Claude Code

Total Entries: 0

Category	Surface / Availability	Name	Syntax	Accepted Values / Range	Example	Description & Notes
1. Session & Startup	Claude Code	Interactive session	`claude ["query"]`	Optional opening prompt	`claude "explain this project"`	Starts the REPL in the current directory. Pass a query to open with that prompt already sent.
1. Session & Startup	Claude Code	Print / headless	`--print, -p`	Toggle + query string	`cat logs.txt \| claude -p "explain"`	Runs one query, writes to stdout, exits. The entry point for every scripted or CI use. Accepts piped stdin.
1. Session & Startup	Claude Code	Continue	`--continue, -c`	Toggle	`claude -c -p "check for type errors"`	Loads the most recent conversation in the current directory, including sessions that added it via `/add-dir`.
1. Session & Startup	Claude Code	Resume	`--resume, -r`	Session ID, name, or blank for picker	`claude -r "auth-refactor" "finish this PR"`	Resumes a specific session, or opens an interactive picker. Background sessions appear marked `bg`.
1. Session & Startup	Claude Code	Session name	`--name, -n`	Text label	`claude -n "my-feature-work"`	Names the session for `/resume` and the terminal title. Rename mid-session with `/rename`.
1. Session & Startup	Claude Code	Fork on resume	`--fork-session`	Toggle (use with -r / -c)	`claude --resume abc123 --fork-session`	Creates a new session ID instead of reusing the original, leaving the source conversation untouched.
1. Session & Startup	Claude Code	Worktree	`--worktree, -w`	Name, `#PR`, or PR URL	`claude -w feature-auth --tmux`	Runs in an isolated git worktree under `.claude/worktrees/` so parallel sessions don't collide.
1. Session & Startup	Claude Code	Maintenance commands	`claude update \| doctor \| install`	`stable`, `latest`, or a version	`claude install stable`	`doctor` prints read-only install and settings diagnostics without starting a session.
1. Session & Startup	Claude Code	Authentication	`claude auth login\|logout\|status`	`--email`, `--sso`, `--console`, `--text`	`claude auth login --console`	`status` exits 0 when signed in, 1 when not — useful as a CI precondition check.
1. Session & Startup	Claude Code	CI token	`claude setup-token`	Toggle	`claude setup-token`	Generates a long-lived OAuth token for CI and scripts. Printed once, never saved. Requires a Claude subscription.
2. Model & Reasoning	Claude Code	Model	`--model`	`opus`, `sonnet`, `haiku`, `fable`, or full ID	`claude --model claude-sonnet-5`	Sets the session model. Overrides the `model` setting and `ANTHROPIC_MODEL`. Aliases resolve to the latest in that tier.
2. Model & Reasoning	Claude Code	Effort level	`--effort`	low, medium, high, xhigh, max, ultracode	`claude --effort xhigh`	How much reasoning the model spends per turn. Available levels depend on the model. `ultracode` starts at xhigh with automatic workflow orchestration.
2. Model & Reasoning	Claude Code	Fallback chain	`--fallback-model`	Comma-separated model list	`claude --fallback-model sonnet,haiku`	Tried in order when the primary model is overloaded or retired. Persist it with the `fallbackModel` setting.
2. Model & Reasoning	v2.1.210+	Advisor	`--advisor`	`opus`, `sonnet`, or a full model ID	`claude --advisor opus`	Pairs a stronger second model that Claude consults at key decision points. Fable 5 isn't offered as an advisor.
2. Model & Reasoning	Claude Code	Subagents	`--agent, --agents`	Agent name, or inline JSON definitions	`claude --agent my-custom-agent`	`--agents` defines subagents on the fly using subagent frontmatter fields plus a `prompt` field.
3. Permissions & Safety	Claude Code	Permission mode	`--permission-mode`	default (manual), acceptEdits, plan, auto, dontAsk, bypassPermissions	`claude --permission-mode plan`	Sets the starting mode; cycle live with `Shift+Tab`. Overrides `defaultMode` in settings.
3. Permissions & Safety	Claude Code	Allow rules	`--allowedTools`	Permission-rule patterns	`--allowedTools "Bash(git log *)" "Read"`	Tools that run without prompting. Same syntax as `permissions` in settings.json — `Bash(*)` does not mean "match all Bash".
3. Permissions & Safety	Claude Code	Deny rules	`--disallowedTools`	Tool names or scoped rules	`--disallowedTools "Bash(rm )" "mcp__"`	A bare name removes the tool from context entirely; a scoped rule keeps the tool and denies only matching calls.
