Account Research and Outbound From Claude Code
The workflow people describe for running account research and outbound from Claude Code, where it breaks, and why your own site data is the best place to start.
5 min read
A common pattern has emerged for running outbound from Claude Code, the version of Claude that works in a terminal. The published walkthroughs differ in detail and mostly agree on shape. This guide lays out that shape, then adds two things most of them leave out: where it goes wrong, and what to use as the starting signal.
You do not need to be a developer to follow it. The same steps work in the regular Claude app with connectors, with less automation.
The Workflow
- Write down who you sell to. A short file describing your ideal customer: industry, size, the problem you solve, and who you do not want. Claude Code reads a file called
CLAUDE.mdat the start of every session, which is where most people put it. - Start from a signal, not a list. Pick accounts because something happened. More on this below.
- Research before writing. Have Claude gather what the company does, what changed recently, and who is likely involved. Keep this as a short brief you can read.
- Check the brief. Read it. This is the step people skip.
- Draft around the reason. One email that refers to something true and specific.
- Send from your normal tool. Claude prepares the draft. A person sends it, from the sequencer or inbox you already use.
Start From Your Own Data
Most published versions of this workflow start from a bought list or a data provider. That works, and everyone else can buy the same list.
The accounts already visiting your site are a better start. A company that came back three times this week and read your pricing page has told you something no list can. If your analytics can show activity by company and your assistant can read it, step two becomes a question: which target accounts were active this week, and what did they look at?
Where It Breaks
- Made-up details. An assistant will sometimes state something about a company that is not true. Ask it to include its sources, and open them.
- Too much output. Claude Code warns when a tool returns a very large result and caps it. Ask for the top ten accounts, not all of them.
- Untrusted content. If a tool fetches pages from the web, text on those pages can try to steer the assistant. The Claude Code documentation warns about this directly. Keep research tools and write tools separate where you can.
- Sending too early. Automating the send on day one is the fastest way to email the wrong person something wrong.
Rules Worth Keeping
Outbound email is covered by law in most places, and personal data about contacts is too. Know which rules apply where your recipients are before you scale anything, and keep a person in the loop on what goes out.
Setting It Up
Claude Code connects to a server with one command and a browser sign-in. For Leapio it is:
claude mcp add --transport http leapio https://api.leapio.com/mcpThen run /mcp and approve the sign-in. In the Claude app, add Leapio under Settings, then Connectors.
How Leapio Fits
Leapio supplies the starting signal and the account history: which companies were active, what they read, and what they are worth. Its own agents can also do steps three and five for you, researching a company, scoring the fit, and drafting outreach. Drafts are only drafts. Nothing is sent on your behalf. See Signals and Agents.