Claude Code hype is louder than the revenue right now

A few months ago, something shifted in the GTM world and I have been watching it closely.

Everyone suddenly became an AI operator. My LinkedIn feed filled up with people who had never managed CRM infrastructure selling outbound automation systems.

People who had never maintained a production workflow launching AI SaaS products.

People who had been running a 3-person lead gen agency six months ago now positioning themselves as the authority on autonomous GTM.

The pattern is always the same.

A lead gen agency owner discovers Claude Code. Gets excited. Drops Clay. Burns through thousands in tokens in the first two weeks because nobody told them about direct API routing. Announces they are building a Clay replacement. A GTM operating system. The future of outbound.

Then the content machine kicks in. Skill files get open sourced on GitHub. A webinar gets scheduled. The new SaaS is built on Claude and Supabase with architecture that would make any senior engineer wince. A Lovable website goes live. A 30-day founder journey newsletter launches.

The original agency… the one that never actually hit a million, gets quietly shut down. A team offsite gets posted to show everyone they are doing things.

I have watched this exact arc play out at least a dozen times.

These are not businesses being built. This is content about building businesses. And the difference between those two things is expensive to learn firsthand.

Here is what I know from building and scaling four B2B companies and leading development teams for over a decade. The expensive part of software is never the first version.

The first version is the easy part.

Claude Code or Lovable or Cursor can generate a working prototype in an afternoon.

The real cost comes after that.

Maintaining reliability when your user base grows.

Handling the edge cases that only show up in production.

Debugging automations that failed.

Cleaning records that got corrupted because nobody built validation logic.

Managing infrastructure costs that compound every month.

Preventing technical debt from accumulating until the entire system needs to be rebuilt from scratch.

AI does not eliminate these problems. It makes the architectural decisions underneath them even more critical. Because when AI builds on a bad foundation, it builds faster, at greater scale, and the problems compound in ways that are harder to unwind than if a human had made the same mistakes slowly.

Meanwhile, the operators who are quietly generating real revenue with Claude Code are doing work that will never trend on LinkedIn.

Here is what that unglamorous work actually looks like.

Fix 1: Clean your CRM before you touch anything else.

Open your terminal.

cd ~/claude-projects
claude

Drop your CRM export into the folder and run:

Read contacts.csv and icp-context.md. Deduplicate
by email. Standardize job titles into consistent
categories. Flag personal email domains. Remove
any contact that does not match ICP criteria.
Export as contacts-cleaned.csv with a removal log.

This takes minutes. But most operators skip it because it is not exciting. Then they build enrichment pipelines on top of dirty data and blame Claude Code when the output is inconsistent.

If your CRM has duplicates, missing fields, and job titles spelled five different ways, nothing you build on top of it will work reliably. Clean the data first.

Fix 2: Stop routing bulk enrichment through MCP.

This is the single most expensive mistake I see. MCP wrappers consume significantly more tokens because they repeatedly fetch documentation on every call. For a quick filter or a Slack notification, MCP is perfect. For enriching 5,000 contacts through Apollo, use the direct API.

In your data-sources.md file, specify it clearly:

## Apollo
- Access: Direct API (not MCP)
- Use for: Bulk contact enrichment, company data
- Rate limit: 100 requests/min on standard plan
- Cost note: Direct API saves ~3-4x on tokens vs MCP

## Gmail
- Access: MCP
- Use for: Filtering already-contacted leads
- Not for: Bulk operations

## Stripe
- Access: MCP
- Use for: Filtering existing customers
- Not for: Bulk operations

## Slack
- Access: MCP
- Use for: Pipeline completion notifications
- Not for: Data operations

Claude Code reads this file and makes the right routing decision every session. You configure it once. The economics improve on every run after that.

Fix 3: Build your ICP scoring from closed-deal data, not assumptions.

If you have Gong or Fathom transcripts, drop them into your project folder and run:

Analyze every transcript in /transcripts. Identify
the top objection patterns, the language buyers use
to describe their pain, the signals that appeared
in deals that closed vs deals that stalled, and
the title/industry/company size patterns across
wins. Output an ICP scoring framework based on
this data.

Then take that output and update your icp-context.md file. Now every enrichment and scoring workflow you run is based on patterns from your actual revenue. Not a spreadsheet someone filled out during onboarding six months ago that nobody has looked at since.

This is the step that changes everything about the quality of your outbound. When your scoring criteria comes from real win/loss data, the leads that score highest actually convert. When your scoring criteria comes from assumptions, you are guessing at scale.

Fix 4: Ship one pipeline before you build a product.

claude --plan

Using icp-context.md, messaging-logic.md, and
data-sources.md, build a pipeline that:

1. Reads contacts-cleaned.csv
2. Enriches via Apollo direct API
3. Scores against ICP criteria
4. Filters via Gmail MCP and Stripe MCP
5. Generates 3-email sequences per Tier 1 lead
   using messaging-logic.md
6. Exports as sequences-ready.csv formatted
   for Instantly import
7. Pauses before final export for my review

Use Plan Mode. Show me the architecture before
building.

That is one pipeline. It runs end to end. It produces real output you can send to real prospects tomorrow. It does not require a SaaS product, a Lovable website, a GitHub repo, or a webinar.

Run it. Look at the output. Refine your .md files. Run it again. The second run is better than the first. The third is better than the second. That is the compounding loop and it is the entire game.

If you cannot get one pipeline running reliably on real leads before you start building a product, you are not ready to build a product.

Fix 5: Understand what you are actually paying for.

Claude Code charges by token. Here is a rough mental model:

→ A CRM cleanup on 1,000 contacts costs a few dollars in tokens

→ Bulk enrichment through MCP on the same list could cost 3-4x that amount

→ Running Plan Mode with Opus for architecture adds cost but saves you from rebuilding broken pipelines that wasted ten times more

→ Session resume means a rate limit at row 3,000 does not waste everything you already spent on the first 2,999 rows

If you’re burning through thousands in tokens without a clear outcome to show for it, start by mapping the workflow in Plan Mode before letting it build.

Review the architecture before you let it run.

That five-minute review can save hours of debugging and hundreds of dollars.

Here is what Claude Code can actually do for you:

→ Automate enrichment across thousands of contacts in minutes

→ Score leads against ICP criteria built from real closed-deal data

→ Filter out existing customers and already-contacted leads automatically

→ Generate persona-targeted sequences with merge variables ready for your sequencer

→ Clean, deduplicate, and normalize your CRM data with one prompt

→ Run pipelines that resume after interruptions without losing progress

→ Connect your entire stack through MCP and direct APIs

Here is what Claude Code cannot do for you:

→ Fix a broken offer that nobody wants

→ Compensate for ICP criteria you never defined

→ Replace the strategic thinking that makes outbound resonate

→ Maintain a production SaaS application as a side project

→ Turn an agency that never scaled into a software company overnight

Make bad data produce good output

→ Substitute for domain expertise you do not have

The world is moving. Claude Code is real. The leverage is real. But only if you build on a real foundation and resist the temptation to perform instead of produce.

I chose to move with it. I hope you do too.

The revenue is in the engineering.

Stay safe out there.

Much Love,

Benjamin Reed

Revyops & NextGen Founder

P.S. Check out ReyvOps at Revyops.com.

P.P.S. Want to help me on my journey to build RevyOps into a $1 Million SaaS? Please share this newsletter with your network so they can follow along. Anyone can sign up 100% free using this link: Click here

The Builder's Playbook is a series by Benjamin Aaron Reed on the tools, systems, and workflows that give founders and GTM operators an unfair advantage. New issues drop regularly.