Created from a single voice note with Agent Craft
13% versus 89%. That's the gap Anthropic just published, and today…
13% versus 89%. That's the gap Anthropic just published, and today they stopped asking your permission. They acted on it on the 14th. If you're on Pro, Max or Team, Claude Code no longer stops to ask before running a command. A classifier checks instead.
1/7Here's the study behind it. A thousand professional developers, mid-session, get a dangerous command slipped in. Who catches it? The developers caught it 13% of the time. The classifier caught it 89% of the time. And people got worse the longer they worked. The classifier never changed.
2/7If you don't write code, this is the part where you tune out. Don't. Because this story isn't really about developers. It's about the little box that pops up and asks if you're sure.
3/7AI that reads your inbox. Drafts your reply. Books the thing. Fills in the form. Every one of those asks for permission. Every one is betting you're reading it. That 13% is the first hard number on how good that bet is.
4/7Think of the last cookie banner you clicked through. The last terms you didn't read. Nobody does. It's the same reflex. And now it's guarding the systems you spend real money on.
5/7Anthropic didn't just remove a safety feature today. They showed it was doing almost nothing to begin with. We're not careless. Click-to-approve just looked like oversight while catching 13%.
6/7The uncomfortable part is what replaced it. Another model, tested by the company that sells it. So when the box stops asking and the classifier starts deciding, who's checking the classifier?
7/7
More content from Agent Craft
- TikTokAnthropic removed permission prompts from Claude Code, and the study behind the decision is eye-opening. 13% is a number worth sitting with. #claudeai #anthropic #aiagents #aisafety #llm #agentcraft
- X (Twitter)After running five companies, I learned marketing can't carry the whole load alone. Companies that move fast hand the mic to executives who actually know the work. Building: AgentCraft (dot) ai Grow your personal brand in minutes per day with voice notes
- LinkedInSay "churn prediction model" in your first engagement and watch the fear cross your client's face. It's a hard sell. And there are two primary reasons. When they bring you on as an external resource, they want this quarter's problem solved. Maybe next quarter's. Rarely is it something 24 or 36 months into the future. The moment you talk about allocating resources that far out, their fear kicks in. Pulling anything away from now means diluted results in the short term to them. The second reason is a serious one to overcome. I've seen CEO's avoid it and actively burn money because of it. If you've spent any time in corporate, you know that keeping churn down, making retention sticky, is a function of the whole business. It's a marketing problem and a product problem and a support problem all at once. You're going to pull in other departments. You're going to ruffle some feathers. Sometimes the cost of those measures is significant, in both time and money. A CEO who doesn't want to rock the ship will be opposed to this, even against his better interests. So keep an eye on the future. But be careful about when you raise churn reduction, or predictive models around churn, with a new client. Highlighting it means you've now identified a problem. And once you've named it, you're on the hook to address it. The easiest way in is to scope out next year's budget for them and look at how they're allocating. The best companies allocate 50% of budget to this years acquisitions, 30% to year 2 revenue protection and uplift and 20% to year 3. That conversation opens the door without setting off the alarm. Even if they negotiate back to 90/10, that's a win Found this useful? Repost it to your network and follow for more Juan Mouton, CMO, Preparing you for Series A Personal Brand Powered by Agent Craft
- X (Twitter)ROAS looks great on a dashboard. But if the cohort you acquired stops paying at month 8, what did that number actually buy you? Juan Mouton, Fractional CMO preparing you for Series A Personal Brand Powered by Agent Craft
- ThreadsA client using Agent Craft for their personal brand told us something we didn't expect. Looking at all their posts together, they genuinely can't tell which ones they wrote and which the AI wrote. And that gap keeps closing. Brand voice in, content that sounds like you out.
- LinkedInRead the first line of any post and ask: could you have predicted the second? If yes, you're reading filler. Most LinkedIn posts fail that test, and it takes about two seconds to run. "Thrilled to announce I've joined..." tells you nothing you couldn't guess. The professional equivalent of clearing your throat. Here's a sharper way to judge whether a post earns attention. Forget the polish and the formatting for a second. Does it say something the writer would only say if it were true? Specifics are hard to fake. "We rewrote one email and trial-to-paid moved" carries weight because someone measured it. Vague uplift language carries none. Does it commit to a claim, or hedge every sentence into mush? "It can be argued that content may sometimes benefit from" is a person saying nothing while sounding busy. Does the rhythm move? Real writing has short sentences next to long ones. Uniform paragraphs, every one the same tidy length, that's the shape of something generated to fill space rather than to make a point. Predictable line two, hollow claims, dead-flat rhythm. Any one of them and you can skip the post. Someone wrote it to be seen, not to say anything. So run it on your own last post. Read line one. Could you have predicted line two? If you can, rewrite it until you can't.