19 Claude Code Mistakes “Pro” Users Are Still Making

A tactical guide to recent changes in Claude (Anthropic) best practices: which prompt elements matter, how connectors and routines affect cost and security, context window and compaction behavior, sub-agent limits, and model-switching pitfalls. The video diagnoses 19 common mistakes, offers practical fixes (auto-compact, tool-access settings, concise Claude.md), and token-cost optimization tips.

– Prompt design: remove persona fluff; instead tell Claude where to look, define done criteria, and include a self-check. Ask for positive formatting rather than ‘do not’.
– Connectors & routines: disable default tool access for routines and set connectors to load when needed to lower token costs and reduce exposure.
– Context & compaction: use auto-compact and ‘summarize up to here’ to manage window size; keep Claude.md concise (≈300–350 words) so rules aren’t lost.
– Agents & models: sub-agents often lack conversation history and are token-expensive; switching models uncaches context and can increase costs unexpectedly.

Quotes:

All of those additional lines you’ve been writing at the top of every prompt for the last year, the model isn’t even reading.

At 1,000,000 tokens the retrieval accuracy dropped to 76% — bigger context can harm recall.

Cheaper models can actually cost you more because each model has its own cache.