The Revenue Leadership Podcast · 22 Jan 2026 · From the week of 19 January
E60: The Claude Code Era: Mastering Autonomous GTM Agents (Jordan Crawford, Founder @ Blueprint)
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Jordan Crawford, founder of Blueprint GTM, joins host Kyle Norton to explain Claude Code and how it differs from ChatGPT and Clay. They cover how Claude Code works with local files, how project folders let context build up over time, and how Jordan runs much of his business through it. Jordan argues that revenue leaders should break roles into atomic tasks an agent can run, and that they need hands-on familiarity with these tools to decide what to automate.
For founders
- Claude Code can work on local files and write code to process data too large for a chat window, such as 100,000 transcripts totalling hundreds of megabytes.
- Give each problem its own local folder and save analysis as documents so context compounds across sessions.
- Start with one small, well-defined task connected to a single system, beginning with read access.
- Jordan keeps AI writes to production systems such as the CRM off the table and designs his work around structured outputs.
- Prompts built in Claude Code can be deployed in Clay for editing and observability, with Claude Code used to build and test them.
For revenue leaders
- Break tasks down to the smallest unit an agent can do, then run that unit across every record.
- Sort team tasks by the time they consume and by whether an agent can perform them, leaving judgment calls such as tone analysis with people.
- A four-BDR pilot aimed at preparing reps for calls produced 85% more calls and 85% more opportunities than baseline, according to the episode.
- Agent job descriptions cannot yet be written reliably, so start by extracting tasks from what each person actually does.
- A CRO should understand both how the team spends its time and what the tools can do in order to find automatable work.
What was said 23, most useful first
Jordan sorts GTM AI into three eras: ChatGPT for word-based tasks, Clay for deterministic workflows, and Claude Code for autonomous work. Listen
He described ChatGPT as good for word-based and data questions, such as categorizing a CSV. Clay workflows run fixed steps with non-deterministic model calls inside them, such as an agent that decides whether a deal is a good fit. Claude Code is the third era, where a goal is given and the agent works on it continuously.
“The clay era was more about workflow automation, and these are deterministic workflows.”
Recording yourself narrating decisions and testing Claude against manual rows is a way to pass on your context. Listen
Jordan suggested recording a solo Zoom where you explain each decision and its reason. You then do about 20 rows manually, have Claude run the same task, and let it ask questions where it is unsure. He said you repeat this testing until the results are reliable.
“you can just work just on testing that context over and over again”
A pilot with four BDRs produced 85% more calls and 85% more opportunities than their baseline. Listen
Kyle said the team ran a pilot last week to better prepare reps for calls and reduce the time between them. The task was fully handled by an agent, and he said it was then scaled across every customer in the database. He said the work took two weeks and could roll out to every BDR once enablement was done.
“this pilot group made 85 % more calls and produced 85 % more opportunities than their baseline”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
Jordan does not let AI write to production systems like the CRM, favouring read access and outputs that go into other tools. Listen
He said he does not trust the tool enough to vibe code writes to his CRM. He cited Jason Lemkin's production database being deleted after someone gave the wrong access token. He said his business is about deploying, not destroying, so he designs his work around structured outputs.
“I don't trust the tool enough to like vibe code and like write to your CRM”
Pointing Claude at call transcripts helped Kyle document decision principles he had not yet written down. Listen
Kyle had written some decision-making principles and connected Claude to Notion through MCP. He asked it to read his call transcripts and find principles he was using but had not documented. Claude surfaced meetings where he disagreed strongly with the proposal under discussion. Jordan then suggested testing Claude's independent choices against Kyle's real choices.
“Here are 10 meetings where I disagreed vehemently”
Listen to the episode Leadership & culture Link to this Report a problem
Only about 2% of negative reviews in a set of about 200,000 contained useful material for the problem Jordan was testing. Listen
Jordan said he had an opinion about negative Google reviews and had Claude process around 200,000 of them locally. He said a chat tool could not take in that volume of data. He said the keyword search found only a small share of reviews that were useful for the problem.
“only in 2 % of cases are there any good, valuable, useful things in the negative reviews for this problem”
Claude Code can pull files in and out of its context window and decide how to process large data. Listen
Jordan contrasted Claude Code with ChatGPT and Gemini, saying Claude Code decides whether to use sub-agents, its context window or code. His example was searching 100,000 transcripts for competitor mentions and pulling the five paragraphs around each one. He said such transcripts are too large to house anywhere except locally, though zipped uploads let ChatGPT write code over more files.
“it can pull things in and out of its context window”
Clay workflows do not know the user's context, so that context has to be written into each prompt or table. Listen
Jordan said a Clay workflow that runs when a HubSpot deal arrives has no understanding of the user's contacts or goals. To improve a prompt he previously had to copy it into a chat tool, fix it, and paste it back into Clay.
“those workflows don't know me. They don't know what I'm trying to do.”
Keeping each campaign in its own local folder lets context accumulate and stay relevant to that task. Listen
Jordan's campaigns each have a folder holding only the transcripts for that campaign, so historical transcripts that do not apply are excluded. He said this makes the context get better over time. Claude Code's ability to work in one folder is what makes this possible.
