Summary
Nick Saraev describes giving Claude Code access to an agency he runs that earns roughly a million dollars a year, and asking it to automate as much of his knowledge work as possible. He says that after years of disappointing attempts with models going back to GPT-3.5 Turbo, the results this time were genuinely strong.
The video lays out a repeatable process. First, audit your work: Saraev had Opus 5.5 on max thinking build a tool that captures compressed screenshots every few seconds, labels them, and stores them in a database, while also asking him questions throughout the day. That produced a full inventory of his tasks, such as lead scraping with Apify, email enrichment, cold email copywriting, domain warmup, and proposals.
Next, he built evals for each task, running each ten times, scoring outputs himself, and iterating until most reached about 8 out of 10. Tasks that cleared the bar became shareable skills that teammates can use even without Claude Code, and he recommends batching test changes in parallel to save time. Finally, he asked Claude which skills are safe to run on a schedule, such as a daily client update draft, and which still need human involvement.
📺 Source: Nick Saraev · Published October 13, 2026
🏷️ Format: Workflow Case Study
