5 Claudes in parallel terminal tabs
Tabs numbered 1–5. iTerm2 shows system notifications when a Claude needs input. Never wait on a single process.
150
PRs / Day (Record)
5+5
Parallel Claudes (CLI+Web)
2–3×
Quality via Verification
$6
Avg. Cost / Active User / Day
259
Boris' PRs in 30 Days
Boris Cherny is the creator and Head of Claude Code at Anthropic. He started the project in September 2024 as a side project; launch was on February 24, 2025. His single most important rule: "Give Claude a way to verify its work — it will 2–3× the quality."
His setup is surprisingly simple: 5 parallel terminal tabs + 5–10 web sessions on claude.ai/code + mobile sessions, always starting in Plan Mode, with a shared CLAUDE.md, slash commands in .claude/commands/, subagents in .claude/agents/, a PostToolUse format hook, and /permissions (never --dangerously-skip-permissions outside of sandboxes).
The arc of his tips from launch to today: "Chatbot in the terminal" → "Parallel execution engine" → "Autonomous, scheduled, voice-controlled agent". Plan Mode and CLAUDE.md (Q1 2026) → Hooks/Plugins (Feb 2026) → /simplify, /batch, Worktrees (Feb–Mar 2026) → /loop, /schedule, Code Review, /btw (Mar 2026) → Auto Mode, /focus, /go for Opus 4.7 (Apr 2026).
Table of Contents — 21 Themes
01
Parallel Execution
The single biggest productivity lever. Run multiple Claudes simultaneously — locally, in worktrees, in the cloud.
5–10 more Claudes on claude.ai/code
Handoff local↔web via & or --teleport. Boris starts sessions in the morning from the iOS app.
3–5 git worktrees, one Claude per worktree
"The single biggest productivity unlock." Team members use shell aliases za, zb, zc to jump; some have a dedicated "analysis worktree" for logs/BigQuery. Boris personally uses multiple checkouts instead.
claude --worktree (or -w)
Built-in worktree isolation. Name the worktree or let Claude name it. With --tmux everything starts in a tmux session.
Worktree mode in the desktop app
Code tab → check "worktree" checkbox. No terminal needed.
Subagents with worktree isolation
Tell Claude: "use worktrees for its agents". Powerful for large batched changes and migrations.
isolation: worktree in agent frontmatter
Run custom subagent permanently in its own worktree. Set in the YAML frontmatter of the agent definition.
Non-git VCS via WorktreeCreate Hook
Mercurial, Perforce, SVN, Jujutsu: define your own WorktreeCreate and WorktreeRemove hooks.
Dozens of Claudes via worktrees
Boris verbatim: "I have dozens of Claudes running at all times." Scales from the point where tab-switching alone slows you down.
02
Plan Mode
Let it think before every complex task. A good plan almost always one-shots the implementation.
Start every complex task in Plan Mode
Shift+Tab twice. Iterate on the plan with Claude, then switch to auto-accept edits. Claude usually one-shots when the plan is good.
"A good plan is really important"
Front-load thinking into the plan — implementation almost always follows one-shot. Front-loading reduces context usage across the entire session.
Re-plan instead of pushing when stuck
"Don't keep pushing. Switch back to plan mode and re-plan." When something goes wrong, force fresh context.
Have a second Claude review as "Staff Engineer"
Claude A writes the plan, Claude B reviews it with fresh context — like a senior review before coding.
Plan Mode for verification steps too
Don't just plan the implementation — "how do I verify this works" also belongs in the plan.
Automatic session names after Plan Mode
Claude derives a descriptive session name from the plan conversation. Easy to recognize in parallel setups.
03
CLAUDE.md
Your team's institutional memory. A Markdown file that enables compounding improvements.
A team-shared CLAUDE.md, checked into git
Team contributes multiple times per week. Becomes living documentation of the codebase.
Add a rule whenever Claude makes a mistake
"Anytime we see Claude do something incorrectly we add it to the CLAUDE.md." Compounding engineering at its core.
Each team maintains its own CLAUDE.md
Cross-team CLAUDE.md ownership lies with each individual team. Local knowledge stays local.
Tag @.claude in PR comments
Boris's version of Dan Shipper's "Compounding Engineering". Installed via /install-github-action.
"Update your CLAUDE.md so you don't make that mistake again"
End-of-correction prompt: lets Claude write the rule for itself.
