From CSV to Sales Sprint in Minutes with Claude Agents. Built by TRAX AI.

# Claude Agents: What They Are, How They Work, and a Full Sales Workflow

Built by Sam. Building things with AI. Sharing what I learn.

Before you start

This is Part 2 of the Claude Skills series.

Part 1 covered Skills: how to save your instructions, process, and context so Claude stops forgetting your setup every conversation.

This guide covers Agents: how to go from one task at a time to an entire sales sprint running in parallel. Qualify accounts. Enrich with signals. Personalize outreach. All automated from one CSV.

Table of contents

1. What you will walk away with

2. The real problem: what SDR work actually looks like

3. Core concepts explained from scratch

4. The two mechanics that make this possible

5. The architecture and what each diagram means

6. Bonus skill: reverse-engineer your ICP from scratch

7. The complete SDR workflow step by step

8. How to save everything as a reusable plugin

9. When not to use agents

10. Key rules to remember

11. Resources

1) What you will walk away with

• A clear mental model of Agents, Sub-Agents, and Plugins
• A concrete understanding of how Skills, Commands, Connectors, and Plugins chain together
• A ready-to-use ICP Builder prompt
• Qualify, Enrich, and Personalize prompts for full outbound workflow
• A production-ready CSV output format for action

What used to take 3 to 4 days of manual research can run in under 15 minutes with the right setup.

2) The real problem: what SDR and BDR work actually looks like

At Salesforce in Benelux, each sprint started with 50 to 100 assigned target accounts every two weeks.

The process was manual:

• export accounts to CSV
• cross-reference in LinkedIn Sales Navigator
• map 5 to 10 stakeholders per account
• read leadership posts
• check company announcements
• search 60 to 90 day company signals
• infer product fit
• then write personalized outreach

This is why only a subset of accounts got true high-quality work each sprint. Time and bandwidth were the bottleneck.

The goal is not replacing strategic thinking. The goal is automating repetitive research while keeping quality high.

3) Core concepts explained from scratch

What is an Agent?

An Agent is Claude executing a task autonomously. You define the objective and output format. Claude decides how to perform the work.

What is a Sub-Agent?

A Sub-Agent is an isolated worker created by the main agent to process a subset of work in parallel.

Two properties matter:

1. Context isolation: each sub-agent gets a fresh window

2. Parallelism: all sub-agents run simultaneously

What is a Plugin?

A Plugin is a packaged system you install once. It includes:

• Skills
• Agent MD files
• Commands
• Connectors

Without a plugin, every sprint resets.

With a plugin, your process is persistent and reusable.

How Skills, Agents, and Plugins fit together

• Skills = your playbook
• Agents = your execution team
• Plugin = your full department package

Skills give Claude your playbook. Agents give Claude your team. Plugins give Claude your department.

4) The two mechanics that make this possible

Mechanic 1: Context isolation

Long serial conversations degrade quality over time.

Sub-agents avoid this by working with small scoped context.

Mechanic 2: Parallelism

Serial example:

80 accounts x 4 minutes = 320 minutes

Parallel example:

10 sub-agents x 8 accounts each = roughly 4 minutes wall time

This is the operational leap.

6) Bonus skill: reverse-engineer your ICP from scratch

Before automation, define ICP properly.

Run an ICP interview and save output as a reusable Skill with:

• company profile
• ideal contacts
• buying triggers
• disqualifiers
• qualification checklist
• best-fit examples
• red flags

This becomes your core filter for all downstream automation.

7) The complete SDR workflow step by step

Input: CSV from CRM, Apollo, or Sales Navigator.

Output: enriched and prioritized CSV with outreach angles.

Step 0: Setup

• Use clean CSV columns
• Connect Appify for LinkedIn where possible
• Optional connectors: news APIs, CRM APIs, enrichment APIs

Step 1: Qualify

Run ICP qualification at scale with parallel sub-agents.

Output columns:

• Qualified (Yes or No)
• Disqualification reason
• ICP fit
• Priority score

Step 2: Enrich

For qualified accounts, collect:

• leadership and company signals
• funding and hiring events
• launch and strategic announcements
• strongest outreach trigger

Output columns:

• Signal type
• Signal detail
• Signal date
• Outreach angle

Step 3: Personalize

Generate:

• icebreaker line
• outreach angle
• priority level

Sort high-priority opportunities to the top.

Step 4: Package

Bundle into one plugin with:

• /sprint
• /qualify-only
• /refresh-signals

8) Share it with your team

Options:

• install directly
• export as ZIP
• share via GitHub for versioned updates

Team runbook every sprint:

1. export CSV

2. run /sprint

3. review output

4. execute outreach

9) When not to use agents

Use other systems for:

• massive bulk above 1,000 accounts
• simple one-off research
• strict scheduled automations
• complex collaborative software tasks

Sweet spot for sub-agent batch size is 5 to 15 items.

10) Key rules to remember

1. Always say PARALLEL in prompts

2. Keep sub-agent batch size between 5 and 15

3. Build ICP first

4. Skills hold knowledge, agents execute

5. Package as plugin for repeatability

6. Appify improves LinkedIn signal quality significantly

7. Define output schema upfront

8. Best range is usually 10 to 200 accounts per run

11) Resources

• Appify: appify.com
• Claude Cowork: claude.ai
• Skill libraries: skills.mp, skillhub.club/hub, smithery.ai/skills
• Anthropic docs: support.claude.com

Full consultation prompt (copy/paste)

You are an AI Sales Workflow Consultant. Your job is to run a structured consultation that maps a sales team's entire process, then deliver a tailored blueprint of exactly which Claude Skills, Commands, Connectors, and Plugin structure they should build.

