Claude Code: Lessons From 3 Weeks of Usage
HerrKader · reddit · 2026-07-11
Over 19 days, the author used Claude Code for 245 sessions, 70,000 messages, and a cumulative 123.1 million tokens, summarizing 7 genuinely useful practices:
- CLAUDE.md and persistent memory are crucial: Pinning down the context reduces repetitive background explanations, evolving the experience from a "smart autocomplete" to a "teammate who remembers decisions."
- Repeatedly rereading context is the real cost: Long sessions constantly review the same context. The author suggests splitting tasks by stages, letting Claude generate a prompt for the next stage after completing one, then using /clear to start a new session.
- Let skills / subagents handle independent tasks: Many early "jobs" are better suited for sub-agents to process in one go. The author also turned them into daily custom commands.
- Add self-verification: Having the system open a browser or preview output to check itself reduces "confidently wrong" results better than mere text comparison.
- Practical limits of pricing tiers: The author believes the $200 tier only lifts the 5-hour limit, not the weekly limit. He confirmed this twice with support; the first time he was told he misunderstood, and support only changed their tune after being asked to check the docs.
- Don't just follow social media "Claude Gurus": Many tutorial creators on X / Instagram lack deep knowledge. Going directly to GitHub repos to let Claude read the code is more valuable.
- Treat Claude as an employee, not an oracle: The author emphasizes questioning, challenging, and delegating; use Fable for planning important decisions first, then switch to Opus for execution.
More from coding & agent
- Soft Clamp cuts tool-call overuse in multi-teacher distillation, from 13.7% to 9.0% — antgroup · 2026-07-21
- Agent harness memory loss and compaction are still a major usability problem — adityaag · 2026-07-21
- SpecJudge runs locally on Ollama to pick the right-sized AI model for your project — jokiruiz · 2026-07-21
- A developer maps out six design rules for CLIs that humans and AI agents can both use — yujiezha · 2026-07-21
- A coding-agent skill that forces ADHD-friendly, answer-first output — ayghri · 2026-07-21
- A set of agent skills for CAD, robotics, and hardware design — earthtojake · 2026-07-21