convez.ai
AgenteMarketing & VendasIntermediário
sosa2601

Agente de Outreach GTM VysionAI

sosa2601/Outreach-Agent

Pipeline de outreach B2B de 5 etapas automatizado em Claude Code

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent automatizado que executa pipeline completo de outreach B2B em Claude Code: identifica sinais de intent de compra, pesquisa empresas contra seu ICP, localiza decision makers e gera sequências de email personalizadas. Transforma um prompt em CSV pronto para Smartlead ou PlusVibe sem dependências externas.

Instalação

$npx convez-ai sosa2601/Outreach-Agent

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar listas de prospecção qualificadas a partir de sinais de intent
  • Pesquisar e pontuar empresas contra critérios ICP customizados
  • Localizar decision makers com contexto de personalização
  • Criar sequências de email 3-touch personalizadas por contato

Veredito

Agent robusto para times GTM que automatiza o trabalho pesado de pesquisa e qualificação. Requer customização inicial dos templates (ICP e email framework) e acesso a Claude Pro/Max, mas elimina dependência de ferramentas caras como Apollo. Implementação sólida do workflow, embora dependa da qualidade dos dados públicos para enriquecer contatos.

README

A tradução para o português deste README ainda está na fila.

VysionAI GTM Outreach Agent

Signal-based outreach pipeline built entirely in Claude Code.
Find companies → research ICP fit → find decision makers → write personalised sequences.
One command. One output folder.


What It Does

This system runs a 5-stage outreach pipeline inside Claude Code:

Signal Finding → Company Research → Contact Finding → Email Writing → OUTREACH-READY.csv

You give it a signal prompt. It finds companies showing buying intent, scores them against your ICP, identifies the right decision maker at each company, writes a personalised 3-email sequence per contact, and compiles everything into a single CSV ready to load into Smartlead or PlusVibe.

No Apollo subscription doing the research. No Zapier. No n8n for the research layer.

One command:

@agents/gtm-outreach-agent.md "B2B SaaS companies hiring Head of Growth in the US"

Requirements

  • Claude Code (v2.0+) — install guide
  • Claude Pro or Max subscription (or Anthropic API key)
  • VS Code (recommended) or any terminal
  • Node.js (for Claude Code CLI)
  • Python 3.8+ (for scripts — optional at start, needed for scale)

Quick Start

1. Clone the repo

git clone https://github.com/yourusername/vysionai-gtm-agent
cd vysionai-gtm-agent

2. Edit your templates ← do this before anything else

templates/icp-criteria.md      ← your ICP: industry, size, stage, tech stack
templates/email-framework.md   ← your proof points, tone rules, CTA format

3. Start Claude Code

claude

4. Test with one company first

@skills/gtm-qualify/SKILL.md https://somecompany.com

5. Run the full pipeline

@agents/gtm-outreach-agent.md "your signal prompt here"

6. Check outputs

outputs/OUTREACH-READY.csv       ← load this into Smartlead or PlusVibe
outputs/sequences/               ← individual email sequences per company

Folder Structure

vysionai-gtm-agent/
│
├── CLAUDE.md                          ← loads every session automatically
│
├── agents/
│   └── gtm-outreach-agent.md          ← main orchestrator, runs all 4 stages
│
├── skills/
│   ├── signal-finder/
│   │   └── SKILL.md                   ← finds companies from job posts, funding, LinkedIn
│   ├── company-research/
│   │   └── SKILL.md                   ← 19-dimension GTM analysis + ICP scoring
│   ├── contact-finder/
│   │   └── SKILL.md                   ← finds decision maker + personalisation anchors
│   ├── email-writer/
│   │   └── SKILL.md                   ← writes personalised 3-email sequence per contact
│   └── gtm-qualify/
│       └── SKILL.md                   ← standalone quick ICP check on one company
│
├── scripts/
│   ├── company_scorer.py              ← ICP scoring at scale (100+ companies)
│   ├── contact_enricher.py            ← contact data processing
│   └── csv_builder.py                 ← compiles all outputs into OUTREACH-READY.csv
│
├── templates/
│   ├── icp-criteria.md                ← ⚠️ edit this first — your ICP definition
│   ├── email-framework.md             ← ⚠️ edit this first — your email rules + tone
│   └── signal-sources.md             ← signal search sources and query templates
│
└── outputs/                           ← all results save here automatically
    ├── SIGNALS-FOUND.md
    ├── COMPANIES-SCORED.csv
    ├── CONTACTS-FOUND.csv
    ├── OUTREACH-READY.csv
    └── sequences/
        └── {company-name}-sequence.md

How It Works

File Types Explained

FileWhat It IsAuto-loads?
CLAUDE.mdYour context file — who you are, ICP, stack, rules✅ Every session
agents/*.mdOrchestrator — manages the full pipeline❌ Invoke with @agents/
skills/*/SKILL.mdTask instructions — one skill = one job❌ Invoke with @skills/
scripts/*.pyPython code — data processing at scale❌ Claude runs when needed
templates/*.mdReference files — ICP, email rules, sources❌ Claude loads mid-task
outputs/Results folder✅ Auto-populated by skills

Skills vs Agents vs Scripts

A skill does one specific task. One input, one output, invoke it standalone.

An agent manages multiple skills in sequence. It launches signal-finder, waits for the output, passes it to company-research, and so on — all the way to the final CSV.

