Prospect Locally
Articles

An agent skill for local lead generation

An agent that has the tools to run lead searches still has to decide how to use them. Left alone, it may run a dozen separate searches where one campaign would do, enrich a hundred domains in one call, or spend credits before you have agreed to anything. The agent skill file is a Markdown document that settles those choices in advance. This article describes what is in it and how to install it for Claude Code, Codex, or Cursor.

What the skill file is

The file lives at /local-prospects-skill.md and is about 10 KB. It is plain Markdown, written to be saved as CLAUDE.md, so an agent reads it as standing instructions for the project. The docs also show it in a viewer with a Copy skill button, for people who prefer to paste it.

The file does not contain your account data or a real key. Its examples use the placeholder key lp_example_key, and you supply the real key when you connect the MCP server or call the REST API.

The four rules it teaches

  1. Use a campaign for many areas. A request for leads across a state, a country, or a list of cities belongs in a campaign. A campaign is saved, resumable, and removes duplicate businesses across its locations, none of which a series of separate searches does.
  2. Use POST /search for one-off searches. A request for a handful of leads in one place, or a search where you want the answer in the same conversation, goes through the search endpoint. It starts at once and returns the businesses while enrichment continues in the background.
  3. Enrich at most 10 leads per call. The enrich endpoint and the enrich_domains tool accept 1 to 10 leads. The server charges the whole batch before it starts, so a batch larger than ten is rejected, and an agent must split a longer list into groups of ten.
  4. Ask before spending credits. Before a search, an enrichment, a campaign start, or a phone verification, the agent tells you the credit cost and waits for a yes. It never repeats a paid call that you did not ask for.

The file also lists the 19 MCP tools with their inputs, what each returns, and what each costs, so the agent does not have to guess a parameter name. It explains the lead fields in plain terms: the owner, the emails with their role-based and free-provider flags, the phones with their line type and carrier, the social profiles, and the website status.

Why the batch limit matters

The ten-lead limit is missed in a long conversation. An agent that is told to “enrich this list” may send the whole list in one request. The request fails with a 400 error, and some agents answer a 400 by retrying with changed parameters. The skill file tells the agent the limit up front, so it splits the list the first time. Each lead that fails to enrich is refunded at one credit and appears in the response under usage.leads_refunded, so a failed domain does not cost you.

Install it for Claude Code

  1. Download the file from the address above, or copy it from the docs with the Copy skill button.
  2. Save it as CLAUDE.md in the root of the project where you want lead searches to happen. If that project already has a CLAUDE.md, append the skill text to it.
  3. Start a fresh Claude Code session in that folder, so the file is loaded at the start.

Claude Code reads CLAUDE.md on its own, so no further setting is needed. The file does not connect the tools for you. You still add the MCP server with your API key, as the connection article describes.

Install it for Codex or Cursor

The file is written to be saved as CLAUDE.md, and the rules in it apply to any agent that can read Markdown. Codex and Cursor each have their own name for a project instruction file, so check which file your version reads. Two approaches work:

  • Save the file as CLAUDE.md and tell the agent in your first message to read and follow it.
  • Paste the text into the instruction file that the tool loads automatically.

Either way, the four rules reach the agent before it makes its first call. If your project already has long instructions, keep the skill text as its own section with a heading, so you can update or remove it later without touching the rest.

Check that it works

After you install the file, ask the agent for leads from three cities and watch what it does. A correct agent proposes a campaign, previews it, states the maximum credits, and asks before it starts. If it runs three separate searches, or starts spending without asking, the file was not loaded: confirm the file name and the folder, and start a fresh session.

The docs list every endpoint the skill file refers to, and the home page has setup prompts for ChatGPT, Claude, GoHighLevel, Cursor, and other agents if you would rather paste a prompt than keep a file.