from a description; we did not watch it or retrieve a transcript, and applying it to general reports is our inference.

Prerequisites

• A written question, the decision the report supports, the audience, and a date range.
• A list of connected accounts you are willing to expose to the run.
• A named approver who is not only the person who requested the research.
• A folder where you save a copy or archive of every cited page, since web pages change (researcher recommendation).

The workflow

1. Frame the question. State the decision, scope, exclusions and what counts as a primary source for this topic. The content-farm finding above is the reason to say so up front.

2. Set the data scope. Because research can use connectors without per-call approval, turn off connectors the question does not need. If you use mobile, assume connectors were skipped and run it on web or desktop when internal data matters.

3. Run the research and request a claim table. Illustrative prompt, not observed output:

Research [question]. Return the findings as a table with columns: claim, source URL, source type (primary, secondary, vendor, forum), publication date, the exact passage that supports the claim, and your confidence. Mark any claim you could not tie to a passage as UNSUPPORTED.

4. Verify claim by claim. A person opens each URL, finds the passage and sets a status:

Check numbers, names, dates and direct quotes against the original, not the summary. Note the publication date and whether the source is primary.

5. Rewrite from verified claims. Keep supported claims; qualify partly supported ones; move unsupported ones to an “open questions” list. Do not ask Claude to “fix” the report by itself, because the point of the exercise is independent checking.

6. Hold the approval gate. A named person signs off on the final text and the recipient list. If a Duet-style agent will send it, the documented approval step shows what it is about to do; it does not tell you the contents were verified, so your claim table is the evidence behind the sign-off (researcher inference).

7. Log it. Record run date, scope, reviewer, saved sources, and changes made.

Validation and rejection checklist

Accept the report when every figure and quotation is Supported or explicitly qualified, every URL opens, the connector scope matches what you intended, and the approver has read the claim table.

Reject or send back when any numeric claim is Unsupported or Not found, a source is a content farm where a primary source exists, the run used connectors you did not intend, or nobody opened the sources.

Failure cases

• Real URL, wrong claim. The citation points to a page that says something adjacent.
• Weak sources win. Anthropic’s own testers saw SEO content ranked over academic or primary material.
• Silent connector skip on mobile. The report lacks internal data and, per the Help Center, nothing warns you.
• Unreviewed private data. Connector content can inform a draft without a per-call prompt; check what ended up in the text before sharing.
• Rubber-stamp approvals. Clicking approve on an action is not verifying a claim.
• Point-in-time results. A saved report goes stale; record the run date.
• API builders: the citations documentation says citations cannot be combined with structured outputs.

Compatibility: Duet versus the Claude ecosystem