3. Permissions & Safety	Claude Code	Restrict toolset	`--tools`	`""`, `"default"`, or a name list	`claude --tools "Bash,Edit,Read"`	Limits built-in tools. Does not affect MCP tools — pair with `--disallowedTools "mcp__*"` to cover those too.
3. Permissions & Safety	Claude Code	Extra directories	`--add-dir`	One or more paths	`claude --add-dir ../apps ../lib`	Grants file access only — most `.claude/` configuration is not discovered from added directories.
3. Permissions & Safety	v2.1.169+	Safe mode	`--safe-mode`	Toggle	`claude --safe-mode`	Disables every customization (CLAUDE.md, skills, plugins, hooks, MCP, themes) to isolate a broken config. Auth, models, tools, and permissions still work.
3. Permissions & Safety	Claude Code	Skip permissions	`--dangerously-skip-permissions`	Toggle	`claude --dangerously-skip-permissions`	Equivalent to `--permission-mode bypassPermissions`. Reserve for sandboxed containers and disposable CI runners.
4. Headless & Scripting	Claude Code	Output format	`--output-format`	text, json, stream-json	`claude -p "list endpoints" --output-format json \| jq`	`json` returns one structured envelope with cost, duration, and turn count; `stream-json` emits live events.
4. Headless & Scripting	Claude Code	Input format	`--input-format`	text, stream-json	`claude -p --input-format stream-json`	Lets you feed structured turns in rather than a single prompt string. Print mode only.
4. Headless & Scripting	Claude Code	Bare mode	`--bare`	Toggle	`claude --bare -p "does this handle null?"`	Skips auto-discovery of hooks, skills, plugins, MCP servers, auto memory, and CLAUDE.md. Leaves Bash, read, and edit available. Fastest, most reproducible scripted start.
4. Headless & Scripting	Claude Code	Turn limit	`--max-turns`	Integer (no limit by default)	`claude -p --max-turns 3 "query"`	Caps agentic turns in print mode and exits with an error at the limit.
4. Headless & Scripting	v2.1.217+	Spend cap	`--max-budget-usd`	Dollar amount	`claude -p --max-budget-usd 5.00 "query"`	Stops once API spend reaches the cap. Subagent spend counts toward it, and running background subagents are stopped.
4. Headless & Scripting	v2.1.205+	Structured output	`--json-schema`	JSON Schema string	`claude -p --json-schema '{"type":"object",...}' "query"`	Returns validated JSON matching your schema after the agent finishes. Invalid schemas now exit with an error rather than falling back silently.
4. Headless & Scripting	Claude Code	Streaming detail	`--include-partial-messages`	Toggle	`claude -p --output-format stream-json --verbose --include-partial-messages "q"`	Emits token-level events. Requires `--print` and `--output-format stream-json`.
4. Headless & Scripting	Claude Code	Cache-friendly prompt	`--exclude-dynamic-system-prompt-sections`	Toggle	`claude -p --exclude-dynamic-system-prompt-sections "query"`	Moves per-machine sections (cwd, env, memory paths, git flag) into the first user message so the prompt cache is reused across users and machines.
5. System Prompt	Claude Code	Append text	`--append-system-prompt`	Text string	`claude --append-system-prompt "Always use TypeScript"`	Preferred for scripts: keeps default tool guidance, safety instructions, and coding conventions, and adds only what differs.
5. System Prompt	Claude Code	Append file	`--append-system-prompt-file`	File path	`claude --append-system-prompt-file ./style-rules.txt`	Same as above but sourced from a file you can version-control and share.
5. System Prompt	Claude Code	Replace prompt	`--system-prompt`	Text string	`claude --system-prompt "You are a Python expert"`	Drops the entire default prompt including tool guidance and safety instructions. Use only when the identity and permission model genuinely differ.
5. System Prompt	Claude Code	Replace from file	`--system-prompt-file`	File path	`claude --system-prompt-file ./prompts/review.txt`	Mutually exclusive with `--system-prompt`. Append flags can combine with either replacement flag.
5. System Prompt	v2.1.205+	Subagent prompt	`--append-subagent-system-prompt`	Text string	`claude -p --append-subagent-system-prompt "Cite file paths" "q"`	Appends to every subagent's prompt, including nested ones. Print mode only.
