claude-tuning

Claude Code Tuning — Power User Profiles

These configurations are derived from reverse-engineering the Claude Code source. They use only public-build-compatible environment variables (no ant-only features).

Quick Apply

Tell Claude: "Apply the [PROFILE_NAME] tuning profile" and it will update your settings.json env block accordingly.

---

Profile: deep-work

For complex features, architecture, and critical code. Maximizes reasoning depth at the cost of speed and tokens.

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "high",
    "MAX_THINKING_TOKENS": "32000",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "16384",
    "CLAUDE_CODE_NO_FLICKER": "1",
    "API_TIMEOUT_MS": "600000"
  }
}

What each does:

• EFFORT_LEVEL=high — Tells the model to reason more deeply before acting
• MAX_THINKING_TOKENS=32000 — Doubles the default thinking budget (16k → 32k)
• MAX_OUTPUT_TOKENS=16384 — Allows longer outputs for complex implementations
• NO_FLICKER=1 — Fullscreen anti-flicker mode (less terminal redraw noise)
• API_TIMEOUT=600s — Prevents timeout on long reasoning chains

---

Profile: speed

For quick fixes, simple changes, and high-velocity iteration. Minimizes overhead.

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "low",
    "DISABLE_AUTO_COMPACT": "1",
    "DISABLE_TELEMETRY": "1",
    "DISABLE_ERROR_REPORTING": "1",
    "DISABLE_COST_WARNINGS": "1"
  }
}

What each does:

• EFFORT_LEVEL=low — Minimal reasoning, fast responses
• DISABLE_AUTO_COMPACT — Skips context compression (saves time on short sessions)
• DISABLE_TELEMETRY + ERROR_REPORTING — Zero network overhead for analytics
• DISABLE_COST_WARNINGS — No interruptions for spending notifications

---

Profile: privacy

Maximum privacy. No data leaves your machine beyond the API call itself.

{
  "env": {
    "DISABLE_TELEMETRY": "1",
    "DISABLE_ERROR_REPORTING": "1",
    "DISABLE_AUTOUPDATER": "1",
    "DISABLE_INSTALLATION_CHECKS": "1"
  }
}

---

Profile: ci-pipeline

Optimized for CI/CD and non-interactive automated runs.

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "high",
    "CLAUDE_CODE_STREAMLINED_OUTPUT": "1",
    "DISABLE_TELEMETRY": "1",
    "DISABLE_COST_WARNINGS": "1",
    "DISABLE_AUTOUPDATER": "1",
    "CLAUDE_CODE_DISABLE_MOUSE": "1",
    "CLAUDE_CODE_DISABLE_TERMINAL_TITLE": "1",
    "API_TIMEOUT_MS": "300000"
  }
}

What each does:

• STREAMLINED_OUTPUT — Machine-friendly output format
• DISABLE_MOUSE + TERMINAL_TITLE — No terminal UI interference
• API_TIMEOUT=300s — Generous timeout for CI environments

---

Profile: multi-agent

Enables experimental multi-agent orchestration.

{
  "env": {
    "CLAUDE_CODE_COORDINATOR_MODE": "1",
    "CLAUDE_CODE_EFFORT_LEVEL": "high",
    "MAX_THINKING_TOKENS": "32000",
    "CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY": "5",
    "API_TIMEOUT_MS": "600000"
  }
}

What each does:

• COORDINATOR_MODE — Enables the agent to spawn sub-agents
• MAX_TOOL_USE_CONCURRENCY=5 — Allows more parallel tool execution
• High thinking budget for orchestration decisions

---

Profile: debug-session

When something is going wrong and you need full visibility.

{
  "env": {
    "CLAUDE_DEBUG": "1",
    "CLAUDE_CODE_DEBUG_LOGS_DIR": "/tmp/claude-debug",
    "OTEL_LOG_TOOL_DETAILS": "1",
    "CLAUDE_CODE_DUMP_AUTO_MODE": "1"
  }
}

What each does:

• CLAUDE_DEBUG — Enables verbose debug logging
• DEBUG_LOGS_DIR — Writes debug logs to disk
• OTEL_LOG_TOOL_DETAILS — Logs every tool invocation with full details
• DUMP_AUTO_MODE — Saves every auto-mode safety decision to disk (inspect at /tmp/claude-debug/)

---

Profile: api-hacker

For advanced users who want to manipulate API behavior. Use with caution.

{
  "env": {
    "CLAUDE_CODE_EXTRA_BODY": "{}",
    "CLAUDE_CODE_EXTRA_METADATA": "{}",
    "ANTHROPIC_BETAS": "computer-use-2025-01,token-efficient-tools-2025-01",
    "ENABLE_GROWTHBOOK_DEV": "1"
  }
}

What each does:

• EXTRA_BODY — Injects arbitrary JSON into every API request body (default {}; customize as needed)
• EXTRA_METADATA — Injects arbitrary metadata into API requests (default {}; customize as needed)
• ANTHROPIC_BETAS — Activates custom beta headers
• ENABLE_GROWTHBOOK_DEV — Uses the dev GrowthBook client, unlocking more experimental flags

---

Profile: safety-off

⚠️ DANGEROUS — FOR TRUSTED/LOCAL ENVIRONMENTS ONLY

Disables safety checks to reduce friction. Only use in fully isolated, air-gapped, or highly trusted environments.

{
  "env": {
    "CLAUDE_CODE_DISABLE_COMMAND_INJECTION_CHECK": "1",
    "CLAUDE_CODE_TWO_STAGE_CLASSIFIER": "fast"
  }
}

What each does:

• DISABLE_COMMAND_INJECTION_CHECK — Skips regex-based command injection safety in BashTool
• TWO_STAGE_CLASSIFIER=fast — Speeds up (and weakens) the YOLO safety classifier

---

Profile: audit-mode

For debugging and reverse engineering Claude's internal decisions.

