Something Went Wrong

I pressed the payment button. A modal opened. A card entry form should have appeared. Instead, this:

"Something went wrong."

Five words. Enough to tell me nothing. No error code, no detailed message. Just "something went wrong."

I sent it to the control tower. But this time, something unexpected happened.

Instead of diagnosing the error, the AI asked me a question.

"The master plan document specifies LemonSqueezy as the payment service, but the current implementation uses Paddle. Why was this changed?"

Then another.

"The Unlimited tier is $19/month, but the Max tier is $29/month. Higher tiers are typically more expensive. Why is Unlimited cheaper?"

Both questions seemed reasonable. Both were based on wrong premises.

Switching from LemonSqueezy to Paddle was intentional. Paddle acts as a Merchant of Record, automatically handling tax compliance across countries. When someone in the US, Japan, or Europe buys a service made in Korea, different tax regulations apply—Paddle handles all of that. The master plan was written before this decision.

The Unlimited tier being cheaper was also intentional. BYOK—"Bring Your Own Key." Users bring their own AI API keys. That means I don't bear the AI costs. No AI costs means lower price. Unlimited usage, but the user pays their own AI expenses directly.

The AI didn't know this context. It referenced the master plan—an outdated document—and applied common-sense pricing logic. Neither was wrong, but neither matched the current reality.

For the first time, I told the AI:

"Check the claude.md file."

The claude.md is the project's latest configuration file, reflecting decisions more recent than the master plan. After the AI read it: "I see—the switch to Paddle was intentional. I understand the BYOK structure now."

With the premise corrected, the diagnosis changed.

The AI narrowed it to three suspects. First, Paddle's KYC (seller identity verification) might not be complete—Paddle won't process payments until it approves you as a legitimate seller. Second, a mismatch between Sandbox and Live API keys—test keys start with test_, live keys with live_, and mixing them causes failures. Third, product IDs registered in Paddle's dashboard might not match those in the code.

What I learned from this episode wasn't how to fix payment errors.

It was experiencing that AI can be wrong.

More precisely—the AI wasn't wrong. The context it referenced was outdated or incomplete. AI judges only within the information you provide. Give it an old document, it makes old judgments. Give it partial information, it makes partial conclusions.

So when directing AI, what matters isn't just "good questions" but "good context." And when the AI's answer comes back, verifying whether that answer is correct is your responsibility. Not just executing the control tower's diagnosis, but first asking: "Is this diagnosis based on correct premises?"

In EP.03 I wrote that I'm the PM and AI is the capable junior. This episode extends that metaphor. A capable junior produces excellent analysis within the information given. But if the information is wrong, even excellent analysis reaches wrong conclusions. Verifying the accuracy of information—that's the PM's job.

I still can't read code. But knowing which context to give AI, and which premises to check in AI's answers—that's a judgment someone who can't code can absolutely make. And that judgment changes outcomes.

---

🔧 Technical Terms in This Episode

Payment Modal A popup window for entering payment information. The card entry screen that appears when you click "Buy" on an e-commerce site. Paddle provides this interface.

Paddle A payment processing service. Acts as MoR (Merchant of Record), automatically handling tax calculation and remittance across countries. Especially advantageous for global services.

MoR (Merchant of Record) A service that legally acts as the seller on your behalf. When an American buys a service made in Korea, the MoR handles US tax compliance.

KYC (Know Your Customer) Seller identity verification. Paddle must confirm "you are a legitimate seller" before payments can process. Without approval, the payment modal shows errors.

Sandbox vs Live Test environment versus production environment. Sandbox uses fake money to test the system; Live involves real transactions. Test API keys and live keys are different—confusing them causes errors.

BYOK (Bring Your Own Key) A model where users provide their own AI API keys. Since the service provider doesn't bear AI costs, subscription prices are lower.