Key Points
- 400x productivity claim: Garry Tan, YC President/CEO, returned to coding after 13 years and rebuilt Posterous in 5 days for $200 versus the original 18 months, $4M, and 6-7 engineers (2008-2012). His open-source gstack project reached ~90K GitHub stars in roughly 60 days.
- Tokenmaxxing philosophy: Instead of minimizing token use, "max out" tokens. Tan frames it as "boil the ocean" — when marginal cost of completeness approaches zero, always pursue exhaustive, full implementations rather than shortcuts.
- Thin Harness + Fat Skills: Offload execution infrastructure (Claude Code, OpenClaw, Codex) to mature tools; invest human effort in detailed markdown skill prompts that encode domain knowledge, workflows, and judgment criteria.
- 23 specialist roles: gstack decomposes Claude Code into a virtual engineering company — CEO, product manager, architect, engineer, security auditor, code reviewer, designer, QA, DevOps, and browser agent — each invoked via a
/commandslash interface. - Conductor parallelism: Tan orchestrates 10-15 simultaneous Claude Code sessions, each in its own Git worktree, allowing independent feature development without interference.
- Persistent browser daemon: A Bun.serve HTTP layer wraps Playwright/Chrome over the CDP protocol. Subsequent commands run in ~100ms versus ~40 seconds for cold-started Playwright, with auto-shutdown after 30 minutes idle.
- Accessibility-tree ref system: Replaces fragile CSS selectors with ARIA-tree traversal and sequential refs (
@e1,@e2), reliably handling Shadow DOM, CSP, and post-hydration frames. - Diff-aware QA: On a feature branch, the QA agent analyzes
git diff, maps affected routes, and tests only changed code paths — roughly 10x faster than full-suite runs. - GBrain knowledge cache: A Brain-First workflow queries a local indexed knowledge graph before calling external APIs; misses are written back. Tan reports 17,888 pages, 4,383 people, and 723 companies in his live usage.
- Three design rules: (1) Always recommend the complete option, not the 90% shortcut; (2) AUTO-FIX mechanical issues, ASK only for genuine judgment calls; (3) Be non-interactive — execute, don't nag for confirmation.
- Replication path: Subscribe to Claude Code/Codex (harness), write one
SKILL.mdper recurring task (skill), think in terms of task delegation rather than writing code, and give AI full context instead of trimming prompts. - Critique and limits: The 400x figure uses logical lines of code, not raw LOC; AI tends toward verbosity. gstack is tightly coupled to Claude Code (requires Bun, MCP, specific Git layout). The 23 roles cover fast-iteration/startup workflows but lack
/compliance-check,/legacy-migration, and/performance-profiling. Critics call it role-switching within one instance rather than true multi-agent orchestration; Tan responds that what matters is shipping, not taxonomy. - gstack GitHub: https://github.com/garrytan/gstack
- YC Lightcone podcast episode (2026-05)
- Augment Code coverage (2026-05-05)
- juejin deep dive (2026-03-18)
- SOTA Sync analysis (2026-05-10)
- Xiaoyuzhou podcast EP98 (2026-05-11)
TL;DR
> "Treat AI as a 15-person engineering team to manage, not a faster senior programmer." Tokenmaxxing is not waste — it is recognizing that in the AI era, information density matters more than code density. Boil the ocean. Implement completely. Let AI do what it is good at (exhaustive search, full test coverage, parallel execution); let humans do what humans are good at (judgment, taste, direction).