How the Creator of Claude Code Actually Uses It — Boris Cherny's 53 Tips

i have seen this thread ongoing few months back but didnt have the time to sit down and go thru them, and happened to see one recommendation on my youtube landing page and just happened to be quite curious to know, to improve on my current workflow!

The Three Core Principles

boris’s entire philosophy boils down to three things:

1. parallelism — turn serial work into parallel work with worktrees + multiple sessions
2. plan first — spend energy on the plan so execution is one-shot
3. verification loops — let Claude verify its own work; quality 2-3x improvement

everything below is a variation on these three ideas.

1. Parallel Execution

the single biggest productivity unlock, according to the entire Claude Code team.

• 5 terminal tabs, each a separate git checkout, numbered 1-5. iTerm2 notifications alert when Claude needs input

personally, i didnt go to that extreme, i had it around 3 tabs: (1) Second brain, (2) Work Assistance, (3) Work Features

• 5-10 more sessions on claude.ai in the browser. use --teleport to hand off between local and web. you can even kick off sessions from your phone.

this is something i’ll be trying too, but for now, no.

• git worktrees over checkouts — the team prefers worktrees. shell aliases (za, zb, zc) for one-keystroke switching.
• subagents also support worktrees — for large migrations, each agent gets its own worktree, runs tests, and creates a PR independently.

this is one of the best things i got from this thread.

2. Plan Mode

start every complex task in Plan Mode (shift+tab twice).

• iterate on the plan with Claude until you’re satisfied
• switch to auto-accept edits mode → Claude typically one-shots the implementation
• one team member has a second Claude review the plan as a “Staff Engineer” before execution
• if things go sideways mid-implementation: re-plan from scratch, don’t patch

3. Model Choice

boris exclusively uses Opus 4.5 with thinking enabled. (currently 4.6 as im writing this blog)

It’s the best coding model I’ve ever used. Even though it’s bigger and slower than Sonnet, since you have to steer it less and it’s better at tool use, it is almost always faster than using a smaller model in the end.

― Boris Cherny

honestly, this is just something i have been doing from the start of my coding career as i truly trust that planning process is the key. even tho it costs the most time during my junior year, but see what we got here? what i spent on clarifying requirements and adjusting scopes built me to have a clear mind to think about how to develop a big task.

4. CLAUDE.md — Compounding Engineering

the team maintains a single CLAUDE.md file checked into git.

• every time Claude makes a mistake → add it to CLAUDE.md so it won’t repeat
• tag @.claude in PR reviews → GitHub Action automatically updates CLAUDE.md
• ruthlessly edit over time to keep it concise
• maintain a notes directory for each task/project, pointed to by CLAUDE.md

5. Slash Commands & Skills

store frequently-used workflows in .claude/commands/, checked into git.

• boris uses /commit-push-pr dozens of times daily
• rule: if you do something more than once per day → make it a command or skill
• examples: /techdebt, Slack/GDrive sync, analytics-engineer agents
• commands support inline Bash for pre-computing context

wow, i just found so much similarities in this. thats sth i have been doing too!

6. Subagents

custom agents live in .claude/agents/ as .md files.

• code-simplifier: reviews code for reuse and quality after Claude finishes
• verify-app: end-to-end testing instructions
• add “use subagents” to any prompt → Claude distributes work automatically
• permission requests can be routed to Opus via hooks for auto-approval

as well as subagents!

7. Hooks

deterministic lifecycle hooks — the infrastructure layer.

this is new to me tho

8. Verification — The Most Important Tip

Give Claude a way to verify its work… 2-3x the quality of the final result.

― Boris Cherny

• UI testing: Claude Chrome extension
• backend: run tests, bash validation
• distributed systems: point Claude at docker logs
• build domain-specific feedback loops for your codebase

9. Permissions & Safety

• use /permissions to pre-authorize safe commands — never use --dangerously-skip-permissions
• wildcard syntax: "Bash(bun run *)", "Edit(/docs/**)"
• /sandbox for file + network isolation
• check settings.json into git so the whole team benefits

10. Tool Integrations

configure in .mcp.json for autonomous access:

• Slack MCP: paste a bug thread → say “fix”
• BigQuery: bq CLI for on-the-fly metrics

I haven’t written SQL in 6+ months.

― Boris Cherny

• Sentry: log access for debugging

11. Prompting Techniques

• challenge Claude: “Grill me, don’t PR until I pass your test”
• demand proof: “Prove this works, diff main vs feature”
• write detailed specs
• never accept the first solution

12. Terminal & UX

• Ghostty terminal — the team’s favorite (synchronized rendering, 24-bit color, proper unicode)
• /statusline — show model, cost, context remaining
• /color — different color per session for visual separation
• voice dictation: macOS fn×2, reportedly 3x faster than typing
• /vim for vim mode
• custom spinner verbs (someone did Star Trek themed 😂)

and because of this, i switched to Ghostty, wow… i think i will stick with ghostty from now on. GPU, faster, multiple sections in a tab, just wow.

13. Power Features

What I’d Actually Adopt

already doing:

• plan mode before execution — been my approach since junior dev days
• slash commands & skills — same instinct, automate the inner loop
• subagents for distributed work
• opus as primary model
• ghostty as terminal

will try:

• git worktrees for parallel isolation — the biggest takeaway from this thread
• --teleport for web ↔ local handoff
• PostToolUse hooks for auto-formatting
• PostCompact hooks to re-inject context after compression

skipping (for now):

• 5+ parallel sessions — overkill for my current projects, 3 tabs is enough
• BigQuery/Slack MCP — not relevant to my stack
• custom spinner verbs — fun but not productive

Sources

• How Boris Uses Claude Code (dedicated site)
• Boris’s original X thread
• Boris’s team tips X thread
• VentureBeat coverage
• Pragmatic Engineer interview