The Bug That Kept Coming Back

Short summary

A developer built a Blockly platform entirely through Claude Code via Telegram and discovered a recurring pattern: the same Spring Security misconfiguration kept resurfacing across four separate endpoints. Each individual fix was correct, but neither the human nor the model had enough context to notice the systemic gap. The core insight is that chat-only development lacks the file-level continuity that catches recurring issues, and human-in-the-loop approval is meaningless when the human has no more context than the model.

• •Chat-only AI coding via Telegram caused the same Spring Security bug to recur four times because neither party tracked file-level history
• •Each individual fix was correct in isolation; what failed was systemic awareness of recurring patterns
• •Human-in-the-loop oversight is hollow when the approver has never read the codebase

Generated with AI, which can make mistakes.