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 Code assumes you have your debugging infrastructure in place. It will try to solve problems with whatever information you hand it.
This sounds reasonable until you think about what it means in practice: if your debugging setup is incomplete, you're feeding incomplete information into a very capable reasoning engine, and getting back confident, well-explained answers that are pointing at the wrong thing entirely. The analysis is sound. The inputs are wrong. The conclusion is confidently, systematically off.
I discovered this during a Firebase debugging session that was going nowhere. I was feeding Claude Code error logs and behavioral descriptions. It was generating hypotheses and producing fixes. Nothing worked. Not because the reasoning was bad — because my logging had gaps I didn't know about, my Firebase emulator wasn't capturing what I thought it was capturing, and the "evidence" I was working from was incomplete in ways that weren't visible to either of us.
When I stopped and asked a different question — "What debugging setup should I have in place before we try to diagnose this?" — everything changed. Claude Code produced a specific list: logging points I was missing, emulator configuration I hadn't done, diagnostic queries I should run against the database before drawing any conclusions about behavior.
Fifteen minutes of instrumentation later, I had real data. The actual debugging session took a fraction of the time the earlier thrashing had taken.
The mental model that fixed it: treat Claude Code like a sharp consultant who's parachuting in to help. A good consultant's first question isn't "what's broken?" It's "what data do we have?" If the answer is "not much," they'll tell you what to go collect before they start analyzing.
The question that unlocks it:
"Before we debug this, what logging, tooling, and diagnostic setup should I have in place to give you the best information?"
Ask it once at the start of any serious debugging session. It'll tell you exactly what to instrument. Then debug from a position of actual visibility instead of informed guessing.
Human-to-Claude calibration setting #6: If Claude's answers aren't landing, check your inputs before you question its reasoning.
This is the broader principle. Claude reasons from what you give it. If the results feel off — if fixes aren't working, if hypotheses don't match what you're seeing — the first question to ask yourself isn't "is Claude wrong?" It's "am I giving Claude complete information?" Usually the answer is no. Fix the inputs first.
We thought we could see what was happening. We could not. Turns out, looking properly is a step.
Aphilaty is a privacy-first community coordination app. aphilaty.com