Stats
Actions
Available In
Tags
claude-prompt-optimizer — Claude Code Plugin for Prompt Engineering & AI Tuning
A power-user Claude Code plugin: sharpen prompts, run agent workflows, tune performance profiles, and control hidden internals — all from slash commands.
What is this?
claude-prompt-optimizer is a Claude Code plugin that gives you direct control over Claude's behavior through 15 slash commands, 2 skills, 9 tuning profiles, and 4 CLAUDE.md templates.
It was built by analyzing Claude Code's internal architecture — cataloguing environment variables, reading the system prompt assembly, and studying the hook schema. Every feature here works in the public build.
This is not a prompt rewriter. It is a control layer: precise commands that expose knobs Claude Code ships with but doesn't surface by default.
Skills
Two skills are available as Claude Code knowledge bases:
prompt-craft
Quick reference for writing effective Claude Code prompts. Teaches the [VERB] [WHAT] in [WHERE]. [CONSTRAINT]. formula, identifies what wastes tokens, and clarifies when to add vs. omit context.
claude-tuning
Complete reference for all tuning environment variables, 9 pre-built configuration profiles, feature flag documentation, and instructions for applying settings temporarily or persistently.
Commands
Prompt Engineering
| Command | What it does |
|---|---|
/sharpen <prompt> | Rewrites a vague or verbose prompt into the [VERB] [WHAT] in [WHERE]. [CONSTRAINT]. formula. Shows before/after and explains the change. Does not execute the sharpened prompt. |
Agent-Driven Workflows
| Command | What it does |
|---|---|
/build <what to build> | Full 5-phase development lifecycle: Explore → Plan → Implement → Verify → Report. Spawns subagents for exploration, planning, and adversarial verification. |
/refactor <what to refactor> | Safe 5-phase refactoring: maps blast radius, captures baseline tests, executes one file at a time with checkpoints, verifies all call sites were updated. |
/pipeline [test|lint|typecheck|build|all] | Local CI. Auto-detects project type (Node, Rust, Go, Python) and runs the appropriate validation chain in sequence with pass/fail reporting per stage. |
Performance Tuning
| Command | What it does |
|---|---|
/tune [profile] | Apply a pre-built performance profile. Run without arguments to see all profiles. |
/super | Maximum power mode: EFFORT_LEVEL=max, 32k thinking tokens, 64k output, 10x tool concurrency. |
/chain [wizard|apply <combo>] | Macro wizard. Apply stacked configurations (deep, fast, swarm, audit, god, privacy) in one shot, with confirmation before writing. |
API & Runtime Controls
| Command | What it does |
|---|---|
/inject [status|body JSON|metadata JSON|reset] | Manage CLAUDE_CODE_EXTRA_BODY and CLAUDE_CODE_EXTRA_METADATA — arbitrary JSON injected into every API request. Deep-merges, never overwrites. |
/mode [bedrock|vertex|opus|fast|set MODEL_ID|reset] | Switch model or provider. Supports AWS Bedrock, Google Vertex, and custom model IDs. |
/env [list|set KEY value|unset KEY|reset] | Inspect and manage the env block in ~/.claude/settings.json. |
/flags [list|enable FLAG|disable FLAG|reset] | Control CLAUDE_INTERNAL_FC_OVERRIDES. Known working flags: tengu_thinkback, tengu_scratch, tengu_tool_pear, tengu_amber_flint. |
/swarm [on|off|status] | Enable experimental Agent Teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS), which spawns separate Claude processes that communicate with each other. |
/yolo [on|off|status] | Tune the auto-mode safety classifier: two-stage mode, model override, and disk auditing of every decision. |
/verify [on|off] | Toggle CLAUDE_CODE_VERIFY_PLAN to enable VerifyPlanExecutionTool where available. |
/debug [on|off|dump|status] | Toggle full debug logging: CLAUDE_DEBUG, OTEL_LOG_TOOL_DETAILS, CLAUDE_CODE_DUMP_AUTO_MODE. Logs write to /tmp/claude-debug/. |
Tuning Profiles
Use /tune <profile> to apply any of these to ~/.claude/settings.json:
