Tokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers

After 13 years away from coding, Tan used Claude Code to rebuild Posthaven in about five days for roughly $200, compared with a year and a half and a team of six the first time. He then adapted the work into Gary's List, a publishing platform that researches topics using web sources, social posts, and retrieval tools.

Tan argues that paying for more model calls can make research more complete. For Gary's List, he says roughly $5 to $10 in model calls can support work that would take a person much longer, and the system compares many sources before drafting. He applies the same approach to coding when extra model effort improves the result.

Tan collected prompts he kept reusing and shaped them into G Stack skills for planning, design, engineering review, and testing. He says asking an agent to map data flows and user journeys before coding helped it handle complex work more completely. He also uses browser based QA and brings in Codex to review plans or code for bugs.

Tan says people need to provide the goals, context, and product judgment, while agents handle work the person would rather not do. His approach puts flexible instructions in Markdown skills and deterministic actions in code, with 80 to 90 percent test coverage as a practical target. He describes OpenClaw as powerful but fragile enough that users should know how to troubleshoot it.

Tan predicts personal AI will become broadly available and frames the key question as whether people control their tools and prompts. He and the host compare token spending to paying for a useful startup advantage, arguing builders should spend more when the work creates real value. Tan says directing many agents helped him produce about 400 times as much code as when he last coded, while emphasizing that he directed the work rather than writing it himself.