GTMcraft Claude Signal: Your secrets are leaking in Claude, Sunday June 14
12 signals from 3 Claude experts. 3 problems. 1 playbook to steal. Implement it on Monday.
Save this. The weekly GTM Claude signal saves you hours of reading and helps you become better with Claude on Monday. It is the one piece worth your Sunday (or Monday) morning.
Send it to 1 founder or GTM operator who is still not mastering Claude for GTM and is struggling implementing any learned AI lessons into action today.
The 90-second version
90% of companies have staff pasting work into personal AI accounts. 57% have typed sensitive data in. On personal plans, Claude and ChatGPT train on it by default and can keep it 5 years.
The fix takes two minutes: turn off model training in settings, default to a temporary chat for work, and never connect a work account to a personal AI.
Building got cheap, so the question flipped from can you build it to should you. 42% of startups still die building something nobody wanted, and that share rises as building gets cheaper.
Most operators run agent-grade work through a chat box. Three surfaces now: Chat for answers, Cowork for file work, Code for builds and loops. Charlie Hills frames the same ladder as three automation levels: start low, climb only when you must.
This week’s build: stand up an AI data-hygiene baseline for your team in 4 weeks.
Read time: 5 min
Lane note: Saturday is GTM execution, Sunday is Claude. Read here the Saturday edition ↓
Here is the one move that matters most this week ↓
This week’s number 1 move
Lock down what your team feeds the AI before it costs you a customer or a lawsuit. Ruben Hassid pulled the numbers this week, and they are not comfortable. Staff at more than 90% of companies use personal chatbots for work, 57% have typed sensitive information in at least once, and 22% still reach for a personal account even when the company pays for a tool. On personal Free, Pro, and Max plans, Anthropic and OpenAI train on your chats by default and can keep them up to 5 years. Samsung banned ChatGPT after engineers leaked source code three times in 20 days, and in Europe pasting a customer name or email into a personal account can be an unlawful data transfer.
Start here: open Claude, go to Settings, then Privacy, and turn off “Help improve our AI models.” Two minutes. Do the same in ChatGPT under Data controls.
Source: Ruben Hassid, “Stop using your own Claude at work.” (Jun 10)
Hi, it is Koen Stam and welcome to GTMcraft: The Future GTM Operator. This newsletter is built from 100,000+ GTM signals collected from 100+ operators and founders, combined with 13+ years of my own lessons and failures from the trenches. I write at the intersection of go-to-market practice and AI-powered systems for founders scaling 0 to 10M+ ARR.
100,000+ GTM relevant signals from LinkedIn, Newsletters and Podcasts indexed. 100+ playbooks structured. 3 GTM operator jobs across 3 GTM motions (SMB, MM, ENT). Same recipe. Now all part of the GTMcraft Operator Room.
Consider joining? (send me a DM for access).
Problem 1: Your team is pasting company secrets into personal AI accounts
At 0 to 10M ARR you have no IT department policing this. The founder is the policy. Hassid’s numbers say the exposure is already live: most of your team is pasting work into personal accounts, and more than half have put sensitive data in. The defaults work against you, because on a personal plan the provider trains on those chats unless you turn it off, and retention runs years. The legal edge is public and growing. The US Defend Trade Secrets Act lets a company sue over leaked secrets, a secret stops being legally secret the moment you hand it to a third party under loose terms, and GDPR adds a second front in Europe.
Start here: turn off model training in every AI account you and your cofounder use, and confirm the toggle saved.
The rule: if you would not post it in the all-hands channel with your name on it, do not paste it into a personal AI. Anonymize it first, or use the company tool.
Source: Ruben Hassid, “Stop using your own Claude at work.” (Jun 10)
Problem 2: The bottleneck flipped from can you build it to should you
When building was expensive, can-you-build-it was the gate, and the people who could ship won. That gate is gone. Ruben Dominguez frames the new shape: the founder stops being the contributor who writes the code and runs the ops, and becomes the orchestrator of agents that carry the work out, with the edge moving from execution to judgment. Anthropic’s own Founder’s Playbook puts numbers on the speed, with validation cycles that took months now taking an afternoon. Here is the trap most people miss. The old CB Insights figure, 42% of startups dying because they built something nobody wanted, does not shrink when building gets cheap. It climbs, because the cost of building the wrong thing dropped too, so more wrong things get built.
