The Claude-Powered Outbound Agent: How to Build an Entire SDR Team in One AI Project

This article is for marketers, founders, and lead gen specialists who have already tried LinkedIn outbound and realized one thing: the process is actually several different jobs. Finding an audience isn’t the same as writing a message. And writing a cold message is a completely different beast compared to responding to “tell me more”.

We’re going to show you how to assemble an AI team of four specialists within a single Claude Project – and ensure everyone stays in their lane.

The Problem: One Prompt Can’t Do It All

Most people trying AI for outbound make the same mistake: they open ChatGPT or Claude, type “write me a cold message for LinkedIn,” and get something generic, boring, and robotic.

The problem isn’t the AI. The problem is that effective outbound requires four distinct competencies:

When one prompt tries to juggle all four roles, you get a muddled mess. The solution: split these roles and give each its own specific context.

How It Works: 4 Modes in One Project

Claude’s Projects feature allows you to store instruction files and a system prompt in one workspace. We use this to keep all four “specialists” in one place.

Mode 1 – STRATEGIST: Finding Your “Who”

The Strategist answers the big question: Who should we write to, and why now?

This agent guides you through:

• ICP Definition: Industry, company size, job title, pain points, and desired outcomes.
• Tier-1 Signals: Hiring news for SDRs, recent funding rounds, or new executive leadership.
• Lead Scoring: Categorizing leads into Tier 1, Tier 2, or Tier 3.
• Quality Control: List criteria and account rotation logic.

A laser-focused list of 200 contacts with the right scoring will always beat a generic list of 2,000.

Mode 2 – ARCHITECT: The Campaign Skeleton

The Architect builds the framework: how many touchpoints, what type of messages, and what happens when someone doesn’t reply.

• Standard Sequence: 4 steps over 10–12 days.
• High-Priority (Tier 1) Sequence: 5 steps over 13 days.
• Rotation Logic: When to hand off a lead or stop attempts.
• Branching Logic: Handling replies vs. silence vs. opt-outs.

Step 1: Connection Request
  ↓ Accepted → Step 2
  ↓ Not accepted after N days → Rotate to a different account

Step 2: First Message (1 day after acceptance)
  ↓ Replied → [DIALOGUE]
  ↓ No reply for 3-4 days → Step 3

Step 3: Follow-up - New angle (Day 4-5)
  ↓ Replied → [DIALOGUE]
  ↓ No reply → Step 4

Step 4: Breakup Message (Day 9-10)
  ↓ Replied → [DIALOGUE]
  ↓ No reply → Close / Re-engage in 60 days

Mode 3 – COPYWRITER: Words That Get Replies

The Copywriter handles the prose: connection notes, initial outreach, and the “breakup” email.

We use three levels of personalization:

The Formula: Hook → Bridge → One CTA question. If it doesn’t fit on a phone screen without scrolling, it’s too long.

Cold outreach isn’t advertising. It’s the start of a conversation.

Mode 4 – DIALOGUE: Mastering the Inbox

Dialogue is where most people leave money on the table. Getting a reply isn’t the goal – the goal is to qualify the lead and book a meeting.

The Dialogue Agent:

• Classifies replies: HOT, WARM, NEUTRAL, OBJECTION, or REJECTION.
• Provides structured rebuttals for “already have a tool,” “no time,” or “send me info”.
• Knows when to push for a call and when to back off.

Metrics: How to Know if You’re Winning

Every project includes an outbound metrics file. Here are the benchmarks:

The One-Change Rule: Never change the segment, the copy, and the sending account at the same time. Otherwise, you’ll never know what actually worked.

Step-by-Step Setup

Step 1: Create a Claude Project

Go to Claude.ai → Projects → New Project. Name it something like “Outbound Team”.

Step 2: Upload Instruction Files

In the Project Knowledge section, upload these 6 files:

1. B1_strategist.md
2. B2_sequence_architect.md
3. B3_copywriter.md
4. B4_conversation.md
5. outbound_metrics.md
6. training_materials_template.md

⚠️ Important: The first five are universal methodology files. The sixth is yours – we’ll cover that nex

Step 3: Set the System Prompt

In Project Instructions, paste the master prompt that tells Claude what this team is, how to switch between roles and which rules to never break:

You are the Outbound Marketing Team — one Claude project with 4 specialist modes.
Channel: LinkedIn only. Approach: tool-independent (logic and decisions, never software click-by-click instructions).
Users: founders, marketers, and students at intermediate level. Adapt depth to context.
Full knowledge for each mode lives in the corresponding project file.

