Boris Cherny
Boris Cherny
One-line summary: Creator of Claude Code at Anthropic. Public voice on how the Claude Code team itself uses the tool.
What he is
Engineer behind Anthropic's Claude Code CLI. Active on X as @bcherny. Interviewed by Charly Wargnier (@DataChaz) in an 11-minute segment where he walked through his personal setup and 15 workflow tips.
Why he matters to this thread
He is the highest-signal source for "how the people who built Claude Code actually use it" — distinct from third-party influencer advice. His claims carry weight as design intent and dogfooding data.
Key positions (from 2026-04-21-boris-claude-techniques)
- Runs 5–10 Claudes in parallel across surfaces (terminal, web, iOS, Android). Kicks off sessions from his phone as soon as he wakes up.
- Uses Opus 4.5 with thinking for everything. Argues it's cheaper in practice than smaller models because it uses fewer tokens to reach a result.
- Almost always starts in plan mode, iterates on the plan, then switches to auto-accept edits and lets the model execute.
- The Claude Code team shares a single CLAUDE.md checked into git; the whole team edits it multiple times a week.
- Tags
@claudeon teammates' PRs to patch CLAUDE.md from review comments — his version of Dan Shipper's "compound engineering." - Does not use git worktrees; keeps separate checkouts in separate terminal tabs instead.
2026-08 Slack routines experiment (from 2026-08-18-autoresearch-effective-all-day-bot-loops)
Mid-August 2026: Slack channel proj-claude-maintains-apps where Claude Tag runs daily routines across iOS, Android, Desktop, web, CLI, and Agent SDK — crash fuzzer, duplicate-abstraction unifier, dead-code remover, "abstraction police." Over "the last few weeks" those routines opened 388 PRs, of which 180 were merged after Claude Code Review plus human review. When a PR is wrong, they ask Claude to tune the routine so the next day is better.
That is scheduled isolated jobs + a review gate (the unattended-all-day-agent-loop outer-loop shape), used as day-to-day app maintenance. It is not evidence that Slack @-mentions are the right unattended scheduler for a research queue, and the 180/388 merge rate still required a human. Vintage: 2026-08 snapshot.
Using the loop, not ranking harnesses (X, August 2026)
From 2026-08-27-best-agent-harnesses-for-programming (primary @bcherny posts; paraphrased in the synthesis — no long verbatim quotes filed here):
- Shared memory got simpler (2026-08-26).
- Opus is used to “hill climb” profiler-closed loops until a numeric target (2026-08-23).
These are using the harness loop, not a ranking of Claude Code vs Codex vs OpenHands. Do not read them as a “best harness” verdict.
Notable quote
"Once the plan is good, the code is good." — 2026-04-21-boris-claude-techniques
Sources
- 2026-04-21-boris-claude-techniques
- 2026-08-18-autoresearch-effective-all-day-bot-loops
- 2026-08-27-best-agent-harnesses-for-programming — @bcherny shared-memory / hill-climb (use the loop, not a ranking)