Debugging a Production API Integration with Claude Code and ICM

3h (edited) • 🏆 Wins

I debugged a broken production API integration — code I didn't write, in a language I don't speak — without hiring a developer. The ICM is why that was possible.

_____________________________________________________________________________________

We run a custom PHP middleware that keeps two platforms in sync. A developer built it. I am not a developer.

It had been running degraded for months. Error logs hit 13 million lines in a 4-minute window. Inventory numbers were inflated. Customers added in one system never appeared in the other. A SQL injection

vulnerability was sitting quietly in the order dashboard. Nobody touched it because nobody felt qualified to.

____________________________________________________________________________________

Before I did any debugging, I built the workspace.

Not "open a chat and paste the error log." Build the folder first so the AI always has the right context for

the right task. That's the ICM move.

CLAUDE.md at the root — one file mapping the entire project. What the system does, where every file lives, what each folder is for, what naming conventions to follow. Every session loads this first.
Source files pulled from the server and organized locally into subfolders by function. The server has everything flat. The local folder gives the AI structure to reason about.
A parsed log summary instead of the raw log. 13 million lines uploaded once, converted into a structured breakdown by file, line number, error type, and frequency. Every future session loads the analysis. The noise is already filtered.
A fixes folder and an archive folder. Every fix gets a doc. Every original file gets archived before it's touched. Messy work happens here without polluting the reference layer.
A routing table. A markdown table that answers "I want to do X — go here, load these." Every session goes straight to what's relevant.

Once that structure was in place, the workflow was repeatable. Load context, describe the symptom, Claude reads the file and writes the fix, archive the original, deploy via the hosting file manager, update the fix doc. No terminal. No git. No local PHP environment.

Eight issues resolved across several sessions — null-safety gaps generating millions of warnings per cron run, a customer sync lockout caused by a success flag written before the API call that was supposed to trigger it, inventory double-counting from a misread field, stale inventory buckets that required a DELETE-then-POST pattern to clear, collapsed multi-PO logic, a SQL injection fixed with a single type cast, and a sync gate to filter internal items from the wholesale catalog.

End state: error log volume near zero, inventory accurate, customers syncing end-to-end, security gap patched, every fix documented.

_______________________________________________________________________________________

What Claude Code allowed was a different kind of collaboration: the operator understood the business outcome (inventory should match, customers should sync, orders should flow), and Claude understood the code. Together, that was enough to fully debug a production integration that had been broken for months - without a developer on call.

The lesson isn't that you don't need developers. It's that with the right tooling, a non-technical operator can be a capable debugging partner for their own systems and get things fixed that would otherwise stay broken.

2 comments