← Back to Blog

How to Teach Your AI Agent to Sell Like You: The Skills Behind One Replit Seller's Inbox

By Andrew Verrilli, Sales at Replit ·

Connecting an agent to your CRM is the easy part. Getting it to know what stage a deal is in, what to do next, and how to write a follow-up that sounds like you is where most teams stall. Andrew Verrilli sells at Replit, and he spent this session showing the custom skills he has layered onto his own agent to do exactly that. He walked through a stage-aware email skill, an inbox triage setup built on Superhuman, the reason conflicting skills wreck agents, and a live moment where he told the agent to send its drafts with no review. He also covered how he uses the agent to answer deep technical questions and to hand engineering a working pull request instead of a vague feature request. This recap covers the build, the guardrails, and the trust curve he recommends.



Why does AI still sound "a little off" when it writes for you?

Andrew's starting point is that humans are tuned to how they sound. A lot of us have been writing and talking for decades, far longer than AI has existed, so even a small mismatch in voice feels wrong. An agent can send an email without trouble and can roughly tell where a deal sits. That is not enough for most people to let it act unsupervised.

His title for the session, "teach your AI agent to sell," was shorthand for something narrower: teach it to sell like you. The goal is to reach the point where you feel good about turning the autonomy dial up, because a lot of people still watch everything their agent does and do not fully trust it. He called that fair, and spent the session on how to earn the trust gradually.

What is a skill, and why does it matter more than the agent you pick?

Andrew treats skills as the semantic layer for your agent. Plenty of companies are focused entirely on building a semantic layer for an organization, and he thinks you can get roughly 70% of the way there just by adding the right skills to your agent.

He builds his skills with AI, and he describes the process as iterative. He did not write the instructions from scratch, and some of the rules in his skills came from the agent telling him about things he did not know existed.

What does a stage-aware sales email skill do?

This is one of his favorites. The idea is that the agent needs to know where in the sales process a prospect sits before it writes anything, because the right reply depends on the stage.

Piece of the skill What it does
Deal-stage awareness Tells the agent where the prospect is, so tone and content match the stage
Product usage lookup Replit is a product-qualified-lead company with many users on the product. The skill knows to log into BigQuery, the company's data lake, and check how well the prospect already knows the platform
Sales knowledge Takeaways from sales books he likes, tagged by where they apply in the deal cycle. For example, Never Split the Difference applies mostly to getting a deal across the line
Reply style How he wants an email answered, plus a separate commercial Q&A skill

The usage lookup is the part he called a big unlock. Digging into product data used to mean going to the data team and asking for a pile of SQL queries. Now the agent pulls it instantly, and those specifics end up in the email, which is what makes it read as relevant.

Doesn't mixing several sales methodologies confuse the agent?

Not when each one is tied to a stage. Because the agent knows where the deal is, it knows which book's lessons to pull out. The methodologies stay out of each other's way because each affects a different part of the process.

Why do conflicting skills ruin an agent?

Andrew was blunt: this is how you ruin your agent ten times out of ten. It can happen inside a single skill with two contradictory instructions, but more often someone builds a skill, then another, then another, and eventually two active skills conflict and the output falls apart. He sees this a lot with customers who are heavily into AI and keep adding skills.

You still want multiple skills that work together, such as one for deal stage, one for how to respond to email, and one for commercial Q&A. The discipline is making sure they do not contradict one another.

His own example shows what a conflict looks like. His agent is connected to both Gmail and Superhuman, and each has its own idea of which emails are important. If the agent considers both, it has two sources of truth and will produce conflicts. So his reply triage skill says to consider only emails Superhuman marks as important, and explicitly not to substitute Gmail's own important label.

How does the email triage skill work?

Superhuman filters his inbox using its own AI. Out of a couple hundred or more emails, it showed five important ones in the demo. Because Superhuman exposes an MCP, he connects it to the agent he uses, which for him is Replit, and the agent acts on whatever Superhuman has already decided matters. In his words, what made Superhuman good for him as a human filtering his inbox is what makes it good for his agent.

The agent has several other connections: his calendar, Gong, Granola, BigQuery, and Gmail. It checks the calendar before sending anything. When he types "email triage," it works through the important emails in priority order and drafts a response to each.

Rules from the triage skill he showed:

  • Only consider emails marked important by Superhuman.
  • Check the latest actual messages in the thread before doing anything.
  • Treat a Superhuman reminder as a reminder, not as a new customer reply.
  • Prioritize by explicit deadlines and by commercial or customer consequences, so a thread about a commercial agreement goes first.
  • Refresh the latest messages before drafting, because a customer may write back while the drafts are being prepared.
  • Show drafts before sending (this is the default).

How does the agent decide when to follow up?

Superhuman can resurface a sent email as important if no reply comes back, and you set how long to wait. In the demo, the agent chose a three-day wait on its own. The reason: it was Friday, so it gave the prospect until Monday. If there is no response, Superhuman pops the email back into the important list, and the agent picks it up from there. That loop is why Superhuman's idea of "important" and the agent's list of things to act on are the same thing.

Can you actually trust an agent to send on its own?

Andrew did it live. After explaining that his triage skill normally shows drafts first, he told the agent to just send the emails, and it did. His point was not that review is unnecessary. It is that he got there over time.

His approach is crawl, walk, run:

  1. Crawl: The skill shows drafts before anything is sent. When he first adopted Superhuman, he ran it side by side with Gmail until it proved itself.
  2. Walk and run: As the agent keeps getting things right, you turn the autonomy up.