“that context gets better and better and better and better.”
Saving analysis as documents means later sessions read summaries rather than re-reading source transcripts. Listen
Jordan had Claude read a set of transcripts and write a document of the reasons he wins and loses. The next time Claude Code opens, it reads that analysis instead of the raw files. He described this as turning unstructured information into structured context.
“it doesn't have to go read all of those transcripts again”
Jordan suggests joining closed-won and closed-lost data to transcripts using a shared key such as the Salesforce ID. Listen
For closed-won and closed-lost analysis, which he said he does not do often, he described what he would do: two CSVs, one per outcome, plus the transcripts for each, all linked by a join key. He said the user does not need to do this setup manually, since Claude can do it. He also said that structuring this context is important so that Claude can connect the pieces.
“there's some join key between them so that Claude can connect it up some Salesforce ID or whatever”
A client's 47-step campaign launch process was stored as a markdown file so Claude follows it each time. Listen
Jordan said a client wrote out the steps to launch a campaign and he gave them to Claude as a markdown file. Each time Claude starts a campaign task, it loads these steps as the plan. Changes are made in the code repository and pulled down from GitHub.
“every time we launch a campaign we should do these 47 steps”
Plan mode breaks a complex task into sub-tasks, and each runs as an independent agent that reports to a coordinating agent. Listen
Jordan described plan mode as producing one to two page documents that list all the steps of a complex task. Sub-task agents then carry out those steps and bring results back to the main agent. He framed the main agent as the coordinator of the work.
“and then deploy those, you know, you could call them chats or just independent agents”
Claude Code can run a process, check the output against your context, and revise it in a loop without you. Listen
Jordan described Claude Code configuring Clay agents, running them, and receiving the results back. It then judges how they did against the context it holds, finds errors such as too many steps, and reruns. He said the user can leave this loop running while doing something else.
“you have this loop that can happen without, you know, go off and have a sandwich”
Agents work best on the smallest indivisible unit of a task, which can then be run across every record. Listen
Kyle compared this to a prime number that cannot be divided further. Once an agent does that unit, he said, you have unlimited effort and unlimited supply of that job. He described going piece by piece through the revenue function at this level before scaling, and Jordan agreed that defining the most leveraged atomic task is key.
“you want to go down to like the smallest unit possible of that work”
CROs should categorize team tasks by time consumed and by whether an agent can perform them. Listen
He said a future CRO needs to understand the number of tasks in the team and have categorized them by time suck and agent capability. He described time suck as the inverse of value. He said this is how to find the work to remove from people.
“I have categorized those tasks Based on time suck and agent capability”
Judgment calls such as tone analysis are not something Jordan gives to AI, while deterministic steps can be broken out for it. Listen
He gave the example of deciding whether a prospect's tone is right to ask for a credit card, which he said AI should not do. He said the tools lack the capability for tone analysis. He said agents do well when tasks are split into deterministic choices and when they handle non-deterministic inputs such as a website.
“tone analysis is like something that, you know, the tools don't have the capability to do”
Job descriptions for agents cannot be written yet because their capabilities are unknown and keep changing. Listen
He said job descriptions describe units of work a human can do, and that agent capabilities make big leaps about every six months. He said that is why leaders must know what their people do and what the tools can do. He said that is the way to extract tasks from the team.
“we're not writing jobs for agents yet”
Listen to the episode Hiring & team building Link to this Report a problem
Jordan suggested a CRO should see how each person spends their day, perhaps by having reps record their screens. Listen
He said a CRO who understands what the team does, and also knows Claude Code, can use the two together to find which tasks to automate. He described this as the way to pull out the tasks that take the most time and that an agent can do. He presented this as a hypothetical.
“why don't you record your screen all day and just send me the recording”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
After building 12 enrichment prompts in Claude Code, Jordan advised a customer to deploy them in Clay. Listen
He said Claude Code was a better tool for building and testing the original prompts. He said Clay is the better place to run them for ongoing use because it is editable, pushes to the user's systems, and gives observability through its interface.
“the best place for you to deploy them is in clay”
Kyle is building an AI chief of staff that reads meeting transcripts to surface decisions and action items. Listen
Kyle said one skill pulls out action items assigned to his team and writes them to a Notion page he monitors. His goal is to scan five hours of meeting summaries with decisions in 20 minutes, and his vision is to be tagged in the meetings most relevant to him. He also built a Slack bot to bring the chief of staff into Slack.
“instead of going to five hours of meetings”
Listen to the episode Leadership & culture Link to this Report a problem
Building your own tools gives a team effectively an engineering department for any problem. Listen
Jordan said the value compounds as you do it yourself and then teach the organization to build its own tools. He called having an engineering department for any problem a wild way to think about it, and Kyle replied that it is a different way to see the world.
“You now essentially have an engineering department for any problem.”
To install Claude Code, search for its bash install script, paste the command into a terminal, then type claude. Listen
Jordan said the install takes about three steps and may prompt a developer tools installer, which the user should accept. He said the user can then change directory into a new folder by dragging it into the terminal. He advised starting really small, such as asking it to research a spreadsheet of your customers on the web.
“if you Google Claude code bash install script, that's the exact Google result”