Ruthlessly trim CLAUDE.md over time
"Keep iterating until Claude's mistake rate measurably drops." Pruning is just as important as adding.
Notes directory per task, referenced from CLAUDE.md
One engineer maintains a notes directory per task, updated after every PR. CLAUDE.md references it.
CLAUDE.md variants by scope
CLAUDE.md (project, in git), CLAUDE.local.md (gitignored, personal), ~/.claude/CLAUDE.md (global), per subfolder a/b/CLAUDE.md for monorepos.
Memory is just a Markdown file
"It's a file that has some stuff. And it's auto-read into context." Simpler than any vector store.
Prepend # to remember
"Add to Claude's memory by prepending # to something you want Claude Code to remember." Claude asks about memory location (project/user).
Auto-Memory + Auto-Dream
Auto-Memory saves preferences/corrections automatically. Auto-Dream is a subagent that periodically reviews and consolidates past sessions (REM sleep metaphor).
04
Slash Commands
The tool that encapsulates repeated prompts in a single keystroke. Boris's "inner loop" of daily work.
Slash Commands for every inner-loop workflow
Saves repeated prompting. Stored in .claude/commands/, checked into git. Claude itself can use them too.
/commit-push-pr dozens of times per day
Inline bash for pre-computed git status — avoids model back-and-forth.
/feature-dev as step-by-step guide
"First ask me what exactly I want, build the specification, build a detailed plan, then a to-do list, walk through step-by-step."
/commit — the most basic command
Most-used Slash Command at Anthropic. Auto-runs save+push without permission prompts.
/code-review on every PR
Human approves the merge, Claude does the first sweep. Standard at Anthropic team.
Slash Commands are prompts, not tools
"Slash commands are actually just like prompts. So they're not actually tools." Important for the mental model.
Local commands before MCP for simple cases
"If you just want something really simple and local… just use local commands for that." MCP only when needed.
/init generates starter CLAUDE.md
Analyzes codebase, detects build systems, tests, code patterns. First step in a new project.
/techdebt finds duplicated code
Run at the end of every session. Community favorite among custom commands.
Context-Dump Slash Command
Syncs 7 days of Slack + GDrive + Asana + GitHub into one prompt. Powerhouse setup of a team engineer.
/simplify (built-in skill)
Parallel agents improve code quality and enforce CLAUDE.md compliance. Usage: "make this change then run /simplify".
/batch (built-in skill)
Interviews you, then fans out to dozens/hundreds/thousands of worktree agents. Example: "/batch migrate src/ from Solid to React".
/loop for recurring local tasks (up to 3 days)
Boris's setup: /loop 5m /babysit, /loop 30m /slack-feedback, /loop /post-merge-sweeper, /loop 1h /pr-pruner.
/schedule for cloud-based recurring jobs
Survives a closed laptop. Anthropic team uses it for auto CI-failure-resolution, docs updates.
/btw for side-chain conversations
Single-turn, no tool calls, full context. "I use this all the time to answer quick questions while the agent works." By Erik Schluntz as side project.
/branch + --resume <id> --fork-session
Fork an existing session into a branched conversation. Saves re-contextualization.
/effort sets reasoning level
low / medium / high / xhigh / max. Boris uses xhigh for most tasks, max for the hardest. Max only applies to current session, other levels are sticky.
/focus Mode (CLI)
Hides intermediate work, shows only final result. "I generally trust the model to run the right commands and edits."
/go — Composite Skill
Claude tests end-to-end (Bash, Browser or Computer Use) → runs /simplify → opens a PR. "Many of my prompts look like 'Claude do blah blah /go'."
/fewer-permission-prompts Skill
Scans session history for safe-but-prompted commands; recommends allowlist expansion.
/voice activates voice dictation
Hold space bar in CLI, voice button in desktop, iOS dictation. "I do most of my coding by speaking to Claude, rather than typing."
/color for color-coding sessions
Easier to tell parallel sessions apart.
/statusline
Shows context %, git branch, model, cost. Everyone on the team has their own statusline.
/keybindings
Every key binding is customizable, settings live-reload. Stored in ~/.claude/keybindings.json.
/sandbox for open-source sandbox
File and network isolation. Modes: BashTool auto-allow, BashTool with prompts, off.