6. Parallel & Background	Claude Code	Background session	`--bg, --background`	Toggle + prompt	`claude --bg "investigate the flaky test"`	Starts detached and returns immediately with a session ID. Cannot be combined with `-p`.
6. Parallel & Background	Claude Code	Shell job	`--exec`	Shell command string	`claude --bg --exec 'pytest -x'`	Runs a PTY-backed background job instead of a Claude session — useful for parking long test runs next to your agents.
6. Parallel & Background	Claude Code	Agent view	`claude agents`	`--json`, `--all`, `--cwd`	`claude agents --json`	One screen showing every parallel session and which ones need input. `--json` makes it scriptable.
6. Parallel & Background	Claude Code	Session control	`claude attach\|logs\|stop\|respawn\|rm`	Session ID	`claude attach 7c5dcf5d`	Manage background sessions from the shell. `respawn --all` restarts every running session, e.g. to pick up a new binary.
6. Parallel & Background	v2.1.186+	Teammate display	`--teammate-mode`	in-process (default), auto, tmux, iterm2	`claude --teammate-mode tmux`	How agent-team teammates render. The default changed from `auto` in v2.1.179.
7. Extensions & Config	Claude Code	MCP servers	`--mcp-config, --strict-mcp-config`	JSON file paths or inline JSON	`claude --strict-mcp-config --mcp-config ./mcp.json`	Loads servers from file. The strict flag ignores all other MCP configuration — the reproducible choice for CI.
7. Extensions & Config	v2.1.186+	MCP auth	`claude mcp login\|logout <name>`	Server name, `--no-browser`	`claude mcp login sentry --no-browser`	Runs a server's OAuth flow from the shell without opening `/mcp`. Use `--no-browser` over SSH.
7. Extensions & Config	Claude Code	Plugins	`--plugin-dir, --plugin-url`	Directory, `.zip`, or URL	`claude --plugin-dir ./my-plugin`	Session-scoped plugin loading. Repeat the flag for multiple plugins.
7. Extensions & Config	Claude Code	Settings override	`--settings, --setting-sources`	File path or inline JSON; `user,project,local`	`claude --settings '{"fastMode": true}'`	Overrides matching keys in settings.json for this session; omitted keys keep their file values. Max 2 MiB.
7. Extensions & Config	Claude Code	Browser & IDE	`--chrome, --ide`	Toggle (`--no-chrome` to disable)	`claude --chrome`	Chrome integration for web automation and testing; `--ide` auto-connects when exactly one valid IDE is available.
7. Extensions & Config	Claude Code	Debugging	`--debug, --debug-file`	Category filter, e.g. `"api,mcp"` or `"!statsig"`	`claude --debug "api,hooks"`	Category filtering keeps the log readable. `--debug-file` implicitly enables debug mode.
8. Slash Commands	In-session	Initialize project	`/init`	None	`/init`	Generates a starter `CLAUDE.md`. Set `CLAUDE_CODE_NEW_INIT=1` for a guided flow covering skills, hooks, and personal memory.
8. Slash Commands	In-session	Inspect context	`/context [all]`	`all` expands the breakdown	`/context all`	Colored grid of what's filling the window, with warnings for context-heavy tools and memory bloat. Run this before compacting.
8. Slash Commands	In-session	Compact	`/compact [instructions]`	Optional focus instructions	`/compact keep the auth flow and test results`	Summarizes the conversation to free space while staying on the same task. Focus instructions control what survives.
8. Slash Commands	In-session	Clear	`/clear [name]`	Optional label for the old session	`/clear auth-work`	Empty context, project memory intact. Aliases `/reset`, `/new`. Recover the prior conversation with `/resume` or the rewind menu.
8. Slash Commands	In-session	Plan mode	`/plan [description]`	Optional task description	`/plan fix the auth bug`	Enters plan mode straight from the prompt so Claude researches and proposes before touching files.
8. Slash Commands	In-session	Model & effort	`/model, /effort`	Model name; low → ultracode, or `auto`	`/effort xhigh`	Both apply immediately without waiting for the current response. In `/model`, press `s` to switch for this session only.
8. Slash Commands	In-session	Code review	`/code-review [level] [--fix]`	low…max, `ultra`; `--fix`, `--comment`	`/code-review high --fix`	Reviews the diff for correctness bugs plus reuse and simplification cleanups. `ultra` runs a deep multi-agent review in the cloud.
8. Slash Commands	In-session	Security review	`/security-review`	None	`/security-review`	Checks the working diff for security vulnerabilities specifically, separate from the correctness pass.
