Documented: Non-Technical PM Stuck in AI Debug Loops for 3 Months

This case study is based on a real published account by a non-technical product manager — not a fictional "waitlist app founder." It explains why "the AI keeps fixing one thing and breaking another" is now a top search term, and why my audit-first model exists.

The documented experience

Author: Andreea Papillon — product leader, explicitly writing for "solo founders and small teams without developer resources."

Published: November 13, 2025 — "Debugging with AI: When It Goes in Circles — and How I Got Unstuck" (Part 2 of a 3-part series on going from prompt to production).

What she built

Three months of building an app with AI tools (Warp agent, ChatGPT for specs). Real production target: App Store + web, Supabase auth, Google Cloud backend.

What broke (her data)

After 3 months she categorized her bugs:

• ~90% backend — auth worst (sessions, token issues, service account permissions).
• ~10% frontend — mostly vague requirements the AI filled with wrong assumptions.
• Auth bugs took 3–5× longer to debug than anything else.

The "black screen" loop (verbatim pattern)

App worked locally. Sent to App Store Connect. TestFlight build: black screen, no error.

She spent days in a loop with Warp and ChatGPT:

Agent: "Now the code should compile!"
Her: "Yes, we have done this several times. The build will compile but the UI won't load… then we remove the optional, it comp compile but the screen will be black. And so on."
Agent: "You're right — we've been going in circles."

Resolution: She stopped, slept, came back with a structured debug summary the AI had written the night before, and fixed it in 20 minutes. Root cause: a React Native configuration change broke rendering.

Her published rules (directly relevant to rescue work)

1. Stop after 3 loops or 20 minutes — start a new conversation with a debug summary.
2. Sleep is a debug tool — AI gets context-confused after 10–15 exchanges.
3. Mock data trap — AI fakes API integrations so code "runs"; you think Stripe/Supabase is wired when it isn't.
4. Environment variables — "works on my machine" because .env loaded from wrong working directory.
5. "Works locally, fails staging" — redirect URLs, CORS, migrations not run in staging.

Industry parallel

This is not one person's quirk. The Cursor community forum documents Claude Sonnet repeatedly applying the same failed fix. Credit-based tools (Lovable, Bolt) charge for each loop, which turns negative-progress debugging into real money lost.

How my service maps to this documented pattern

Papillon's "debug summary + fresh eyes" is exactly what my $299 triage formalizes — except done by a senior engineer who reads the whole repo, not the last 15 chat messages:

Her self-serve playbook	My engagement
Write debug summary	Written report + recorded video walkthrough
Stop after 3 AI loops	Audit before any more prompts or credits
Compare to last working Git state	Root-cause read across auth, env, deploy
"90% backend"	Priority Fixes on auth, RLS, webhooks — not UI symptoms

I don't claim I was her engineer. I claim this is the documented failure mode my service is designed for — and it's searchable.

What you should search to verify

• andreea papillon debugging with AI goes in circles
• cursor claude stuck in loops forum
• vibe coding non technical founder backend bugs

Sources

• Debugging with AI: When It Goes in Circles (Andreea Papillon)
• From Product Manager to AI Builder (series)
• Claude Sonnet 4.0 gets stuck in loops (Cursor forum)

---

Stuck in the same loop on a Lovable/Bolt app? When to stop prompting or $299 triage.