A script handles data processing at scale — scoring 200 rows, calling an API, merging CSV files. Claude Code writes and runs these when a skill needs them. You don't write scripts manually.

The Pipeline

Stage 1 → signal-finder skill
          Searches job boards, funding news, LinkedIn content
          Output: outputs/SIGNALS-FOUND.md

Stage 2 → company-research skill
          Fetches each company website, 19-dimension GTM analysis, ICP scoring
          Output: outputs/COMPANIES-SCORED.csv (A/B/C/Skip grades)

Stage 3 → contact-finder skill
          Finds decision maker at each A/B grade company
          Identifies personalisation anchors (recent posts, job changes, funding)
          Output: outputs/CONTACTS-FOUND.csv

Stage 4 → email-writer skill
          Writes 3-email sequence per contact using their anchor
          Applies rules from templates/email-framework.md
          Output: outputs/sequences/{company}-sequence.md

Stage 5 → csv_builder.py
          Merges all data into one send-ready file
          Output: outputs/OUTREACH-READY.csv

Signal Types Supported

The signal-finder skill searches across four signal types:

Signal TypeWhat It FindsQuality
Hiring signalsCompanies posting GTM/RevOps/Growth roles⭐⭐⭐⭐⭐
Funding signalsSeries A/B raises in last 90 days⭐⭐⭐⭐
Content signalsFounders posting about GTM pain on LinkedIn⭐⭐⭐⭐⭐
Tech stack signalsJob posts revealing automation gaps⭐⭐⭐

Example signal prompts:

"B2B SaaS companies hiring Head of Growth in the US"
"GTM tools startups that raised Series A in last 90 days"
"RevOps SaaS companies posting Head of Growth roles"
"Founders writing about building outbound from scratch"

ICP Scoring

Every company is scored 0–100 across five dimensions:

DimensionWeightWhat It Measures
ICP Firmographic Match25%Industry, size, stage, geography
GTM Pain Visibility20%Evidence of the problem we solve
Budget Signals20%Funding, tech stack, hiring
Tech Stack Fit20%HubSpot, Apollo, modern tools
Outreach Readiness15%Can we find the right person?

Grade bands: A (80–100) → hot. B (60–79) → warm. C (40–59) → nurture. Skip (<40) → pass.

Only Grade A and B companies proceed to contact finding and email writing.


Email Output Format

For each Grade A/B company, the email-writer produces a sequence file:

outputs/sequences/company-name-sequence.md
├── Subject line options (2)
├── Email 1 — The Opener (60–80 words, signal-based first line, yes/no CTA)
├── Email 2 — The Value Add (50–70 words, new insight, softer ask)
└── Email 3 — The Breakup (30–40 words, warm close, no pressure)

All emails are copy-paste ready. No placeholders. Every Email 1 opens with the specific personalisation anchor found for that contact.


Customisation

Edit Your ICP

Open templates/icp-criteria.md and update:

  • Target industries and company sizes
  • Funding stage preferences
  • Tech stack signals (which tools indicate fit)
  • Scoring weights per dimension
  • Buyer persona profiles

Edit Your Email Voice

Open templates/email-framework.md and update:

  • Your proof points and outcomes you deliver
  • Tone rules and words to avoid
  • CTA format preferences
  • Anchor quality standards

Add New Signal Sources

Open templates/signal-sources.md and add:

  • New job boards to search
  • Industry-specific funding newsletters
  • LinkedIn search queries that work for your ICP

Standalone Skill Commands

Use individual skills without running the full pipeline:

# Quick ICP check on one company (fastest test)
@skills/gtm-qualify/SKILL.md https://company.com

# Find signals only
@skills/signal-finder/SKILL.md "B2B SaaS hiring RevOps"

# Research a specific company
@skills/company-research/SKILL.md https://company.com

# Find contacts at a company
@skills/contact-finder/SKILL.md https://company.com

# Write email sequence (reads from CONTACTS-FOUND.csv)
@skills/email-writer/SKILL.md

Output Files

FileContentsUse For
SIGNALS-FOUND.mdRaw list of companies found with signal type and sourceReview before committing to full research
COMPANIES-SCORED.csvAll researched companies with ICP scores, GTM gap, tech stackUnderstand why each company was prioritised
CONTACTS-FOUND.csvDecision makers with personalisation anchors and email patternsInput for email writing
OUTREACH-READY.csvAll contacts + Email 1 copy in one fileLoad directly into Smartlead or PlusVibe
sequences/*.mdFull 3-email sequence per companyReview before sending, or load into sequences manually

Limitations

This pipeline handles the research and writing layer only.

It does not:

  • Auto-send emails
  • Auto-load contacts into your CRM
  • Run on a schedule without you present

For scheduled, unattended automation: connect OUTREACH-READY.csv to an n8n workflow that loads contacts into Smartlead and triggers the sequence. Claude Code builds. n8n runs.


Stack

Built with Claude Code only — no external tools required for the research pipeline.

ToolRole
Claude CodeResearch, scoring, writing, orchestration
Python (optional)Data processing at scale (100+ companies)
n8n (optional)Scheduled sending after pipeline runs
Smartlead / PlusVibeEmail delivery (load OUTREACH-READY.csv)

About

Built by Saswati Gorai — Growth AI Strategist at VysionAI.

VysionAI builds AI automation and GTM systems for B2B SaaS companies.


Licence

MIT — fork it, adapt it, build on it. If you share what you build, a mention is appreciated.

Ver o original no GitHub

Bom para

Relacionados