8. Slash Commands	In-session	Side question	`/btw [question]`	Question text, or blank to reopen	`/btw what was that config file called?`	Ephemeral answer in an overlay — sees the full conversation, has no tools, never enters history. Works while Claude is mid-turn.
8. Slash Commands	In-session	Rewind	`/rewind`	None (or `Esc` `Esc`)	`/rewind`	Rolls code and conversation back to a checkpoint, or summarizes part of the conversation. Can restore context from before a `/clear`.
8. Slash Commands	In-session	Goal	`/goal [condition\|clear]`	Completion condition	`/goal all tests pass and lint is clean`	Claude keeps working across turns until the condition holds. `clear`, `stop`, or `cancel` ends it early.
8. Slash Commands	In-session	Loop	`/loop [interval] [prompt]`	Interval like `5m`; omit to self-pace	`/loop 5m check if the deploy finished`	Repeats a prompt while the session stays open. Omit the prompt to run `.claude/loop.md`. Alias `/proactive`.
8. Slash Commands	In-session	Batch	`/batch <instruction>`	Change description	`/batch migrate src/ from Solid to React`	Decomposes a codebase-wide change into 5–30 units, then runs one background subagent per unit in its own worktree, each opening a PR. Requires git.
8. Slash Commands	In-session	Branch / fork	`/branch, /fork, /background`	Optional name or prompt	`/fork try the Redis approach instead`	`/branch` switches you into a copy; `/fork` runs a copy as a separate background session; `/background` detaches this one and frees the terminal.
8. Slash Commands	In-session	Memory	`/memory`	None	`/memory`	Edit `CLAUDE.md` files, toggle auto-memory, and review what Claude has learned automatically.
8. Slash Commands	In-session	Permissions	`/permissions`	None	`/permissions`	Interactive allow / ask / deny rule editor by scope, plus working directories and recent auto-mode denials. Alias `/allowed-tools`.
8. Slash Commands	v2.1.181+	Config	`/config [key=value ...]`	Any settable key	`/config theme=dark model=sonnet`	Sets settings without opening the UI. Run `/config --help` to list every key. Alias `/settings`.
8. Slash Commands	In-session	Diagnostics	`/doctor, /debug, /usage`	Optional issue description	`/doctor`	`/doctor` checks install health, trims bloated CLAUDE.md files, and flags unused skills and slow hooks before changing anything. `/cost` is an alias of `/usage`.
8. Slash Commands	In-session	Diff & export	`/diff, /export, /copy`	Optional filename; `/copy N`	`/export review-notes.txt`	`/diff` opens an interactive viewer for uncommitted and per-turn changes; `/copy` picks individual code blocks.
9. Keyboard & Input	In-session	Cycle permission mode	`Shift+Tab`	`Alt+M` on some Windows runtimes	`Shift+Tab → plan mode`	Cycles Manual → acceptEdits → plan, plus any modes you've enabled such as auto or bypassPermissions.
9. Keyboard & Input	In-session	Interrupt	`Esc`	Single press	`Esc → redirect mid-turn`	Stops the response or tool call mid-turn and keeps the work done so far. Closes a dialog instead when one is open.
9. Keyboard & Input	In-session	Rewind menu	`Esc Esc`	Double press on empty prompt	`Esc Esc → checkpoint list`	On an empty prompt, opens the rewind menu. With text in the prompt, clears the draft and saves it to history.
9. Keyboard & Input	In-session	Transcript viewer	`Ctrl+O`	Toggle	`Ctrl+O → full tool detail`	Shows detailed tool usage with timestamps and per-message model, and expands collapsed MCP calls.
9. Keyboard & Input	In-session	History search	`Ctrl+R`	`Ctrl+S` cycles scope in fullscreen	`Ctrl+R → search past prompts`	Reverse-searches prompt history across all projects. In fullscreen it opens a dialog scoped to session, project, or everything.
9. Keyboard & Input	In-session	Background a task	`Ctrl+B`	Press twice in tmux	`Ctrl+B → detach the build`	Moves a running Bash command or agent to the background so the turn continues.
9. Keyboard & Input	In-session	External editor	`Ctrl+G`	`Ctrl+X Ctrl+E` also works	`Ctrl+G → write prompt in $EDITOR`	Composes long prompts in your default editor. Optionally prepends Claude's last reply as commented context.
9. Keyboard & Input	In-session	Toggle thinking / fast	`Option+T, Option+O`	`Alt+` on Windows and Linux	`Option+T → extended thinking off`	Thinking toggle has no effect on Fable 5, which always uses extended thinking. `Option+P` switches model without clearing your prompt.
