Engineering notes on AI hallucination prevention and honest AI.
Claude Code, Cursor, Copilot, Windsurf — none of them verify AI output against your actual project. Here's what's missing and how to fix it.
22 June 2026The default for AI-generated code is trust. AI writes, code ships, bugs surface at runtime. Nobody is verifying the output against the actual project.
22 June 2026AI models know about 169.254.169.254, localhost, and private IPs. When you give them agent tasks, they will target these endpoints.
22 June 2026One command. Every IDE. Install Check and start catching hallucinated imports, phantom functions, and bad commands.
22 June 2026Your AI coding agent has no idea what packages are in your project. It generates code based on training data, not your dependency tree.
22 June 2026Better prompts reduce hallucinations. They don't eliminate them. Verification is deterministic. Prompting is probabilistic.
22 June 2026Security firewalls, LLM guardrails, and eval platforms all verify the AI. Check verifies the code the AI writes. Different axis.
22 June 2026A single bad command triggers retries that compound inference cost. One Check call costs $0.0068 and prevents the entire chain.
22 June 2026AI models invent function names that look real but don't exist in your project. Check compares every call against your actual exported symbols.
22 June 2026AI coding agents suggest packages not in your project. Check reads your package.json after every file write and catches hallucinated imports.
22 June 2026