⚡ copt — Claude Prompt Optimizer

Optimize your prompts for Claude 4.x models (4.5, 4.6, and 4.7)

Claude 4.5 models do exactly what you ask — no more, no less. Prompts that worked with Claude 3.x may need adjustment. copt analyzes your prompts for anti-patterns and optimizes them using Claude 4.5 itself. With Claude 4.7, prompts also need to account for literal instruction following, adaptive thinking, and calibrated response length — copt handles these automatically when you target an opus-4.7 model.

---

Installation

Pre-built Binaries

Platform	Download
macOS (Apple Silicon)	copt-macos-arm64
macOS (Intel)	copt-macos-x64
Linux (x64)	copt-linux-x64
Windows (x64)	copt-windows-x64.exe

# macOS / Linux — user-local install (no sudo, XDG-compliant)
mkdir -p ~/.local/bin
curl -L https://github.com/praveenc/copt/releases/latest/download/copt-macos-arm64 -o ~/.local/bin/copt
chmod +x ~/.local/bin/copt

# Ensure ~/.local/bin is on PATH (skip if already there):
#   zsh:  echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && exec zsh
#   bash: echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc && exec bash

copt --version

Prefer a system-wide install?

curl -L https://github.com/praveenc/copt/releases/latest/download/copt-macos-arm64 -o copt
chmod +x copt && sudo mv copt /usr/local/bin/

/usr/local/bin is on PATH by default on macOS and most Linux distros, but requires sudo. On Apple Silicon Homebrew setups, /opt/homebrew/bin is an alternative — but don't write there directly unless the tool is a formal Homebrew formula.

⚠️ macOS Gatekeeper note

If macOS blocks the binary, either:

1. Go to System Settings → Privacy & Security and click "Allow Anyway"
2. Or run: xattr -d com.apple.quarantine ~/.local/bin/copt (replace with /usr/local/bin/copt if you installed system-wide)

From Source

cargo install --git https://github.com/praveenc/copt

---

Quick Start

1. Set Up Credentials

# Bedrock API key (recommended — simplest setup)
export AWS_BEARER_TOKEN_BEDROCK="your-bedrock-api-key"

# Or Anthropic API
export ANTHROPIC_API_KEY="sk-ant-..."

# Or AWS credential chain (SigV4 fallback — no API key needed)
# Uses ~/.aws/credentials, AWS_PROFILE, IAM roles, SSO, etc.

To generate a Bedrock API key: Bedrock console → API keys → Generate long-term key (up to 1 year).

2. Create a Config File (optional)

copt --config-init    # Creates ~/.config/copt/config.toml

Edit the config to set your preferred defaults (including API key):

[default]
provider = "bedrock"
model = "us.anthropic.claude-sonnet-4-5-20250929-v1:0"

[bedrock]
region = "us-west-2"
api_key = "your-bedrock-api-key"   # or use AWS_BEARER_TOKEN_BEDROCK env var

[output]
format = "pretty"
show_diff = false

CLI arguments always take precedence over config values.

3. Optimize a Prompt

copt "Create a dashboard"           # Direct input
copt -f my-prompt.txt               # From file
copt -f prompt.txt --offline        # Analyze only (no API calls)
copt -f prompt.txt -i               # Interactive TUI mode

---

What It Does

copt detects Claude 3.x patterns and rewrites them for Claude 4.5:

Pattern	Problem	Fix
Don't use X	Negative framing	Reframe positively
Can you help...	Indirect command	Direct instruction
NEVER do X	Aggressive caps	Normal casing
Create something	Vague instruction	Add explicit requirements

Example transformation:

- Don't use placeholder data. Can you help me create a dashboard?
+ Use real data from the API. Create an analytics dashboard with:
+ - User metrics visualization
+ - Date range filtering
+ - Export functionality

See docs/RULES.md for the full list of 25 analysis rules across 8 categories.

---

CLI Reference

View full --help output

⚡ Optimize prompts for Claude 4.5 models

Usage: copt [OPTIONS] [PROMPT]

