When AI Slaps You in the Face

When AI Slaps You in the Face: Debugging Claude-Generated Code in Adama

Ever let an AI "vibe-code" a complex feature, only to spend hours debugging subtle bugs it missed? As a dev building Adama—a platform for stateful serverless apps—I've been using Claude to accelerate development. It nails the big picture but trips on low-level gotchas like async timing and pipeline configs. Here are three real debugging war stories, with lessons on keeping AI in check.

WebSocket Activation: Netty Pipeline Nightmares

I tasked Claude with building an MCP (Message Control Protocol) server in Adama to let AI compile Adama docs on-the-fly. It spat out a solid first pass: a "book generator" that parses and validates every code snippet in my docs, plus WebSocket support via Netty for real-time interaction.

But unit tests exploded. Claude's code used multiple WebSocketServerProtocolHandlers in the Netty pipeline—impossible, as Netty only allows one. I prompted fixes, but it spiraled: reverted, added abstraction layers, more test failures.

The real killer? Tests weren't waiting for the channel write before handshake. Logging revealed the async mismatch. Claude burned through my context window without spotting it—likely because it couldn't build/run the code itself.

Lesson: AI excels at boilerplate but lacks runtime introspection. Fast local builds and rich logs are essential. Once fixed, the MCP keystone worked beautifully, letting Claude compile Adama snippets seamlessly.

Teaching Adama: Build Times and Context Loss

Next, I had Claude rewrite Adama docs and generate samples. My new Markdown-to-site engine reports compile success: 80% on first pass, auto-refined from source.

But monorepo builds take 10 minutes—fine for human IDE probing via unit tests, deadly for AI iteration. Claude went off-rails on partial features (e.g., unfinished type system), lacking full context.

Lesson: Mirror your IDE for AI. Optimize builds for speed, enrich compiler errors with doc hints. Prompt with "mainline + experiments" summaries to simulate team handoffs. Shipping > experimenting.

Large Refactors: When Context Overflows

Adama's type system is dense—even compact prompts lead to hallucinations on transcribes. $10k+ tokens later? Still broken. AI chokes on full-language refactors without specialist "agents."

Lesson: Break into micro-tasks. Shape prompts like org charts: specialists for subsystems. Devs, this is your new skill—managing AI like a distributed team.

Keep AI on a Leash: Actionable Takeaways

AI accelerates if you enforce:

• Fast feedback loops: Sub-minute builds/tests.
• Rich signals: Logs, errors-as-docs.
• Scoped prompts: No monoliths; chunk wisely.
• Human oversight: Review everything.

Huge wins: Full doc generator, MCP for AI self-improvement, 80% compile rate. But debugging slapped me awake—AI isn't magic; it's a junior dev on steroids.

Try Adama yourself and star adama-core to support stateful dev tools. What's your worst AI-debug story?

Original post: https://nexivibe.com/posts/ai-generated-code-debugging.html