Anti-patterns while working with LLMs (instavm.io)

After 15 months working with LLMs the author lays out five practical anti-patterns to avoid: wasting scarce context by resending near-duplicate data (e.g., every tiny screenshot of a moving cursor) instead of only state-change frames (they built open-source click3 to avoid duplicate screenshots); asking models to do tasks they’re weak at (e.g., brittle character-counting or precise layout quirks) instead of asking them to generate code or a structured tool call which is often more reliable; overfilling long-context sessions (128k+ tokens) where models can “forget” or invent values like database URLs; expecting high accuracy on obscure or post-training-cutoff topics; and becoming a passive “vibe-coder” who stops reviewing model outputs, risking logic or security mistakes (example: LLM inserting a User object with password_hash into a generated invoice JSON). Significance for practitioners: these are system-design and workflow rules of thumb that materially affect reliability, cost, and security. Technical implications include treating context as a limited resource (use compression, deduplication, state-change sampling), preferring code/tool-calls for deterministic behavior, aggressively validating long-context outputs, flagging low-confidence domains, and keeping human-in-the-loop checks to catch data leaks or logic errors. Following these patterns improves robustness, reduces token costs, and lowers attack surface when deploying LLMs in production.