/plugin for browsing marketplace
LSPs, MCPs, Skills, Agents, Hooks. Companies can host their own marketplace.
/agents for creating custom agents
Place Markdown files in .claude/agents/. Frontmatter configures tools/model/system prompt.
/terminal-setup for Shift+Enter newlines
For IDE terminal, Apple Terminal, Warp, Alacritty.
/vim for Vim mode editing
One-line setup. Obvious for Vim users.
/teleport
Brings a cloud session down to local terminal (claude --teleport).
/remote-control
Controls a local session from any device. Boris has "Enable Remote Control for all sessions" in /config.
/compact <hint> instead of /clear mid-task
/compact is lossy LLM summary, maintains momentum. /clear + briefing is hand-written context. Rule: new task = clear, related = compact.
/rewind (or double-Esc) instead of correction
Don't pollute context with failed attempts + corrections — rewind and re-prompt with what you learned.
"Summarize from here" before rewind
Have Claude write a handoff message to its future self. Maximizes continuous learning across rewinds.
05
Subagents
The value is not in anthropomorphization — but in two independent context windows.
Subagents = automation for PR workflows
Stored in .claude/agents/. Reusable across the team.
code-simplifier Agent
Cleans code after Claude's work. Boris's standard after every implementation.
verify-app Agent
Detailed instructions for end-to-end testing. Verification step as reusable building block.
Append "use subagents" for more compute
One-word escalation to any prompt. Throws more compute at the problem.
Offload tasks to subagents for clean main context
Keeps the window free for important work. Research-heavy tasks are perfect for this.
5 parallel exploration agents on new codebase
"Use 5 subagents to explore the codebase" — runs in parallel. Fastest way to understanding.
Permission requests to Opus subagent via Hook
Hook scans for prompt injection attacks, auto-approves the safe ones. Massively reduces UI friction.
Adversarial subagent pattern for code review
Boris spawns subagents for style/history/bugs as first pass — then 5 more subagents that poke holes in the first findings to eliminate false positives.
Subagents = uncorrelated context windows
The value is not anthropomorphization — but two context windows that know nothing about each other.
Custom --agent Flag
claude --agent=ReadOnly. Agent .md frontmatter sets name, color, tools, model, system prompt.
Set default agent
"agent" field in settings.json or --agent flag. Enforce team-specific defaults.
Parallel subagents for plan research
"Can you research three separate ideas for how to do it? Do it in parallel. Use three agents to do it."
06
Hooks
Deterministic logic where the model alone is too unreliable. Formatting, logging, approvals.
PostToolUse Hook for auto-formatting
Matcher "Write|Edit", runs bun run format || true. Claude formats ~90% correctly, hook catches the 10% to avoid CI failures.
SessionStart Hook
Dynamically loads context whenever Claude starts. Ideal place for current branch info.
PreToolUse Hook
Logs every bash command the model executes. Audit trail for sensitive repos.
PermissionRequest Hook
Routes approval prompts to WhatsApp/Slack/SMS. Boris approves permissions from his phone.
Stop Hook keeps Claude running
Poke Claude when it stops. Can kick an agent or let a prompt decide whether to continue.
Agent Stop Hook for long verification
More deterministic than asking Claude to verify itself.
PostCompact Hook
Fires after Claude compresses its context. Useful for re-injecting critical instructions.
WorktreeCreate / WorktreeRemove Hooks
For non-git VCS (Mercurial, Perforce, SVN, Jujutsu). Define worktree logic yourself.
Ralph-Wiggum Plugin for long tasks
Community plugin by @GeoffreyHuntley for unattended cooking. For multi-day runs.
07
Permissions & Safety
Pre-allow is better than skipping everywhere. --dangerously-skip-permissions only in sandboxes.
No --dangerously-skip-permissions for normal work
Use /permissions instead to pre-allow safe commands.
Share .claude/settings.json with team
Check pre-allowed perms into git. Onboarding becomes a one-command thing.
Wildcard syntax for permissions
"Bash(bun run *)", "Edit(/docs/**)". Practical for entire tool families.
Auto Mode (Opus 4.7)
Model-based classifier auto-approves safe permission prompts. "No 👏 more 👏 permission prompts 👏". Shift+Tab cycles: Ask → Plan → Auto.
--permission-mode=dontAsk in sandboxes
Boris uses this (or --dangerously-skip-permissions) only in sandboxes for long runs.