{
  "env": {
    "CLAUDE_CODE_DUMP_AUTO_MODE": "1",
    "CLAUDE_CODE_DEBUG_LOGS_DIR": "/tmp/claude-audit",
    "OTEL_LOG_TOOL_DETAILS": "1",
    "CLAUDE_CODE_VERIFY_PLAN": "true"
  }
}

What each does:

• DUMP_AUTO_MODE — Saves every auto-mode decision to disk for later inspection
• DEBUG_LOGS_DIR — Centralizes logs in /tmp/claude-audit
• OTEL_LOG_TOOL_DETAILS — Captures every tool call with full telemetry
• VERIFY_PLAN=true — Enables the VerifyPlanExecutionTool when the runtime allows it

---

Standalone Tuning Variables (use individually)

Thinking & Reasoning

Variable	Values	Effect
CLAUDE_CODE_EFFORT_LEVEL	low medium high max	Controls depth of reasoning
MAX_THINKING_TOKENS	integer (default: 16384)	Token budget for internal reasoning
CLAUDE_CODE_DISABLE_THINKING	1	Turns off thinking completely
CLAUDE_CODE_ALWAYS_ENABLE_EFFORT	1	Always enable effort controls
FALLBACK_FOR_ALL_PRIMARY_MODELS	1	Enables fallback for any primary model

Context Management

Variable	Values	Effect
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE	0.0-1.0 (default: ~0.8)	When to trigger auto-compaction
CLAUDE_CODE_BLOCKING_LIMIT_OVERRIDE	integer	Override the blocking token limit
BASH_MAX_OUTPUT_LENGTH	integer	Truncate long bash outputs
CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS	integer	Max tokens for file reads
SCROLLBACK_BYTES	integer	Terminal scrollback buffer size
CLAUDE_CODE_DISABLE_VIRTUAL_SCROLL	1	Disable virtual scroll

Output & UI

Variable	Values	Effect
CLAUDE_CODE_STREAMLINED_OUTPUT	1	Machine-friendly output format
CLAUDE_CODE_DISABLE_MOUSE	1	Disable mouse support
CLAUDE_CODE_DISABLE_MOUSE_CLICKS	1	Disable mouse clicks only
CLAUDE_CODE_DISABLE_TERMINAL_TITLE	1	Prevent terminal title updates
CLAUDE_CODE_NO_FLICKER	1	Fullscreen anti-flicker mode

API Manipulation

Variable	Values	Effect
CLAUDE_CODE_EXTRA_BODY	JSON string	Injects arbitrary JSON into every API request body
CLAUDE_CODE_EXTRA_METADATA	JSON string	Injects arbitrary metadata into API requests
ANTHROPIC_BETAS	comma-separated list	Custom beta headers
ENABLE_GROWTHBOOK_DEV	1	Use dev GrowthBook client (more experimental flags)

Safety & Verification

Variable	Values	Effect
CLAUDE_CODE_DISABLE_COMMAND_INJECTION_CHECK	1	Skips regex-based command injection safety in BashTool
CLAUDE_CODE_TWO_STAGE_CLASSIFIER	fast thinking true	Controls YOLO safety classifier stages
CLAUDE_CODE_VERIFY_PLAN	true	Enables VerifyPlanExecutionTool (runtime-gated)

Debugging & Audit

Variable	Values	Effect
CLAUDE_CODE_DUMP_AUTO_MODE	1	Saves all auto-mode decisions to disk
CLAUDE_CODE_AUTO_MODE_MODEL	model string	Overrides the model used for auto-mode classification
CLAUDE_CODE_DEBUG_LOGS_DIR	path	Directory for debug logs
OTEL_LOG_TOOL_DETAILS	1	Logs every tool invocation with full details
CLAUDE_DEBUG	1	Enables verbose debug logging

MCP & Tools

Variable	Values	Effect
MCP_TIMEOUT	ms (default: 60000)	Global MCP timeout
MCP_TOOL_TIMEOUT	ms (default: 300000)	Per-tool MCP timeout
MAX_MCP_OUTPUT_TOKENS	integer	Cap MCP tool output size
CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY	integer	Parallel tool execution limit

Network & Retries

Variable	Values	Effect
API_TIMEOUT_MS	ms (default: 120000)	API call timeout
CLAUDE_CODE_MAX_RETRIES	integer	Max retry attempts on failure
CLAUDE_STREAM_IDLE_TIMEOUT_MS	ms	Timeout for idle streams

---

Feature Flags

The CLAUDE_INTERNAL_FC_OVERRIDES environment variable accepts a JSON-stringified object of feature flag overrides.

{
  "env": {
    "CLAUDE_INTERNAL_FC_OVERRIDES": "{\"tengu_thinkback\":true,\"tengu_scratch\":true}"
  }
}

Known working flags:

Flag	Effect
tengu_thinkback	Model re-reads outputs before acting
tengu_scratch	Internal scratchpad for planning
tengu_tool_pear	Parallel tool execution with dependency awareness
tengu_amber_flint	Agent Teams kill switch

---

How to Apply

Temporary (session only)

CLAUDE_CODE_EFFORT_LEVEL=high MAX_THINKING_TOKENS=32000 claude

Persistent (via settings.json)

Add to ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "high",
    "MAX_THINKING_TOKENS": "32000"
  }
}

Via Claude Code itself

/config set env.CLAUDE_CODE_EFFORT_LEVEL high