9. Keyboard & Input	In-session	Input prefixes	`/ ! @ :`	First character of the prompt	`! npm test`	`/` command, `!` shell mode, `@` file path autocomplete, `:` emoji shortcode. Shell output lands in context and Claude responds to it.
9. Keyboard & Input	In-session	Multiline input	`Shift+Enter, \+Enter, Ctrl+J`	Terminal dependent	`Ctrl+J → newline anywhere`	`Ctrl+J` works in any terminal with no setup. For VS Code, Cursor, Alacritty, and Zed, run `/terminal-setup` to install the Shift+Enter binding.
10. Claude API	Messages API	Model	`model`	claude-opus-5, claude-sonnet-5, claude-haiku-4-5-20251001, claude-fable-5, claude-opus-4-8	`model="claude-opus-5"`	Required. Opus for complex agentic work, Sonnet for balanced throughput, Haiku for fast and cheap.
10. Claude API	Messages API	Max tokens	`max_tokens`	Integer, model dependent	`max_tokens=16000`	Required. Hard ceiling on total output for the turn, thinking tokens included.
10. Claude API	Messages API	Messages	`messages`	Array of `{role, content}`	`[{"role":"user","content":"Hello"}]`	The API is stateless — send the full conversation every request. Content can mix text, images, documents, and tool results.
10. Claude API	Messages API	System prompt	`system`	String or content-block array	`system="You are a technical editor."`	Top-level, not a message role. Use the array form to attach `cache_control` breakpoints for prompt caching.
10. Claude API	Messages API	Thinking	`thinking`	`{"type":"adaptive"\|"enabled"\|"disabled"}`, `display`, `budget_tokens`	`thinking={"type":"adaptive","display":"summarized"}`	Adaptive lets Claude decide how much to reason and is the path forward on newer models. With manual `enabled`, `budget_tokens` must be ≥1024 and below `max_tokens`.
10. Claude API	Messages API	Tools	`tools`	Array of `{name, description, input_schema}`	`tools=[weather_tool]`	Description quality is the single biggest driver of tool-use accuracy: say what it does, when not to use it, and what each parameter means.
10. Claude API	Messages API	Tool choice	`tool_choice`	auto, any, tool, none	`tool_choice={"type":"tool","name":"get_weather"}`	With extended thinking, only `auto` and `none` are supported. `disable_parallel_tool_use` forces one call at a time.
10. Claude API	Messages API	Streaming	`stream`	true / false (default false)	`stream=True`	Server-sent events: `message_start` → content blocks → `message_delta` → `message_stop`. Token counts in `message_delta` are cumulative.
10. Claude API	Messages API	Stop sequences	`stop_sequences`	Array of strings	`stop_sequences=["\n\nHuman:"]`	Ends generation when a sequence appears. The matched value comes back in `stop_sequence` on the response.
10. Claude API	Messages API	Beta headers	`betas`	e.g. `interleaved-thinking-2025-05-14`	`betas=["interleaved-thinking-2025-05-14"]`	Interleaved thinking lets Claude think between tool calls, and is the one case where `budget_tokens` may exceed `max_tokens`. Newer models enable it automatically with adaptive thinking.
11. Deprecated & Removed	Legacy — API	Sampling controls	`temperature, top_p, top_k`	Default value only on newer models	`temperature=0.3 # 400 error`	Not supported on Claude 4.7 and later or Mythos Preview — a non-default value returns 400. Omit them and steer behavior through the prompt instead.
11. Deprecated & Removed	Legacy — API	Response prefill	`assistant turn as last message`	Removed on 4.6+	`{"role":"assistant","content":"The answer is ("}`	Returns 400 on Claude 4.6 and later and on Mythos Preview. Use structured outputs or system-prompt instructions to shape the response instead.
11. Deprecated & Removed	Removed v2.1.111	Auto mode flag	`--enable-auto-mode`	—	`claude --permission-mode auto`	Auto mode is now in the `Shift+Tab` cycle by default. Use `--permission-mode auto` to start in it.
11. Deprecated & Removed	Removed v2.1.91	PR comments	`/pr-comments`	—	`"show me the review comments on this PR"`	Just ask Claude directly now. Related: `/review` still gives a fast read-only pass over a GitHub PR.
11. Deprecated & Removed	Deprecated alias	Remote flag	`--remote`	Task description	`claude --cloud "Fix the login bug"`	Deprecated alias for `--cloud`, which creates a web session on claude.ai. Pull it back down locally with `--teleport`.