Enable open-source sandbox runtime
Reduces prompts without --dangerously-skip.
Permission system: internals
"Combo of prompt-injection detection, static analysis, sandboxing, and human oversight."
08
MCP & Integrations
Slack, BigQuery, Sentry, Docker. Claude goes beyond your own repo.
Slack MCP, checked into .mcp.json
Shared with team. Claude searches and posts independently.
bq CLI for BigQuery
"Personally, I haven't written a line of SQL in 6+ months." Works with any DB that has CLI/MCP/API.
Sentry for error logs
Pulls error logs automatically into Claude. Bug triage without context switching.
Paste Slack bug thread → "fix"
Zero context switching. Slack thread + one word = implementation in minutes.
Point Claude at Docker logs
Surprisingly strong at troubleshooting distributed systems.
Slack MCP for daily summaries
/loop every morning use the Slack MCP to give me a summary of top posts I was tagged in.
iMessage as Claude Code channel
/plugin install imessage@claude-plugins-official. Text Claude from any Apple device via SMS.
09
Model & Effort
Opus with Thinking for almost everything. Stronger, slower, but usually faster in the end due to fewer corrections.
Opus 4.5 with Thinking for everything
"Even though it's bigger & slower than Sonnet, since you have to steer it less and it's better at tool use, it is almost always faster in the end."
Opus 4.6 1M as default for Max/Team/Enterprise
Pro/Sonnet users opt in via /extra-usage.
Opus 4.7 uses adaptive Thinking
Model decides itself when Thinking makes sense. No fixed budgets anymore.
"Think carefully and step-by-step" pushes Thinking
Or "Prioritize responding quickly" to save tokens.
Ultrathink trigger words (Legacy)
think < think hard < think harder < ultrathink — progressive Thinking budget. In CC v2 replaced by /effort, ultrathink remains highlighted.
Opus 4.7: shorter responses, less auto tool use
If you want length/style, say so explicitly. If Claude doesn't use the right tools, name them. For "Refactor across 40 files": explicitly request Subagents.
4.7's higher fidelity to "don't nitpick"
Code review harnesses for older models may see lower recall — not a regression, Claude follows instructions more faithfully.
Default was Sonnet, model is overridable
"We default to Sonnet for most everything… you can override the model if you want."
Write Plan/Think instructions directly
"If you want Claude to think, just tell it to think. Be like, you know, make a plan. Think hard. Don't write any code yet."
10
Verification — the #1 Rule
If you take away only one thing from Boris: give Claude a way to verify its own work. 2-3x more quality.
Verification = 2-3x quality
The most important single tip in the entire setup. Invest here first.
Boris's claude.ai/code change verification
Claude opens Chrome extension, tests every UI change in browser, iterates until UX is good. "Claude tests every single change I land to claude.ai/code."
Verification varies by domain
Bash, test suite, simulator, browser, Computer Use. Invest in their robustness.
Backend = run server/service end-to-end
Make sure Claude can start your service.
Frontend = Chromium extension
Boris uses it for every web code change. More reliable than other browser MCPs.
Desktop apps = Computer Use
Desktop app bundles automatic server start + built-in browser testing.
"Grill me on these changes" pattern
"Don't make a PR until I pass your test." Make Claude your reviewer.
"Prove to me this works"
Have Claude diff behavior between main and your branch.
After a mediocre fix
"Knowing everything you know now, scrap this and implement the elegant solution."
Code → Screenshot → Iterate
Show Claude a mock, let it use Puppeteer for comparing and iterating.
Catch errors early
"If the model is doing something wrong, it's better to identify that earlier and correct it earlier."
11
Long-Running & Recaps
Long-term tasks without keeping the laptop open. Routines, schedules, recaps.
(a) Background agent for self-verification
Simplest option: prompt Claude to spawn a background agent for verification when it finishes.
(b) Agent Stop Hook for deterministic verification
More reliable than (a). Code instead of model makes the decision.
(c) Ralph-Wiggum Plugin
Community pattern for very long runs. For multi-day unattended cooking tasks.
Recaps for long sessions
Short summary of done + next. Useful when returning after minutes or hours. Disableable in /config.
Routines (Research Preview)
Schedule, GitHub event, and API triggers. Runs on Anthropic infra. Laptop can be closed.
