The Fleet Commander Strategy
Boris Cherny treats Claude Code as reliable capacity—not a chatbot. The result: 100% AI-written production commits, achieved by managing parallel sessions, planning first, and enforcing verification loops.
Direction over syntax.
Terminal, web, mobile.
From proof-driven loops.
1 Parallelism: run wide
The bottleneck is often the human waiting for the model. The fix is operational: run multiple sessions in parallel and only context-switch when an agent needs input.
- ▸ Numbered terminal tabs: each tab owns a single task (refactor, tests, docs).
- ▸ Web + mobile monitoring: extra capacity for quick reviews and follow-ups.
- ▸ Async notifications: context-switch only when a decision is needed.
Active session distribution
Illustrative concurrent load
Time allocation shift
Typing collapses; planning + verification dominate.
2 Plan mode first
The goal is “one-shot” execution: refine the architecture until implementation is largely deterministic.
Enter plan mode
Stop coding. Start specifying.
Refine & iterate
Converge on a clear spec.
Auto-accept
Implement in one pass.
The “one-shot” goal
Refine the plan so thoroughly that execution is a single successful pass—minimising back-and-forth debugging and wasted tokens.
3 Shared memory: CLAUDE.md
Every recurring mistake is a permanent rule. The repo becomes an operating manual that agents follow—reducing repeat errors and compounding productivity.
# Project guidelines & memory
## Coding standards
- Prefer type over interface
- Avoid enum; use union types
- Use vitest instead of jest
# Updated 2026-01-08: capture recurring failures as permanent rules.
Knowledge compounding
Abstract efficiency score over time
4 Verification loops
Code only counts when it’s proven: write tests, run them, fail, fix, pass—then human review. This is where quality lifts by 2–3×.
Verification impact
5 Tooling
Claude Code
Treat it as schedulable capacity: tasks, queues, sessions.
Slash commands
Standardise actions like /verify and /simplify.
Post-tool hooks
Auto-run lint/format to save tokens and reduce drift.