How Do You Know If You're Supervising Claude Enough?

Ask most people whether they check Claude's work carefully enough, and they answer with a feeling: careful, or not careful enough. This video argues that's the wrong measurement. Supervision should scale with how much decision-making you actually hand over on a given task, not stay flat as a personal habit, and the only way to know if it does is to log it.

The problem with treating supervision as one habit

Most people treat being careful with AI as a single, flat trait: you either double-check things or you don't, and that habit follows you everywhere. But not every task hands Claude the same amount of trust. A quick copy-paste edit and a real decision made mostly by the model are not the same kind of interaction, and a flat habit can't tell them apart. The mismatch between how much you're trusting Claude and how long you're actually checking stays invisible until something forces it into view.

Log, classify, audit

The fix isn't trying harder to be careful. It's building a small tool that does three things: log each interaction as it happens, classify how much decision-making was actually handed to Claude (a quick edit, an assisted task, or a real decision made mostly by the model), and once a week, audit for the gap, specifically the high-autonomy interactions that got almost no verification time.

The anchor: twelve interactions on a grid

The worked example plots twelve real interactions from one week on a grid, how much you relied on Claude across the bottom, how long you actually spent checking up the side. Nine of them land near the diagonal, where usage and verification roughly match. Three sit in a danger corner: heavy reliance on the model, almost no time spent checking. Those three don't just get counted, they get named, and each one gets a suggested next step matched to what it actually was, a quick code check for a script nobody reviewed, a second opinion for a decision that leaned hard on Claude's judgment. The log turns from a report into a to-do list.

What a rising score does and doesn't prove

Run the audit again two weeks later and the danger corner should empty out. That's real progress, the gap closing. But the video is careful about what this does and doesn't prove: a rising score only shows you logged more checking time against those interactions, not that the checking itself caught anything. And an interaction that stays outside the danger corner isn't automatically fine, it just means this pass didn't flag it.

Why this matters

Supervision isn't a switch you either have on or off. It's a gap between how much trust you're extending and how much verification you're actually doing, and that gap is invisible until you measure it directly, task by task, instead of trusting your own sense of how careful you've been.

Key takeaways

• Supervision should track how much decision-making you hand to Claude on each task, not stay constant as a personal habit.
• A simple log-classify-audit loop turns a vague feeling of carefulness into a measurable pattern.
• Plotting reliance against verification time on a grid exposes a specific "danger corner": high trust, low checking.
• A shrinking danger corner over time is real progress, but it proves you checked more, not that the checking caught real problems.
• Anything outside the danger corner still isn't guaranteed safe, it just wasn't flagged by this pass.

Who this is for

Anyone who uses Claude regularly for work and has only a gut feeling about whether they're checking its output enough, especially people handing over increasingly high-stakes tasks without a clear sense of whether their review habits have kept up.