Arguments:
  [PROMPT]  Prompt text to optimize

Options:
  -f, --file <FILE>              Read prompt from file
  -o, --output <FILE>            Save optimized prompt to file
      --output-dir <DIR>         Output directory for auto-save
      --no-save                  Disable auto-save
  -p, --provider <PROVIDER>      Provider: anthropic, bedrock
  -m, --model <MODEL>            Model ID or alias
  -t, --target <FAMILY>          Target Claude family: auto (default), 4.5, 4.6, 4.7
      --region <REGION>          AWS region for Bedrock
      --format <FORMAT>          Output format: pretty, json, quiet
      --diff                     Show before/after diff
      --show-prompt              Display optimized prompt
  -q, --quiet                    Quiet mode (prompt only)
      --analyze                  Analyze only, no optimization
      --offline                  Offline mode (no API calls)
      --check <CAT>              Check specific categories
      --no-suggest               Disable auto-suggestions for vague prompts
  -i, --interactive              Launch full-screen interactive TUI mode
  -e, --editor                   Open editor for multi-line input
      --skip-connectivity-check  Skip connectivity check
      --config-init              Create default config file and exit
  -v, --verbose                  Verbose output
  -h, --help                     Print help
  -V, --version                  Print version

Model Aliases

Use short aliases instead of full Bedrock ARNs:

Alias	Model
sonnet / sonnet-4.5	us.anthropic.claude-sonnet-4-5-20250929-v1:0
opus / opus-4.5	us.anthropic.claude-opus-4-5-20251101-v1:0
opus-4.6	global.anthropic.claude-opus-4-6-v1
sonnet-4.6	global.anthropic.claude-sonnet-4-6
opus-4.7	global.anthropic.claude-opus-4-7
haiku / haiku-4.5	us.anthropic.claude-haiku-4-5-20251001-v1:0

copt -f prompt.txt -m opus            # Use Claude Opus 4.5
copt -f prompt.txt -m haiku           # Use Claude Haiku 4.5

Claude 4.7 Support

Opus 4.7 is generally available as of 2026-04-16. Sonnet 4.7 and Haiku 4.7 are not yet released.

When you target opus-4.7, copt routes optimization through a Claude 4.7-specific meta-prompt that applies eight 4.7 behaviors: literal instruction following (the model does exactly what you say), adaptive thinking guidance (explicit reasoning directives), effort-level awareness (xhigh/high/medium/low/max markers), scratchpad and memory directives (working-memory tags), condensed context (drops 4.5/4.6 scaffolding that 4.7 doesn't need), tone specification (explicit tone markers), response-length calibration (explicit length bounds), and vision-aware instructions (high-resolution image handling up to 2576px/3.75MP with 1:1 pixel coordinates).

Selecting sonnet-4.7 or haiku-4.7 produces a clear "not yet released" error unless you run in --offline mode (offline mode still works for analysis).

Usage:

copt -f prompt.txt -m opus-4.7

Targeting a family explicitly (-t / --target)

By default copt infers the target Claude family from --model (-m). To decouple the rewriter model from the target family — for example, to use a fast, cheap Sonnet 4.5 to produce a prompt tailored for Opus 4.7 — pass -t:

# Opus 4.7 rewrites for Opus 4.7 (auto)
copt -f prompt.txt -m opus-4.7

# Sonnet 4.5 rewrites for Opus 4.7 family best-practices
copt -f prompt.txt -m sonnet -t 4.7

# Opus 4.7 rewrites a prompt that will run on Sonnet 4.5
copt -f prompt.txt -m opus-4.7 -t 4.5

Accepted values for -t: auto (default), 4.5, 4.6, 4.7.

Common Examples

copt -f prompt.txt --offline          # Analyze without API calls
copt -f prompt.txt --diff             # Show before/after diff
copt -f prompt.txt -p anthropic       # Use Anthropic API
copt -f prompt.txt --format json      # JSON output for scripting
copt -f prompt.txt --no-suggest       # Disable suggestions (for CI)
copt --config-init                    # Create default config file

---

Interactive Mode

Launch with -i for a full-screen TUI:

copt -f prompt.txt -i

Keyboard shortcuts:

Key	Action
q	Quit
d	Toggle diff view
c	Copy to clipboard
s / e	Save & open in editor
?	Help

---

Prompt Storage

Optimized prompts are auto-saved to ~/.copt/prompts/YYYY-MM-DD/ with descriptive filenames:

~/.copt/prompts/
└── 2026-02-15/
    ├── dashboard-analytics-api_143022_original.txt
    ├── dashboard-analytics-api_143022_optimized.txt
    └── dashboard-analytics-api_143022.json

Use --output-dir <DIR> to override, or --no-save to disable.

---

Raycast Integration

Optimize prompts directly from your clipboard. See scripts/raycast/README.md for setup.

---

Documentation

• Analysis Rules — All 27 rules across 8 categories
• Migration Guide — Claude 3.x → 4.5 patterns
• Contributing — Development setup
• Rust for Pythonistas — If you're coming from Python

---

License

MIT — Copyright (c) 2026 Praveen Chamarthi

---

_{Built for the Claude developer community ❤️}