Prompt Architecture & Working Vocabulary

[Context & Role] + [Concrete Task] + [Constraints & Guardrails] + [Reference Material] + [Output Contract]

1. Context & Role

Say what world this lives in and who's asking (e.g. this is a Rails 7 monolith; I'm the on-call engineer, it's 2am). Persistent context belongs in CLAUDE.md, not every prompt.

2. Concrete Task

One verb, one outcome. Find why the checkout webhook times out beats look at the webhook. Vague asks produce vague diffs.

3. Constraints & Guardrails

State the boundaries up front (e.g. don't change public APIs, smallest diff that passes tests, ask before installing anything).

4. Reference Material

Point at real artifacts: @src/auth.ts, a pasted stack trace, ! git diff. Specific beats descriptive every time.

5. Output Contract

Name the shape you want back — a plan, a unified diff, a JSON object, three options ranked. Ambiguous format is the most common cause of rework.

Task Verbs (A–Z)

Audit Benchmark Bisect Consolidate Document Extract Instrument Migrate Profile Refactor Reproduce Trace

Constraint Phrases (A–Z)

Ask before editing Don't add dependencies Don't touch public APIs Match existing patterns No new abstractions Smallest diff that works Stop after the plan Tests must stay green

Output Contracts (A–Z)

Bulleted list, no prose Comparison table JSON matching this schema One paragraph, no headers Plan first, then confirm Ranked shortlist of 3 Unified diff only

Reasoning Depth (A–Z)

Consider the failure modes Flag what you're unsure about List 3 approaches, then pick Read before you write State assumptions first What would break this?

Verification Demands (A–Z)

Cite file:line Prove it with a script Reproduce before fixing Run the tests, show output Show the failing case Verify, don't assume

Context Anchors (A–Z)

@path/to/file ! git diff ! npm test /context CLAUDE.md Pasted stack trace Screenshot paste

Top 20 High-Signal CLAUDE.md Entries (A–Z)

Branch naming convention Build command Commit message format Deploy is manual — never trigger Directories that are generated Error-handling convention Files never to edit Formatter and lint command Known flaky tests Local dev setup gotchas Logging convention Migration workflow Package manager (npm vs pnpm) Preferred test framework Secrets live in X, never inline Single test command Style rules that differ from defaults Test command Type-check command Where new modules go

Cool Tips & Tricks

Diagnose Before You Compact

/compact is a blunt instrument. Run /context first to see what's actually eating the window — it's often one enormous file read or an MCP server you forgot you connected. /context all → /compact keep the auth flow and test results Same task → compact. Different task → clear.