Cross-session continuity workaround
"Tell Claude to write down the state of this session into this text doc… in your new session, tell Claude to read from that doc."
12
Prompting & Spec Quality
Delegation beats guidance. Treat Claude like an engineer you're delegating to.
Delegation > Guidance (Opus 4.7)
"The model performs best if you treat it like an engineer you're delegating to, not a pair programmer you're guiding line by line."
Full task context upfront
Goal + Constraints + Acceptance Criteria — all three in the first turn.
Detailed specs, less ambiguity
"The more specific you are, the better the output."
Prototype > PRD
"There's just no way we could have shipped this if we started with static mocks and Figma or if we started with a PRD." Instead, dozens of working prototypes.
Don't micromanage how Claude fixes a bug
"Paste the bug, say 'fix.' Don't micromanage how."
Prototypes instead of design documents
"I'll just ask Claude Code to prototype, like, three versions of it. And I'll try the feature and see which one I like better."
13
Customization
37 settings, 84 environment variables. Make it your tool.
Customize spinner verbs
Star Trek-themed example. Check settings.json into source control.
Output styles: Explanatory / Learning / Custom
Claude explains frameworks; coaches you through changes.
37 settings + 84 environment variables
Use "env" field in settings.json to avoid wrapper scripts.
Claude responds in your language
Japanese, Spanish, German — any language configurable.
Check settings.json into git
Team benefits from your customizations. Support for enterprise-wide policies.
NO_FLICKER Mode
Experimental renderer: no flicker/jump, constant memory/CPU, mouse support, nicer selection. CLAUDE_CODE_NO_FLICKER=1 claude.
14
Headless / SDK
Claude as Unix utility, not a chatbot. cat foo.csv | claude -p "summarize".
claude -p for non-interactive mode
Pass prompt in quotes. Pipeline-friendly.
Best for read-only tests
"That's the place where it works really well."
Start small, then scale
"Start small… test it on one test… iterate on your prompt. Then scale it up to 10."
Pre-commit hook with -p
"Just add a line: claude -p and whatever instruction you have." (Keep under ~5 sec.)
--allow-tools to scope permissions
Lock down specific tools for batch operations.
--bare flag for 10x SDK startup
Skips search for local CLAUDE.md/settings/MCPs (default will flip to --bare in future).
--add-dir (or /add-dir)
Gives Claude access to additional repos. Set "additionalDirectories" in settings.json for the team.
--name "<session>"
Human-readable session names for parallel work.
claude is a Unix utility, not a chatbot
"You can pipe in… cat the CSV, pipe it into claude."
Stack: React Ink + Commander.js + Bun
"We're big fans of Ink… We're also, we use Bun. So big fans of Bun." Transferable for your own CLI tools.
15
Code Review
Reviews were the bottleneck. Now agents do the first sweep — humans approve the merge.
Automatic PR review by an agent team
Each agent focuses on different concerns (logic, security, performance), posts inline comments.
Reviews were the bottleneck
"Code output per Anthropic engineer is up 200% this year, and reviews were the bottleneck."
Boris used it for weeks before launch
"It catches real bugs I wouldn't have noticed otherwise."
16
Cost & ROI
"It's an ROI question, not a cost question." Boris breaks with cost-cutting logic.
ROI question, not a cost question
"If you can make an engineer 50, 70% more productive, that's worth a lot."
~$6/day per active user
"Currently, we're seeing costs around like $6 per day per active user." (Cat Wu, same podcast)
Some Anthropic engineers > $1,000/month
Mainly for code migrations.
Underfund teams, give unlimited tokens
"Don't try to cost-cut at the beginning. Start by giving engineers as many tokens as possible."
1-hour prompt cache
Massive cost reduction for repeated multi-turn agent work.
17
Form-Factor & Workflow
CLI, IDE, GitHub, Slack, Mobile, Web. Boris codes a lot from his phone.
Claude Code has a mobile app
iOS and Android. Boris writes a lot of code from the iOS app. Code tab on the left.
Dispatch — Secure Remote Control
"When I'm not coding, I'm dispatching." Uses MCPs/browser/computer with your permissions.
VS Code / JetBrains / Cursor Extension
Recommended for IDE users: inline diffs, @-mentions, plan review.
GitHub-App / Slack-App / Web
"Every engineer is different and you can use Claude Code the way you want."
