prompt-smith

Prompt quality management for Claude Code

한국어 문서 (Korean)

Core Features:

• 8-Point Quality Check (LINT / BUILD / Review / Intercept)
• Eval Mode: A/B testing with test datasets
• Long Context: 200K token optimization
• Token Management: Cost estimation & context validation
• Structured Outputs: 100% JSON validity

When to Use Which Mode?

Situation	Recommended	Command
"Is this prompt good enough?"	LINT	/ps:lint <prompt>
"Make this prompt better"	Review	/ps:r <prompt>
"Just fix it and run"	Intercept	/ps:a <prompt>
"Design a new prompt"	BUILD	/ps:build <goal>
"Compare prompt variants"	Eval	See Eval Mode

Quick Start

Commands at a Glance

Command	Description	Best For
/ps:r <prompt>	Review Mode - Show improvements, await approval	Daily use (safe)
/ps:lint <prompt>	LINT Mode - Full diagnosis with test cases	Debugging prompts
/ps:a <prompt>	Intercept Mode - Auto-improve and execute immediately	Quick tasks
/ps:build <goal>	BUILD Mode - Design from requirements	New prompts
/ps:help [topic]	Help - Show usage guide and command reference	Getting started

Quick Example:

/ps:r Write a function to parse JSON

Shows score (3/10 → 8/10), changes, asks "Proceed? (y/n/e)"

---

Option 1: Plugin Install (slash commands enabled)

1. In VS Code, type /plugin to open plugin terminal
2. Press Tab to navigate to "Add Marketplace"
3. Enter: joseph0926/prompt-smith
4. Press Tab to navigate to "Install Plugin"
5. Select ps@prompt-smith

Option 2: Local Plugin (development)

git clone https://github.com/joseph0926/prompt-smith
claude --plugin-dir ./prompt-smith

Option 3: Skill Only (natural language triggers)

git clone https://github.com/joseph0926/prompt-smith
cp -r prompt-smith/skills/prompt-smith ~/.claude/skills/

Note: This method only enables natural language triggers (use prompt-smith -r), not slash commands.

Usage

Slash Commands (recommended):

/ps:r your prompt here            # Review Mode (single line)
/ps:r your prompt here
with multiple lines               # Review Mode (multiline)
/ps:a your prompt here            # Intercept Mode
/ps:lint your prompt              # LINT Mode
/ps:build requirements            # BUILD Mode

Note: Triple backticks (```) are optional. Plain text and multiline input are fully supported.

Natural Language (alternative):

use prompt-smith -r Your prompt here

Real Examples

LINT Mode (Diagnose)

/ps:lint Analyze user feedback

Output: Score 3/10 → Top 3 issues → Improved prompt → 5 test cases

Review Mode (Safe Improvement)

/ps:r Write a function to parse JSON

Output:

+----------------------------------------------------------+
| Score: 3/10 -> 8/10 (+5)                                 |
+----------------------------------------------------------+
Changes:
- [+] Added: Software engineer role
- [+] Added: Error handling examples
- [~] Modified: Specified return format

Proceed? (y/n/e)

Intercept Mode (Auto-fix)

/ps:a Summarize this document

Output: Auto-improved if +2 points possible → Executes immediately

What's New in v3.x

v3.3.1 (Latest)

Feature	Description
Examples & TOCs	Expanded examples and added TOCs to long references
Structured Outputs	Stronger guidance for schema mismatch and JSON mode caution

v3.3.0

Feature	Description
ps:a Token Efficiency	Minimal output: [Prompt Smith] 활성화됨 (X→Y점)
ps:r Intent Capture	AskUserQuestion for format/detail/constraints/goals

v3.2.0

Feature	Description
Eval Mode	A/B prompt comparison with test datasets
Long Context	200K token optimization (chunking, placement)
Token Management	Cost estimation, context window validation
Structured Outputs	100% JSON validity techniques

v3.0.0 - v3.1.0

Feature	Description
Progressive Loading	Skills load on-demand
Hooks	Auto-LINT on prompt submit (optional)
Agents	3 specialized: optimizer, reviewer, test-generator
GitHub Actions	CI/CD integration

See CHANGELOG.md for full release notes.

Modes

Mode	Slash Command	Natural Language	Description
Review	/ps:r <prompt>	use prompt-smith -r	Show improvements, await approval
Intercept	/ps:a <prompt>	use prompt-smith -a	Auto-improve and execute
LINT	/ps:lint <prompt>	lint this prompt	Diagnose existing prompts
BUILD	/ps:build <goal>	build a prompt for...	Design from requirements

8-Point Quality Check

#	Dimension	Score	Note
1	ROLE	0-2	Base
2	CONTEXT	0-2	Base
3	INSTRUCTION	0-2	Base
4	EXAMPLE	0-2	Base
5	FORMAT	0-2	Base
6	SUCCESS_CRITERIA	0-2	Base (v2.7+)
7	STATE_TRACKING	0-2	Extended
8	TOOL_USAGE	0-2	Extended

Base Score: Dimensions 1-6 (max 12 → normalized to 10) Extended Score: All 8 dimensions (normalized)

Documentation

Getting Started

• Quick Start
• First LINT
• First BUILD

Mode Guides

• LINT Mode
• BUILD Mode
• Review Mode
• Intercept Mode
• Eval Mode

Advanced (v3.2.0)

• Long Context Optimization
• Token Management
• Structured Outputs
• Prompt Chaining

Reference

• 8-Point Quality Checklist
• Anti-Patterns
• Technique Priority (Anthropic recommended)
• Claude 4.x Best Practices

Troubleshooting

Slash commands not working

1. Verify plugin: In VS Code, type /plugin → check installed list
2. Restart VS Code or reload Claude Code extension
3. Try reinstalling: Remove plugin, then reinstall

Score seems wrong

• Extended items (STATE_TRACKING, TOOL_USAGE) only apply to multi-step or tool-using prompts
• If N/A, they're excluded from score calculation

Platform

• Claude Code (CLI / VS Code extension)

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

• Report a bug
• Security issues

License

MIT License - see LICENSE