Bare Mode for Scripts

Skipping auto-discovery of hooks, skills, plugins, MCP servers, and CLAUDE.md makes scripted calls start noticeably faster and behave reproducibly across machines. claude --bare -p "does this handle null inputs?" < src/parse.ts Docs note this will become the default for -p.

Plan Before Big Changes

For anything spanning more than a couple of files, enter plan mode first so Claude researches and proposes before editing. Cheaper to redirect a plan than to unwind a diff. Shift+Tab → plan | /plan fix the auth bug

/btw vs. Subagent

They're exact inverses. /btw sees your whole conversation but has no tools — for asking about what Claude already knows. A subagent has full tools but starts empty — for going and finding out something new. /btw which file did you say held the retry logic? Both keep the main context clean.

Parallel Work Without Collisions

Worktrees give each session its own checkout, so two agents can't stomp each other's files. Pair with claude agents to watch them all from one screen. claude -w feature-auth --bg "implement the token refresh"

Shell Mode Round-Trip

Prefix with ! to run a command directly. The output lands in context and Claude responds to it — so you get an explanation of the failure without a second prompt. ! npm test Set respondToBashCommands: false to silence the auto-response.

Cap Spend in CI

Unattended runs deserve a hard ceiling. Subagent spend counts toward the cap, and running background subagents get stopped when it's reached. claude -p --max-budget-usd 3.00 --max-turns 10 "triage this failure"

When Config Goes Weird

--safe-mode disables every customization at once, so you can tell in one run whether a plugin, hook, or skill is the culprit. Unlike --bare, permissions and models still work normally. claude --safe-mode Then narrow it down with /doctor.

Project Standards & Framework

Rule 1: CLAUDE.md Discipline

CLAUDE.md loads into every session, so every line costs context forever. Keep only what Claude can't work out by reading the code:

Keep the non-derivable: pitfalls, rationale, and conventions that differ from tool defaults. Claude can read your directory layout; it can't guess that make test is broken and you use pytest -x.
Cut the derivable: directory listings, dependency lists, and architecture overviews belong in the codebase, not in permanent context. /doctor will offer to trim these for you.
Push detail down: move on-demand guidance into skills and nested per-package CLAUDE.md files so it loads only when relevant.
Never mid-session: CLAUDE.md edits don't apply until the next session — the prompt cache is already built. Use /memory to edit and restart to apply.

Rule 2: Permission Posture by Environment

Match the mode to the blast radius, not to your patience:

Exploring an unfamiliar repo: plan mode. Claude reads and proposes, nothing is written.
Normal feature work: acceptEdits or auto, with a project allowlist so read-only commands stop prompting. Build it with /fewer-permission-prompts.
Unattended CI: --strict-mcp-config, an explicit --tools list, --max-turns, and --max-budget-usd. Constrain the agent rather than trusting the prompt.
bypassPermissions: containers and disposable runners only. It skips the checks that stop a bad command, and the mode persists across supervisor restarts for background sessions.

Rule 3: Session Hygiene

Context quality degrades long before the window is full. Defaults worth adopting:

New task, new session. /clear beats fighting drift from an unrelated conversation. Project memory survives.
Name sessions you'll come back to with -n or /rename, so /resume is a list of tasks rather than a list of timestamps.
Branch instead of arguing. Trying a second approach? /branch preserves the first so you can compare rather than reconstruct.
Review before you ship: /diff to see it, /code-review for correctness, /security-review for vulnerabilities.

Standard CLAUDE.md Template

# Project: [name] ## Commands - Test: [single command] - Type-check: [single command] - Lint: [single command] - Build: [single command] ## Conventions that differ from defaults - [e.g. we use tabs, not spaces — the formatter enforces it] - [e.g. errors are returned, never thrown, outside the HTTP layer] - [e.g. new modules go in src/features/<name>/, not src/lib/] ## Gotchas - [e.g. the auth tests are flaky in CI; rerun once before investigating] - [e.g. never edit db/schema.rb by hand — it's generated by migrations] ## Do not - Deploy, or run anything under scripts/prod/ - Add dependencies without asking - Touch [generated directory]