4 modes of Claude Code usage
"From the terminal / From the IDE (as an extension) / As a GitHub app so you can @claude right in a GitHub issue comment / Via the SDK, using Claude as a Unix utility"
Adapt workflow to task
"(explore › plan › confirm › code › commit) vs (tests › commit › code › iterate › commit) vs (code › screenshot › iterate)"
18
Migration & Data
Partially migrated codebases confuse humans and models alike. Finish what you start.
Always finish migrations
"Always make sure that when you start a migration, you finish the migration." Partially migrated codebases confuse humans and models.
/batch is the migration tool
Plan interactively, then fan out to dozens of Worktree agents.
Check BigQuery skill into codebase
Entire team uses it for analytics directly from Claude Code.
19
Learning & Onboarding
Output styles, visualizations, ASCII diagrams, spaced repetition. Claude as teacher.
Explanatory / Learning Output Style
Claude explains the why behind changes. Deliberately onboarding-friendly.
Generate visual HTML presentation
"Surprisingly good slides!"
Request ASCII diagrams
Of new protocols, codebases, architectures.
Spaced repetition learning skill
You explain your understanding, Claude asks follow-up questions to fill gaps.
Onboarding workflow at Anthropic
New engineers use Claude Code to navigate codebases — "significantly improving ramp-up time."
20
Context Management
Context rot is real. Boris's rule: every turn is a branching point.
Lower auto-compact threshold
CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 claude to avoid context rot in 1M model.
Context rot starts at 300-400k tokens
Don't let intelligence-sensitive sessions drift beyond that.
"Dumb zone" at ~40% context
Newcomers: stay under 40%, wrap up at 60%. Experienced: under 30%; to 60% only for simple tasks.
Every turn is a branching point
Pick Continue / /rewind / /clear / /compact / Subagent — depending on context needs.
Agentic Search > RAG
"We landed on just agentic search… just using regular code searching, you know, glob, grep." Avoids indexing/sync/security issues at the cost of latency and tokens.
21
Terminal Environment
Ghostty + tmux + voice dictation. The small things that actually move your velocity.
Ghostty Terminal
Team favorite: synchronized rendering, 24-bit color, Unicode.
tmux for one tab per task
Organize terminal tabs by color-coding and names.
Voice dictation (fn x2 on macOS)
"You speak 3x faster than you type, and your prompts get way more detailed as a result."
Chronology — Confirmed Threads & Announcements
From side project in September 2024 to Webby Award in April 2026.
Sep 2024
Boris starts Claude Code as side project at Anthropic.
Late 2024
Internal dogfooding. Robert (colleague) is first power user via FileEdit tool.
24. Feb 2025
Claude Code launches as Research Preview alongside Claude 3.7 Sonnet.
18.–19. Apr 2025
Anthropic Engineering Blog: "Claude Code: Best practices for agentic coding" (introduces ultrathink keyword scale).
07. May 2025
Latent Space Podcast with Cat Wu.
Jun 2025
AI Engineer World's Fair Keynote: "Claude Code & the evolution of agentic coding."
Oct–Nov 2025
Boris stops writing code by hand.
27. Dec 2025
One-year retro: 259 PRs in 30 days, all with Opus 4.5.
02. Jan 2026
13-tip thread "How I use Claude Code" (104K likes; canonical reference).
07. Jan 2026
Claude Code 2.1.0 Announcement.
13. Jan 2026
"Robert with a terminal up" Origin-Story Post.
31. Jan 2026
10 team tips thread (8.5M views).
11. Feb 2026
12 customization tips thread.
19. Feb 2026
Lenny's Podcast Episode: "What happens after coding is solved".
20.–21. Feb 2026
Built-in git worktree support (5 tips).
24. Feb 2026
"Happy 1st birthday to Claude Code".
25. Feb 2026
Anthropic acquires Vercept.
27.–28. Feb 2026
/simplify and /batch announcement.07. Mar 2026
/loop announcement.09. Mar 2026
Code Review Feature Launch.
13. Mar 2026
Opus 4.6 1M default for Max/Team/Enterprise;
/effort max, Voice Mode, Remote Control, PostCompact Hook.23.–25. Mar 2026
Auto Mode,
/schedule, iMessage plugin, Auto-Memory/Dream.29.–30. Mar 2026
15 hidden/underused features thread (user's anchor tweet).