He is a skeptic by habit and said Superhuman earned his trust by doing the job well. He also said that sometimes he shows up to a meeting and wonders how it got booked, then finds his agent sent the prospect several emails. Usually, he says, those are good meetings.

What about the worry that the system filters out something important?

This is a fair concern when you cannot see what the filter dropped. His answer was practical. You fine-tune the filter early on, and the side-by-side period is how you catch misses before you rely on it. He also guessed that much of what a tool like Superhuman does behind the scenes is simpler pattern matching than people assume, and that you train it toward inbox zero over time.

Why connect a sales agent to the product's code base?

This is the part of the session that goes beyond email. Andrew's agent is also connected to Replit's code base. Many sellers would ask why they would need that, but he sells to technical buyers who ask specific questions about source control, or about where a query goes when it is handed to inference and which control plane handles it. With the code base connected, he can answer those questions quickly, even live on a call.

That does two things. It keeps momentum going in a deal, and it builds trust with the buyer, who feels comfortable emailing him anytime if he can answer anything. The answers also go deeper than the documentation. In a security review with a CSO, or with a CTO as the buyer, the docs will not contain what they need.

He still brings a sales engineer to big accounts as backup. For smaller deals where a sales engineer's time is not worth it, he can now take those calls himself.

Can a seller hand engineering a pull request instead of a feature request?

Andrew calls this his 10x edge. Doing everything, in his definition, includes drafting a pull request to change the product and fix something for a customer. Sales teams are the closest to customers, and they know what customers will pay for, so he argues sellers should be the start of the product development life cycle. That used to be hard. Now he can draft a PR and give it to engineering with the case that the idea is validated and customers are lined up.

A few details of how it works:

  • The PR he showed was never going to be merged because the diff was far too large. That is fine. He gives engineering the PR number, and they can run it locally or in Replit and see exactly what he means.
  • It removes the game of telephone between product and engineering. Instead of several meetings to explain a request, the working example is in front of them, and the only question is whether there is engineering time this week.
  • He summed it up as "PRs are the new PRD." No design file and no spec document are needed to start the conversation.
  • It works for non-technical people because Replit runs the environment in the cloud. Instead of local development, he does cloud development, in an isolated version of the same environment engineers use, so he can test changes as he builds them.
  • Most sellers on his team are new and still ramping, so not everyone does this, but some do. The practice is generally accepted. A merged seller PR is rare, but it happens more as the habit spreads.

Why build apps as well as agents?

Before agents, Andrew's way to "do everything" was building apps, and he still thinks they matter. One example was a vendor form-filling app: upload a form and it fills it out. A new seller needs no process knowledge, just the app. In the traditional version, a seller goes to finance with a list of questions. He still goes to finance for niche cases, but the menial tasks of every department should be handled by an app or agent the seller already has.

His view on where the return is highest: internal apps. Engineers writing more code with coding assistants is one thing. The bigger unlock is people who have never been able to code, and who sit close to specific business problems, building solutions instantly. He pairs that with a venture-style approach to ideas. Since it is so cheap to invest in ideas now, you would be foolish not to invest in as many as possible, and the same applies at a small company if you are the one generating them.

Is "drift" a model problem or a harness problem?

When people say agents drift, Andrew's take is that it is mostly a misalignment problem, and that agent harnesses have largely solved drift. A harness sits on top of the models. It can filter what goes to the agent, filter what comes back, double check a response and have it redo it, and pick the right frontier model for each task.

His example is the email triage run. Reading the inbox is an API call, so it may not need a model at all. Reading the emails, drafting responses, and applying the skills are three different jobs. He said he was assuming one model with a large, cheap context window handles the reading and another that is strong at writing handles the drafts, and he was clear that he did not know what runs for every step. The point stands: if you feel misalignment, you may be sending everything straight to a single model, including a coding model doing a writing job.

He also made the case for not doing AI on hard mode. If a tool picks models and splits up tasks for you, you do not need to learn a terminal to get the work done. In his words, we are going forward in time, not backward, and we should focus on getting the task done.

Key Takeaways for GTM Teams

  • Teach the agent where the deal is. A stage-aware skill that checks product usage data lets every email match the prospect's actual situation.
  • Tie each methodology to a stage. Sales books help the agent only if it knows which parts of the process they apply to.
  • Audit your skills for conflicts. Two skills telling the agent different things is the most common way to break it. Pick one source of truth, such as one definition of "important."
  • Build skills with the agent, and iterate. It will often tell you rules you did not know to write.
  • Earn trust in stages. Start with drafts shown before sending, run new tools side by side with the old ones, and raise autonomy as the agent proves itself.
  • Connect the agent to your product. Answers deeper than the docs build trust with technical buyers and free up sales engineers for bigger deals.
  • Hand engineering a working PR, not a request. Even an unmergeable pull request beats several meetings of explanation.
  • Put the right model on each task. If output feels off, check whether a harness is picking models, or whether one model is doing everything.

Want to Attend the Next Session?

GTM Clawllective is a community where go-to-market operators show how they are actually using AI in the field, live and unscripted. The live discussion stays off the record, and every recap includes the speaker's demo video so you can watch the build yourself. Come see the next one live and join the community.


About the Speaker

Andrew Verrilli works in sales at Replit, where he has been for about eighteen months. He has been a Replit user since 2019. Before moving into sales he was a teacher, and he taught through COVID. He sells to technical buyers and builds his own agents, skills, and apps to cover work that used to need a data team, a finance team, or a sales engineer.

Connect with Andrew on LinkedIn.

Watch the session replay: Learn from Replit: Teach Your AI Agent to Sell