ACTIVATE MODE by user command:
  [STRATEGIST]   → B1_strategist.md       — ICP, signals, scoring, list criteria
  [ARCHITECT]    → B2_sequence_architect.md — sequence structure, steps, timing, rotation
  [COPYWRITER]   → B3_copywriter.md       — all message copy, personalisation, offer framing
  [DIALOGUE]     → B4_conversation.md     — live replies, objections, negotiation, qualification
  No command → ask which mode is needed before proceeding.

NON-NEGOTIABLE RULES:
Tool-independent — give logic and criteria; the user applies them in whatever tool they use.
Stay in active mode — no cross-mode work without an explicit handoff statement.
LinkedIn account safety is law — never recommend volume or behaviour that risks restriction.
Quality over volume — a clean, scored, relevant list beats a large dirty one every time.
Explicit handoffs — when mode work is done, state exactly what the next mode needs and why.
One clarifying question max — state your assumptions and proceed.
Always reply in the language the user writes in. The project files are in English — that does not affect your response language.

Metrics reference lives in: outbound_metrics.md

Step 4: Craft Your Training File

This is the most important step that separates a generic agent from your personal SDR. Let’s dive into this file.

How to Write Your Training File

The training materials file is a document that explains your context to the agent. Without it, the agent works on general principles but doesn’t know your product specifics, your clients, or your communication style.

What should be in the file:

• Product Profile:
  What it is, who it’s for, pricing.
  What you DO NOT do – so the agent doesn’t overpromise.
• Accounts / Sender Profiles:
  Who is writing on whose behalf and how it affects communication.
  Example: if the founder is writing, invite them to a demo with the founder.
• Pipeline Stages:
  Names of stages in your CRM and when to use each.
• Booking Links:
  Different Calendly links for different audiences and warm-up levels.
  When to provide a link and when it’s too soon.
• Lead Classification:
  HOT, WARM, NEUTRAL, SERVICE PROVIDER, REJECTION, NOT ICP –
  Signals and specific actions for each type.
• Inbox Workflow:
  Step-by-step: what to do after each type of response.
  Task timeframes.
• Communication Rules:
  Language, tone, message structure.
  Personalization for different lead types.
• Standard Templates:
  Ready-made phrases for repetitive situations.
• Competitors:
  Who they are, their strengths, and your specific advantage.
  Not just “we are better” – but exactly what and why.
• Mistakes to Avoid:
  Things that have cost you real leads in the past.

Tips:

• Don’t be vague. “We help businesses grow” is useless. “We help SaaS teams of 10+ people automate LinkedIn outbound without hiring an SDR” is useful.
• Add real objection examples. Include actual pushback you’ve received and the responses that worked.
• Update the file quarterly. The market changes – keep the agent current.

Next Level: Connecting Grinfi MCP

The configuration described above is already a full-fledged outbound agent. But there is a next level.

If you use Grinfi as your platform for LinkedIn + Email outreach, the agent can do more than just offer advice – it can perform actions directly.

What is MCP?

MCP (Model Context Protocol) is a standard that allows Claude to connect to external tools and perform real-time actions. Instead of “write what I should say,” it becomes “read the inbox, classify responses, send a message to this lead, and set a task”.

Grinfi has its own MCP server: mcp.grinfi.io

What the Agent can do with Grinfi:

With Grinfi MCP connected, the agent gains direct access to the platform without manual tab-switching:

Inbox:

• Open all unread conversations.
• View the full chat history for each lead (across all accounts).
• Classify a response, suggest text, and send it after your confirmation.

Contacts and Pipeline:

• Change a lead’s stage.
• Stop or resume automation for a specific contact.
• Create a task with the required timeframe and time zone.
• Mark a conversation as read.

Control:

• Check active automations.
• View account and mailbox statuses.
• Find a contact by name or company.

How to Connect

In Claude.ai, go to Settings → Integrations and add the MCP server: https://mcp.grinfi.io

Once connected, Claude will recognize the Grinfi tools and can use them in any conversation within the project.

About Action Confirmation

The agent does not act autonomously. Before every action – sending a message, changing a stage, creating a task – it shows you what it intends to do and waits for your confirmation. This is not an autopilot; it’s an assistant that removes manual work while keeping you in the driver’s seat.

What You Get as a Result

Basic Configuration (files + prompt, without MCP) – A Claude project that:

• Helps define the ICP and score the list before launch.
• Builds a sequence structure with rotation logic.
• Writes personalized messages at three levels of detail.
• Reacts to live responses: classifies, handles objections, and leads toward a call.
• Tracks metrics and suggests where to troubleshoot.

With Grinfi MCP Connected – All of the above, plus:

• Reads the inbox and displays unread messages directly in the chat.
• Sends messages after your confirmation.
• Changes stages, stops automations, and creates tasks – no manual tab-switching required.

This isn’t magic or “AI instead of humans”. It’s a structured tool for those who already understand outbound and want to remove the grind without losing the methodology.

Files for Download