A few weeks ago I posted four words that stopped people mid-scroll: Claude just learned to copy how you work. The idea is almost too simple. You record your screen once while you do a task by hand, Claude watches, and it turns the whole thing into a skill it can run again forever. I tested it the hour it came out. I recorded myself turning a sales call transcript into my recap email, the exact thing I do after every demo. Four hundred and fifty-seven steps captured. Then I handed it a transcript it had never seen, and it produced the same structure, the same tone, my exact process, without me touching it. I run Growth Cab, a GTM advisory, and that was the moment claude for sales stopped being a chatbot I talk to and became something that does my work while I do something else.
in
“Claude just learned to copy how you work. Record your screen while you do a task once. Claude watches and turns it into a skill it can run forever.”
What Claude for Sales Actually Looks Like Now
For two years, using AI for sales meant sitting in a chat window and writing a good prompt every single time. You described the task, you got an answer, and the next day you described it again from scratch. The new feature, tucked under Record a skill in the plus menu of the Claude desktop app, breaks that loop. Instead of explaining the job in words, you do it once with the screen recording on. Claude watches where you click, what you copy, how you decide, and compresses the whole demonstration into a reusable skill. My recap-email skill is not a prompt I rerun. It is a recording of how I actually write that email, frozen into something I can point at a new transcript and walk away from. It lives on the Pro, Max, and Team plans, so this is not an enterprise-only toy that never reaches the people actually doing the selling.
The Difference Between Prompting and Teaching
This is the part worth slowing down on, because it is the difference between prompting and teaching. A prompt is an instruction you give from memory. You have to remember every rule, every exception, every bit of formatting you care about, and type it out again each time. A recorded skill captures the parts of your process you would never think to write down. When I recorded the recap email, Claude picked up that I pull the next steps to the top, that I quote the prospect's own words back to them, that I keep the whole thing under two hundred words. I never told it any of that. It saw me do it. Teaching by demonstration reaches the tacit knowledge that lives in your hands, the stuff you do without narrating. That is why the output read like mine instead of like a generic assistant guessing at my voice.
Once I saw it work on the recap email, the list of things I wanted to record got long fast. The call-prep brief I build before every demo. The way I qualify an inbound lead before it reaches my calendar. The weekly account review where I decide who is worth a fresh outbound touch. Every one of those is a task I do the same way each time and have never managed to hand off cleanly, because writing the instructions down always lost something in the gap between how I think and how I type. Recording the task keeps the something. You are not documenting the process anymore, you are handing over a copy of yourself doing it.
Where Claude for Sales Still Breaks
I am not going to pretend this runs your whole funnel by Friday. The first place it breaks is judgment. A recorded skill copies your steps, and it will copy them even when the situation underneath has changed. My recap skill writes a clean, confident email whether the call went well or quietly fell apart, because it does not know the difference between those two calls. It reproduces the process, and the process was never the entire job. Reading whether a deal is actually alive is still on me, and no recording of my clicks can carry that.
The second place it breaks is the trap every automation shares, and it is worth saying out loud. A person writing one wrong follow-up is a bad morning. A skill firing four hundred wrong follow-ups because the input changed and nobody checked is a brand problem. So anything that touches a live deal at Growth Cab still has me on the approval step. The skill drafts, I read it, I send. Only the low-stakes slice, the internal notes and the first-draft prep, runs fully on its own. A recorded skill is unsupervised the moment you stop watching it, and unsupervised is a very different word from autonomous. Most of the disasters I have seen with sales automation came from people treating the two as the same thing.
How to Record Your First Claude for Sales Skill This Week
Do not try to record your whole job. Pick one narrow task you already do the same way every week and quietly resent. For most founders that is the post-call recap or the pre-call brief. Open the Claude desktop app, find Record a skill in the plus menu, and just do the task once the way you always do it, slowly and cleanly, with none of the corner-cutting you would not want copied. Then hand it a fresh input and watch the first three outputs like a hawk. You will find the gaps immediately, and each gap is a rule you can add in a sentence. Once it stops surprising you, let it run on the safe work and keep yourself on anything with revenue attached. The teams getting real leverage out of claude for sales are the ones who moved from doing the task to approving the output, and who stayed honest about which tasks are safe to let go and which ones still need a person reading every line.
If you want the workflows I have already recorded, with the exact task and the one rule that keeps each of them honest, I break down one GTM play a day in The Revenue AI Brief, my newsletter. The original post that kicked this off, with the full demo, is on my LinkedIn. Come tell me which task you would record first, or where you think I am wrong about keeping a person on the send.
