What separates people who tinker from those who deliver: actively managing context and mastering advanced session commands.
/context /compact /clear
/btw /fork /loop /schedule
reasoning budget
/clear /compact · opusplan · hooks
competency map
Boris's 15 secrets
A context window is the session’s working memory—and it’s finite. Every file read, search, command output, and failed attempt takes up space in it. When it fills up, Claude has less room to reason about the main task.
In long sessions, quality seems to drop—but it's not the model that gets worse; it's the context that gets cluttered. Anyone who masters /clear e /compact keeps Claude sharp for hours. It's the most underrated skill in Claude Code.
| Command | What it does | When |
|---|---|---|
| /context | Shows how much of the window is in use | To decide when to act |
| /compact | Summarizes the conversation without losing the thread | ~70% full, same task |
| /clear | Clears the context completely | Switched tasks |
Context hygiene: keep the CLAUDE.md lean, avoid dumping huge files into the conversation, and give /clear when changing topics. Clean context = better answers.
In addition to clearing context, Claude Code has commands to control the flow without cluttering the main conversation: parallel questions, branching paths, running on its own in a loop, and automatic permission mode.
Long tasks require context management actively. Instead of asking a secondary question that fills the window, use /btw. Instead of risking a path and derailing the thread, ask a /fork.
| Command | What for |
|---|---|
| /btw | One-turn parallel question — without breaking your flow, but with all the context |
| /fork | Branch the session to explore another path without losing your place |
| /loop | Runs a task at intervals on its own for up to a few days |
| /schedule | Schedules runs (e.g., review PRs every 30 min) |
| /buddy | "Pair programming" mode: reasons alongside you and teaches you instead of just executing |
| auto-mode | Risk classifier: safe actions go through, risky ones are flagged |
# forkar uma sessão pelo CLI: claude --resume <session-id> --fork-session # rodar a tarefa em loop: /loop a cada 30 min, verifique novos PRs e resolva comentários
⚠️ Several of these are skills/commands you install (some from the community, popularized on the channel: /btw /buddy). The context ones (/clear /compact) e o --fork-session are native.
Claude Code recognizes trigger words that increase the reasoning budget (extended thinking). The "stronger" the word, the more tokens Claude spends thinking before taking action.
A difficult problem — architecture audit, major refactor, tricky debugging — gets much easier when you deliberately ask Claude to think more deeply. But reasoning costs tokens; use it only when it’s worth it.
~10,000 reasoning tokens
Triggers: think hard, think deeply, megathink
Medium/large refactors, design documents, structured planning
~32,000 reasoning tokens
Triggers: ultrathink, think harder, think very hard
Full architecture audits, in-depth multistep planning
💡 Tip: say "ultrathink" when you want a deep, multistep plan; "think hard" for detailed refactors. Leave it without a trigger for everything else—it’s faster and cheaper.
Commands that help you improve your own usage of Claude Code: learn the fundamentals, measure how you work, and discover where to improve.
The biggest lever isn’t a standalone command; it’s the improvement loop: measure → adjust rules and hooks → repeat. Those who do this stay months ahead of those who just “use” it.
The 2 biggest levers: clear context when switching tasks, compact when you pass ~70% of the /context. Always a sharp session.
Opus plans, Sonnet executes (/model opusplan). Quality in planning, low cost in execution.
Adjust output styles and permissions in /config; automate the agent cycle with hooks in the settings.json.
Deep planning + extended reasoning before touching the code — a popular custom skill on the channel.
Continuous improvement loop: measure how you work (your logs stay in ~/.claude/projects) → turn what you've learned into rules in the CLAUDE.md and hooks → save it in a second brain (track 6) → repeat.
One competency map to find your way around: the pillars of effective Claude Code use and the levels every user moves through—from “improved chat” to “agent orchestrator.”
Without a roadmap, you stay stuck at level 1 without realizing it. Knowing the next step (e.g., "I don’t use subagents yet") turns random use into deliberate progress.
Conversation — use it like a chat in the terminal, with targeted edits
Context — CLAUDE.md, /clear and /compact, plan mode
Extension — skills, MCP, hooks, custom commands
Parallel — subagents, worktrees, several Claudes at the same time
Orchestration — agent teams, scheduled automations, Cowork (track 6)
The 6 pillars that support every level.
A collection of rarely used resources — including the famous thread from Boris Cherny, creator of Claude Code, with his 15 favorite hidden features. This is where the "real leverage" is.
Most people have never used hooks, worktrees, or the mobile app. These features aren’t just for show: they’re what separates people who run one Claude for those who run it dozens in parallel.
| Resource | What it does |
|---|---|
| Mobile app | "Code" tab in the Claude app (iOS/Android) — full sessions from your phone |
| /loop + /schedule | Run automatically at intervals — track PRs, code review (skills) |
| Hooks | Deterministic logic in the agent cycle (SessionStart, PreToolUse, Stop...) |
| git worktree | Multiple Claudes in parallel, each on a branch/worktree — no conflicts |
| --fork-session | Resume by branching: claude --resume <id> --fork-session |
| .claude/commands | Custom slash commands — automate recurring team workflows |
| --add-dir | Access multiple repositories without switching sessions |
| /agents | Subagents with their own prompts and tools (.claude/agents/) — track 6 |
| /mcp | Connects external tools via MCP servers — Track 5 |
| claude -p | Headless mode (print) — run Claude in scripts and pipelines |
🌐 Boris’s #1 tip: give Claude a way to verify its own results — the Chrome extension for web coding. “If you don’t give it a browser, the visual result probably won’t be good.”