Summary
After spending more than $31,000 and roughly 1,000 hours on Claude Code, Nick Saraev shares the lessons he wishes he had known at the start. The tips are ordered from simplest to most advanced, so both newcomers and heavy users can pick up something useful.
He begins with the point that language models are not deterministic, so a prompt that works once should not be standardized into a business process. Instead, he recommends running prompts through repeated evals, such as ten runs, counting successes and keeping changes only when the score improves. He also explains why models such as Opus 5.5 should be given a way to check their own work, much like drafting an essay in multiple passes.
Other advice covers how to direct the model. Saraev suggests treating Claude like a contractor by defining what done looks like instead of dictating every step, and dropping the heavy “never do X” instructions left over from older models. He also recommends asking Claude to list problems first and only then choosing which ones to fix, which saves tokens and time. The video is a practical guide for developers and teams using Claude Code in real workflows.
📺 Source: Nick Saraev · Published September 25, 2026
🏷️ Format: Workflow Case Study
