Series: Weekend at Claude's — misadventures in building a production app with an AI anyone could mistake for the person who's going to make the whole thing happen
Claude gives confident answers. That is mostly a feature.
When a capable person gives you a clear, well-explained answer, you believe them. That's rational — they're usually right. Claude works the same way. The explanations are thorough. The reasoning is coherent. The answer arrives with complete assurance.
The thing to understand is that this is equally true whether Claude is right or wrong.
I've been building Aphilaty with Claude Code for months. In a long debugging session, "I found the issue" means Claude has identified what it believes to be the most probable cause based on available evidence. It does not mean the issue is definitely found. After the eighth time it announces this in a session where the issue remains stubbornly present, you develop a more nuanced relationship with the phrase.
This isn't a criticism. An AI assistant that hedged every answer with uncertainty disclaimers would be exhausting and useless. Confidence is the right default. The user just needs to understand what it actually means.
Here's how I think about it now:
Human-to-Claude calibration setting #2: "I found the issue" is a strong signal — verify before you move on.
"I found the issue" means Claude has identified the most probable cause based on available evidence. It does not mean confirmed. Try the fix. See it work. Then commit and move on.
Human-to-Claude calibration setting #3: A longer explanation doesn't mean higher certainty.
Sometimes Claude is elaborating a wrong hypothesis very thoroughly. The detail signals engagement, not correctness. When the explanation gets longer but the fix still isn't working, that's your cue to ask for alternative hypotheses rather than more analysis of the current one.
Human-to-Claude calibration setting #4: Ask for alternative hypotheses explicitly.
When a fix doesn't work, ask: "If that wasn't the cause, what are the next two most likely explanations?" This breaks the single-theory loop. Claude won't volunteer this reframe on its own — it will keep working the current hypothesis until you redirect it.
Human-to-Claude calibration setting #5: Your domain knowledge is the check Claude doesn't have.
Claude knows the tools, the patterns, the libraries. You know your codebase, what changed just before the bug appeared, and what "feels wrong" in a way you can't always articulate. Both inputs matter. Claude is working correctly when it applies its knowledge. You're working correctly when you apply yours alongside it.
The point of all of this isn't skepticism. Claude Code has saved me hundreds of hours. The calibration is about becoming a better partner to it — understanding what its signals actually mean so you can use them more effectively. It's like learning to read any new instrument. Once you understand the readings, you use it with a lot more confidence.
Bernie always seemed like he knew what was going on too. The trick was knowing when to verify.
Aphilaty is a privacy-first community coordination app. aphilaty.com