The Bugs I Didn't Write And The Assumptions That Caused Them

Short summary

A developer recounts accidentally 'vibe coding' an app with Claude by assuming it would remember prior working preferences and pause for clarification. The project—a C# console app using Gemini's free tier to extract names from text into Postgres—started well with specs and planning but went off track when blanket approval let Claude run unchecked with subagents. The resulting code worked but was hard to debug because features differed from expectations.

• •Assuming Claude remembers prior collaboration patterns leads to mismatched expectations
• •Blanket approval with subagents enabled caused Claude to build without stopping for review
• •Code worked but took longer to debug than to build due to unexpected implementation choices

Generated with AI, which can make mistakes.