vibe-prospecting

Este extracto es una copia guardada durante la revisión. La fuente externa contiene la versión completa y actual.

---
name: "vibe-prospecting"
description: "Find company & contact data. Turn your agent into a prospecting platform. Get contact information, roles, tech stack, business events, website changes, intent data. Build lead lists, research prospects, identify talent. 150M+ companies, 800M+ professionals, 50+ data sources."
metadata:
  version: "0.1.108"
---

# Vibe Prospecting

B2B prospecting — companies, contacts, enrichment, events.

## Platform

**Before any prospecting work, detect your host using the rules below, read exactly one platform guide, and follow it for auth, workflow, flags, and troubleshooting.**

### Detecting your runtime (mandatory, in order)

1. **Host identity in your system context** (check this first — overrides tool names):
   - **Claude Code** — context says you are Claude Code / Anthropic's CLI, or shows a CLI **Environment** block (cwd, shell, platform). → read [`claude-code.md`](platforms/claude-code.md) only. **Do not read `claude-chat.md` or `cowork.md`.** MCP tools in the list (`mcp__*__fetch-entities`) do **not** change this.
   - **OpenAI Codex** — context identifies Codex or `CODEX_SHELL=1`. → [`codex.md`](platforms/codex.md)
   - **OpenClaw** — context identifies OpenClaw or `OPENCLAW_CLI=1` / `OPENCLAW_SHELL`. → [`openclaw.md`](platforms/openclaw.md)
2. **Only if step 1 did not match** — no CLI host identity, no shell:
   - **Claude Cowork** — Cowork workspace, `.plugin` install, connector store. → [`cowork.md`](platforms/cowork.md)
   - **Claude Chat** — claude.ai or Claude desktop MCP connector (not Cowork). → [`claude-chat.md`](platforms/claude-chat.md)
3. **Fallback:** terminal, scripts, CI, or unknown host → [`other.md`](platforms/other.md)

**Never infer platform from MCP tool names alone** (`mcp__claude_ai_*`, `mcp__*__fetch-entities`, etc.). Tool prefixes are not a host signal when step 1 already identified Claude Code, Codex, or OpenClaw.

| Platform | Read now |
|----------|----------|
| **Claude Code** | [`claude-code.md`](platforms/claude-code.md) — `vpai` CLI |
| **OpenAI Codex** | [`codex.md`](platforms/codex.md) |
| **OpenClaw gateway** | [`openclaw.md`](platforms/openclaw.md) |
| **Claude Cowork** | [`cowork.md`](platforms/cowork.md) — MCP connector; **not** Claude Code |
| **Claude Chat** (claude.ai, Claude desktop) | [`claude-chat.md`](platforms/claude-chat.md) — MCP connector; **not** Cowork or Claude Code |
| **Other** (terminal, scripts, CI, generic hosts) | [`other.md`](platforms/other.md) |

**CLI hosts** use `vpai` — each platform guide is self-contained (Sample Gate, stdout/`csv_path`, CSV upload). **Chat/Cowork** follow each MCP tool's description and input schema only.

## Hard Rules

1. **`tool_reasoning` (or platform equivalent)** on every real call. Use the user's request verbatim. Reuse across the whole workflow. Skip only when inspecting a tool's input schema with no real execution.
2. **Chain via session + CSV, never paste IDs.** Each step returns **`session_id`** and **`csv_path`** (names may vary slightly by host). **Never invent `session_id`** — on the first step omit it and use the value MCP returns; on later steps copy it exactly from the prior tool JSON. Pass **`--csv-path`** (CLI) from the prior **`csv_path`**. Do not paste raw ID lists into tool arguments when the platform reads IDs from the source CSV. For **`fetch-entities`** prospects scoped to earlier companies, pass the prior business output's **`csv_path`** via **`--businesses-csv-path`** per your platform guide.
3. **You may edit the CSV between steps — keep the ID column.** On CLI hosts, read the file at **`csv_path`**, filter/dedupe/add columns, or rewrite rows, then pass the result to the next step via **`--csv-path`**. **Always retain the entity ID column:** **`business_id`** for business workflows, **`prospect_id`** for prospect workflows. Every row you want processed next must have a non-empty ID in that column. Do not rename or drop those headers — the CLI reads them verbatim for ID injecti