Hypotheses, logging, and binary search

Once the AI proposes a cause, don't just accept it. Test it cheaply. The cheapest test is usually a log line.

Ask the AI to add temporary logging:

"Before we change anything, add a console.log (or print) that shows the value of user right before line 14, so we can see whether it's actually undefined."

Run it, paste the output back. Now you have evidence, not a theory. Maybe user is undefined — confirming the hypothesis. Maybe it's fine and the real culprit is one level up. Either way you've moved forward. Log generously while you investigate: the value of the variable, and a label saying which one it is and where the log sits. A bare console.log(user) that prints undefined is far less useful than console.log("user at validate.js:14 =", user). And remember to pull the temporary logs back out once the bug is fixed — leftover debug noise is its own small mess.

When the bug lives somewhere in a long process and you can't tell where, use binary search. Cut the suspect region in half: add a log at the midpoint.

The idea: each check throws away half of what's left, so a huge search space collapses fast. Here a bug hides somewhere between step 1 and step 8 — three checks corner it:

   step:  1  2  3  4  5  6  7  8      a bug is somewhere in here
          ░░░░░░░░░░░░░░░░░░░░░░

   check 1 ── log at 4 ── OK ──▶  bug is AFTER 4, drop 1–4
                                  ░░░░░░░░░░░  (5 6 7 8 left)

   check 2 ── log at 6 ── BAD ─▶  bug is AT/BEFORE 6, drop 7–8
                                  ░░░░░  (5 6 left)

   check 3 ── log at 5 ── OK ──▶  bug is at step 6   ✔ found

   8 steps → 3 checks.  1000 steps → about 10 checks.
``` Did the value look correct there? Then the bug is *after* it — search that half. Did it look wrong already? The bug is *before* — search that half. Each step halves the search space, so even a thousand-line path narrows down in about ten checks. Tell the AI explicitly: "Help me bisect this — where's the halfway point I should log?" The same technique works on history, not just code: if a feature worked last week and broke today, bisect your commits to find the exact change that introduced the bug.