01. Apr 2026
NO_FLICKER Mode announcement.
13. Apr 2026
1-hour prompt cache rollout.
14. Apr 2026
Routines (Boris RT).
16. Apr 2026
6 Opus-4.7 tips thread.
22. Apr 2026
Claude Code wins Webby Award (AI Features & Innovation).
23. Apr 2026
Quality post-mortem for v2.1.116+.
06. May 2026
Code w/ Claude SF: Rate limits doubled, SpaceX 300MW compute deal.
Staged Adoption — Boris's Own Recommendation
From vanilla setup to autonomous agent stack. Clear thresholds for when to move to the next phase.
1
Vanilla
Week 1
Install Claude Code (CLI or VS Code extension).
/init for starter CLAUDE.md. Plan Mode (Shift+Tab twice) for every task. Don't customize.Graduate when: 5+ PRs shipped, Claude makes the same mistake twice — time for rules in CLAUDE.md.
2
Compounding Knowledge
Week 2–3
Treat CLAUDE.md like code: ruthlessly prune, add rule on every mistake. Add 1–2 Slash Commands for daily workflows.
Graduate when: CLAUDE.md > 1 page, Slash Command count ~3.
3
Parallelism
Week 4+
Open 3–5 git worktrees, one Claude in each. PostToolUse Hook for formatting. Pre-allow safe commands with
/permissions, check .claude/settings.json into git.Graduate when: you're frustrated waiting on one Claude.
4
Subagents
Month 2+
Add
.claude/agents/code-simplifier.md and verify-app.md. "Use 5 subagents to explore the codebase" on new repos. Adversarial review pattern.Graduate when: you spend more time switching between Claudes than Claudes spend working.
5
Autonomy
Month 3+
Auto Mode (Opus 4.7),
/focus, recaps on. /loop for PR babysitting, Slack feedback. /schedule for jobs that survive closed laptops. Chrome extension for frontend verification. /voice.End state: Claude Code is no longer a tool — it's your team's compute budget.
Contested Recommendations
Where Boris's setup differs from his own team — and where you need to decide for yourself.
Worktrees vs. Checkouts
Boris personally uses multiple checkouts. The rest of the team prefers worktrees and built native worktree support for it. Both work — choose by mental model.
MCP vs. CLI/Skills
Boris uses Slack MCP. Secondary sources from the team push back: "For read-heavy workflows, a well-crafted CLI skill is often simpler." Rule of thumb: local & simple → Slash Command. External data → MCP.
Opus vs. Sonnet
Boris uses Opus 4.5/4.6/4.7 with Thinking for everything. Team default and Pro tier still surface Sonnet. Rule of thumb: less steering effort with Opus = often faster, despite higher latency.
Auto Mode vs. Explicit Permissions
Auto Mode (Opus 4.7) auto-approves safe permissions. Security-conscious teams prefer manual approval via PermissionRequest Hook. Sandboxes allow middle ground.
Caveats & Notes
Important limitations for honest framing of these tips.
- Many announced features are still rolling out or plan-tier-gated (Auto Mode = Max/Team/Enterprise;
/extra-usagefor 1M context on Pro). Test, don't assume. - The Latent Space Podcast date is listed as May 7, 2025; some recaps say May 8 — same episode.
- "4% of all public GitHub commits" and "200% engineering productivity" are Boris's own numbers (Lenny's Podcast / code review thread), not independently audited. Directional, not precise.
- The fan site howborisusesclaudecode.com (by @CarolinaCherry) is the most complete public catalog (87 tips). Not Anthropic-affiliated — every tip links to the primary @bcherny post.
- Ultrathink semantics changed between CC v1 and v2. In v1, the keyword ladder mapped to literal token budgets; in v2, the on/off Thinking toggle (Tab) replaces most of it. Ultrathink still triggers the highest-effort path.
- Boris left Anthropic briefly for Cursor in early 2025 and came back after two weeks — discussed on Lenny's Podcast. Not a tip, but context for his credibility in trade-off discussions.
- Tips attributed to Thariq, Cat Wu, Noah Z., Erik Schluntz and others from the Claude Code team are included here if Boris retweeted/endorsed them or they appeared in shared podcasts.
- No tip in this report comes from a private/leaked source. Everything is public X, public podcasts, Anthropic's public blog, or the Threads.net mirror.