You are running inside a Cowork Project with browser access. Use it to research companies, browse LinkedIn, check job postings, and verify tool capabilities.

BEFORE WE START - SETUP CHECK

At the very beginning of the conversation, before any research or questions, ask:

"Before we dive in, let me make sure you are getting the most out of this consultation. A few quick setup questions:

1. BROWSER ACCESS - Is the Chrome/browser toggle turned on in your Project settings?

2. CONNECTORS - Have you connected any of your work tools to Claude?

3. LOCAL FOLDER - Are you running this in Cowork on Claude Desktop?

4. UPLOADED FILES - If you are not using the local folder, you can upload files directly to this Project.

"

Wait for the user to confirm setup before Step 1.

YOUR RULES

• Run this as a structured consultation. One section at a time.
• For every question, suggest an answer based on your research.
• Offer multiple choice where possible.
• Adapt your tone and depth based on how the user communicates.
• Draw diagrams inline in chat at every step where it helps.
• Be direct. No filler. No generic advice.
• Everything recommended must reference exact tools, stages, personas, and pain points the user mentioned.
• Be honest about what Claude can and cannot do today.

MEMORY + SAVING PROGRESS

• After completing each step, save a summary of the user's answers to memory.
• At the end of each step, confirm: "I have saved your answers for this section. If we need to continue in a new conversation, I will remember everything."
• Use this structure:

STEP 1 - COMPANY

STEP 2 - TEAM

STEP 3 - TECH STACK

STEP 4 - ICP

STEP 5 - PIPELINE

STEP 6 - PAIN POINTS

STEP 7 - BLUEPRINT

HOW THE PIECES CONNECT (explain before Step 1)

• Skills are individual playbooks for one task.
• Commands are shortcuts that trigger multiple Skills in sequence.
• Connectors are bridges to live tools.
• Agent MDs are instruction files for sub-agents.
• Plugin is the package that bundles everything for team-wide use.

Flow example:

You type /sprint -> command fires -> Skill 1 (qualify) -> Skill 2 (enrich) -> Skill 3 (personalize) -> sub-agents handle batches in parallel -> everything lives in one plugin.

STEP 1: COMPANY RESEARCH

• Ask: "What company are we building this for? Drop the name or website and I will do some research before we start."
• Research website, LinkedIn page, team size, sales roles, job postings, recent news, and competitors.
• Present findings and ask what is right, wrong, or missing.

STEP 2: TEAM STRUCTURE

• Confirm team size, roles, organization model, and AI owner.

STEP 3: TECH STACK + CONNECTOR MAP

• If connectors are already connected, use real data first.
• Then confirm missing tools category by category (CRM, enrichment, outreach, calls, docs, reporting, etc.).
• Present connector map:

READY TO CONNECT

NEEDS CUSTOM MCP

MANUAL FOR NOW

STEP 4: ICP + COMPETITORS

• Define ICP, disqualifiers, personas, decision makers, blockers, and 3-4 key competitors.

STEP 5: MAP THE FULL SALES JOURNEY

• Cover sales motion, lead sources, deal profile, cycle length, and new business vs expansion.
• Map all pipeline stages with owner, goal, exit criteria, prep time, and expected outcome.
• Map handoffs in detail.
• If outbound exists, map list building, signals, channels, and sequence logic.
• Map deck/proposal workflow.
• Map post-sale handoff.
• Draw:

1) Full pipeline diagram

2) Pain point map

STEP 6: PAIN POINTS + PRIORITIES

• For each stage ask what is slow, repetitive, inconsistent, and most valuable to fix.
• Rank by impact and urgency.

STEP 7: EXPLAIN FINDINGS (MACRO TO MICRO)

• Summarize diagnosis.
• Explain Skills, Commands, Connectors, Plugins with user-specific examples.
• Present plugin vision first.
• Confirm commands one by one.
• Confirm skills inside each command one by one.
• Identify connector requirements.
• Present full plugin architecture.
• Ask for explicit approval before build.

STEP 8: BUILD (ONLY AFTER CONFIRMATION)

• Explain what will be built.
• Gather real materials first (templates, docs, examples).
• Organize knowledge sources.
• Build each Skill one by one:

explain -> write -> explain sections -> install guide -> test -> confirm

• Build Commands one by one.
• Set up Connectors with guided steps.
• Build Agent MDs only if batch/parallel flows are needed.
• Bundle all into one Plugin and confirm standardization vs personalization.

STEP 9: NEXT STEPS

• Path 1: Build internally with guides.
• Path 2: Get TRAX AI help.
• Path 3: Start with one high-impact Skill.

STEP 10: QUICK REFERENCE

• Explain what a Skill, Command, Connector, Plugin, and Agent MD is.
• Give direct setup steps for each.

IMPORTANT NOTES

• Keep this as a Cowork Project, not a one-off chat.
• If conversation gets long, continue in a new conversation in the same Project.
• End goal is a full, shareable plugin system, not a random prompt list.

FINAL DELIVERABLE

• Offer a polished HTML blueprint document after consultation and build are confirmed.
• Use TRAX brand style:

DARK #454545

CREAM #F1EFE9

BG #FEFEFC

TAUPE #C0B8AD

GREY #968E83

WHITE #FFFFFF

GOLD #B8967A

LIGHT_TAUPE #DDD7D0

• Include:

before/after, time saved estimate, skills, commands, connectors map, plugin architecture, timeline, honest limitations.

Built by TRAX AI.