Build a Claude research workflow you can still inspect

A research agent that produces a polished message without showing its sources has moved the work around. You still have to check the person, the company and the claim before sending it.

The useful output is a short research brief you can inspect: why the person might fit, what actually happened, what remains uncertain and what a reasonable next step would be. Claude can help assemble that brief, compare it with your existing records and prepare the next piece of work.

Harry the Marketer's qualify, homework, send workflow supplies a practical starting sequence. His connected-tool tutorial adds the broader idea of retrieving context from the places where the business already works. Put those together carefully and you get a research process worth using beyond one clever DM.

First identify what Claude is actually connected to

“Claude is connected to LinkedIn” can describe several different arrangements. They do not expose the same data or permit the same actions.

The first two can overlap technically. MCP is a protocol for exposing tools and context, not a promise that every service has an integration or that every integration is trustworthy.

In Harry's LinkedIn connection demonstration, the mechanism is a Chrome extension operating an open browser tab. That is not a native LinkedIn API connection. An outreach workspace exposing an MCP tool is another distinct service, even when its records contain LinkedIn profile URLs.

Check the current Claude connector documentation and remote MCP guidance against the client you use. Plans, interfaces and supported tools change. Our Claude setup guide covers the connection process; this article concerns the work after authentication.

There is also a platform boundary. LinkedIn prohibits third-party software and extensions that automate activity on its website. A demonstration of browser control does not settle permission to use it. You can apply the research method to authorized CRM records, selected source files and manually gathered public evidence without handing an agent the controls of a logged-in LinkedIn session.

Give the research a job and a stopping point

Do not begin with “find leads and start outreach.” That combines several decisions while leaving the criteria undefined.

For an illustrative campaign, suppose an agency helps specialist consultancies improve project handoffs. The brief could say:

Review these ten existing prospect records. Look for current founders or operations owners at specialist consultancies whose public material indicates a possible handoff problem. Exclude existing customers, prior refusals and firms outside our delivery region. Produce a sourced fit assessment and questions to check. Do not import records, edit the CRM or contact anyone.

Now the agent has a defined batch, a purpose, exclusions and a deliverable. It can also report that none qualifies. That is a legitimate result.

The service description should explain what you can deliver and what would make a poor fit. Include current offer scope, relevant proof and any limits on capacity. Otherwise the model can find plausible ways to describe almost anyone as a prospect.

Start with enough records to inspect personally. A ten-record task is useful when you are still testing the rule; a thousand-record task mostly makes errors harder to see. The number is an operating choice, not a platform allowance.

Keep facts, interpretations and missing information separate

Harry's homework step asks for three recent observations about the person or company. That can prevent empty personalization, provided each observation has a source and a reason to matter.

Ask Claude for a record like this:

• Fact: what the source actually says, with URL, date and matched identity.
• Possible relevance: how that fact could relate to your service.
• Unknown: what the source cannot establish.
• Question: what you would ask before treating the hypothesis as a need.

For the consultancy example, the public fact might be a founder discussing work arriving without a clear project owner. The possible relevance is a handoff process. Unknowns include frequency, business impact and whether a fix is already underway. A reasonable question asks how ownership is assigned now.

That record is stronger than “growing consultancy needs operational help.” The latter sounds decisive while concealing several assumptions.

Require the agent to flag stale roles, conflicting company names and inaccessible sources. A search snippet or an old biography can help locate evidence; it should not quietly become proof of someone's current position. Ask for fewer good observations rather than padding a required total with trivia.

Here is a reusable research instruction, written for this workflow rather than taken from a source template:

For each record, show evidence for fit and evidence against it. Cite the source for every recent event. Label inference explicitly. Use “unverified” where you cannot establish a fact. Recommend proceed, research further or exclude, with a reason. Do not draft a claim that assumes an unverified problem exists.

Human review still matters. A model can cite a real page and misunderstand what it means.

Put approved client knowledge in a separate workspace

Research becomes more useful when it can refer to the company's actual relationships and promises. Harry's agency AI-tool discussion describes a client knowledge workspace containing briefs, notes, brand guidance and decisions. This is a more substantial idea than giving the agent every app in the business.

Keep client context separated. An approved brief, a dated decision log, relevant meeting notes, the current offer and the responsible people can be enough. Add correspondence only when it is necessary and you are entitled to use it. A meeting recording or transcript needs appropriate consent and handling; it is not automatically reusable training material.

Make the current decision easy to find. If an early call suggested a weekly newsletter but a later approved brief specifies fortnightly delivery, the agent should cite the change rather than combine both into a new promise. Record the date, owner and source of the approval.

This helps handoffs between copy, design, development and account management. Someone who missed the call can retrieve the actual decision instead of asking the client to explain it again. It also gives a human reviewer a place to check whether a proposed action contradicts an agreement.

If a client has not approved a claim or direction, label it as a draft. Do not let a plausible AI summary quietly become the new source of truth.

Draft from the brief, then review the words

Only after the research is checked should Claude prepare an opener, call brief or suggested response.

For a first message, ask it to connect one relevant observation to a proportionate question. Give it representative writing samples from the actual sender and a factual description of the service. You do not need an entire mailbox to establish a clear voice.

Harry asks for both a text message and a voice-note script, then chooses one. That is useful when the format decision remains with the sender. It is not a reason to send both to every prospect or infer a person's preferences from their photo.

Review the draft against the source, not just against how natural it sounds. Did “mentioned an issue” become “struggling with a serious problem”? Did “new product” become “rapidly scaling”? Did a general company event become a claim about the individual's priorities? Repair those changes before considering style.

When the next step is a meeting, ask for a preparation note: confirmed company context, previous conversation, open commitments, plausible questions and missing information. Harry includes this in his sales-workflow video. Preparation can improve the conversation, but it does not replace asking the prospect about their situation. A public profile does not reveal an entire business.

Triage the inbox without throwing away ambiguity

Inbox review is another useful application in Harry's three-pillar Claude workflow. The agent can group threads and suggest where a person should spend time.

Avoid a crude “hot or worthless” classification. An existing customer with a delivery issue may need attention before a new sales enquiry. An interested prospect saying “after our launch” needs a timing note. A refusal needs to stop solicitation. A message that is hard to classify belongs in a review queue.

Ask for the thread, a short evidence-based explanation and a proposed next action. Preserve the conversation date and owner. Verify that the connected inbox actually includes the relevant threads: a campaign tool may show only conversations created by that tool, not all personal account messages.

The agent should not decide that an ambiguous reply authorizes the next sales step. Read the exchange before making a commitment or moving someone to another channel.

Approve an action you can describe precisely

“Go ahead” is vague if the agent has accumulated several proposed changes. An actionable approval names the operation: update these records, save this reviewed draft, assign these threads to this owner or perform another specifically authorized task.

For an external communication, review the sender, recipients, exact text and context. For a CRM change, review the fields and records. If the tool cannot show the change before applying it, reduce the batch or keep that step manual.

Connected documents, web pages and emails remain source material. A sentence inside them telling an agent to ignore its rules or send information elsewhere is not an instruction from you. Give the agent access appropriate to the job and treat unexpected requests in retrieved content as something to flag.

After an authorized action, verify what happened. A draft is not a sent message. A successful sign-in is not a successful import. A tool timeout is not proof that nothing changed. Record the actual result, any uncertainty and the next owner so a colleague does not repeat the operation blindly.

The operating loop is practical: retrieve and cite, draft or analyze, verify, take the authorized action, then record the change. If the workflow cannot show you those stages, simplify it until it can. The point of connecting tools is to reduce repeated work while keeping the business decision understandable.