Start here: take the last thing you shipped or were about to. Write one sentence on what evidence said someone wanted it before you built it. If you cannot, that is the gap.
The rule: no build starts without one line of evidence that someone wants it. If you cannot write that line, you are not ready to build it.
Source: Ruben Dominguez, “The one-person company stopped being a meme.” (Jun 11, free section)
Problem 3: You run agent-grade work through a chat box
Most operators do everything in one chat window, including work the chat window is the wrong tool for. Two creators map the way out this week. Ruben Dominguez lays out the three surfaces: Chat for a quick answer you copy out, Cowork for work on your own files where the AI reads a folder and writes the real spreadsheet back to your machine, and Code for builds and for loops that run without you. Charlie Hills draws the same ladder as three levels of automation: Level 1 is a Skill you fire from Chat on demand, Level 2 is that same Skill on a schedule in Cowork, Level 3 is Claude Code running unattended on Anthropic’s cloud with your laptop closed. His discipline is the part to steal: start at Level 1, climb only when the level you are on stops doing the job, and keep a human on the publish button at every level.
Start here: list your 3 most-used Claude tasks, label each one Chat, Cowork, or Code, and move the most painful one to its right surface this week.
The rule: match the work to the lowest surface that does the job. Quick answer is Chat, file work is Cowork, builds and loops are Code. Start one level lower than you think, and a human approves every publish.
Sources: Ruben Dominguez, “The one-person company stopped being a meme.” (Jun 11), Charlie Hills, “The 3 levels of Claude automation.” (Jun 7), Ruben Dominguez, “The head of Claude Code stopped prompting Claude.” (Jun 12, free section)
Also on the radar
Ruben Hassid, How to AI: connectors are the most dangerous thing on a personal AI account. A connected AI can be hijacked by the content it reads, so never connect a work account to a personal AI. Link
Ruben Dominguez, The AI Corner: Claude Code shipped a goal command (v2.1.139, May 11) that runs a loop until your condition is true, grading the work each turn. Most builders have not tried it once. Link
Ruben Dominguez, The AI Corner: Anthropic published a 36-page Founder’s Playbook, with validation cycles that took months now taking an afternoon. Link
Ruben Hassid, How to AI: vibecode a side project on the company’s AI and the company can own it. The line between work and personal AI is also an IP line. Link
Steal this move
Stand up an AI data-hygiene baseline in 4 weeks. For you if: founder or GTM operator, 0 to 10M+ ARR, lean team with no IT or security function, and staff running customer data, deal terms, or forecasts through AI accounts.
Week 1, Define: Inventory who uses which AI account, personal or company, and what data each one touches. List every connected app across the team.
Week 2, Build: Turn off model training on every account. Set the temporary chat as the default for work. Write the paste rule and the anonymize habit, and share both.
Week 3, Validate: Disconnect every work account from every personal AI. Spot-check the last week of usage against the paste rule and fix the gaps.
Week 4, Operate: Decide the company-tool question. Put a 15-minute connector and settings review on the calendar, monthly.
Why this matters now
The cost of AI dropped to near zero, so everyone is feeding it more: more of your pipeline, your customers, your deal terms. The input side is where the risk and the edge both moved.
This week one creator showed the risk, which is that your team is already pasting secrets into accounts that train on them. Two more showed the edge, which is that when building is cheap, the judgment about what to build and which surface to use becomes the scarce thing.
This builds on Software Got Cheap from June 7. Cheap building was the upside. This week is the bill: what you feed the machine, and whether you should have built the thing at all. It also extends Taste Is the Moat from May 24, now with a number on it, because as building gets cheaper the 42% who build the wrong thing grows rather than shrinks.
For a lean team at 0 to 10M+ ARR, the move is simple. Govern your inputs, sharpen your judgment, and match the work to the lowest surface that does the job. The typing is no longer the hard part.
See you inside GTMcraft’s Operator Room,
Koen
Your (human) GTM Agent
Full source list
Ruben Hassid, How to AI:
Ruben Dominguez, The AI Corner:
The one-person company stopped being a meme. (Jun 11) (free section used; full playbook behind the paywall, not reproduced)
The head of Claude Code stopped prompting Claude. (Jun 12) (free section used; full playbook behind the paywall, not reproduced)
Charlie Hills, MarTech AI: