Lwy Prompt Optimizer

Prompt Optimizer

A pre-flight prompt checker that analyzes, critiques, and improves prompts before they are sent to an AI system. It plays the role of a coach — teaching users to write better prompts by showing what can be improved across 7 key dimensions.

Core principle: analyze before acting. This skill is invoked at the start of a conversation, whenever the user's message looks like a prompt or instruction for an AI system. Analyze first, then improve, then let the user decide.

Prerequisites

• No runtime dependencies (a pure methodology skill, no scripts)
• Works for any prompt — code generation, writing, analysis, creative tasks

When to use

Triggers:

• User says "optimize my prompt", "improve this prompt", "check my prompt", "review my prompt"
• User says "make this more specific", "help me write a better prompt"
• User provides a raw requirement and asks for help refining it
• User is about to send a long instruction to an AI and wants feedback first

Do not trigger when:

• The user is having an ordinary conversation
• The user asks the AI to do something (rather than improve a prompt)
• The input is clearly code, a file path, or a technical command

7-Dimension Analysis Framework

Every prompt is analyzed across these 7 dimensions:

Score for each dimension:

• ✅ Strong — well covered, no action needed
• ⚠️ Weak — partially addressed, can be improved
• ❌ Missing — not addressed at all, must be added

Workflow

[1. Receive draft prompt]
       ↓
[2. Dimension analysis] → score each dimension (✅ strong / ⚠️ weak / ❌ missing)
       ↓
[3. Show the critique card] → a structured report with scores
       ↓
[4. Suggest improvements] → a concrete rewrite for each weak dimension
       ↓
[5. Show the optimized prompt] → the full rewritten version
       ↓
[6. User decision] → use original / use optimized / edit manually

Response Format

When analyzing a prompt, always respond with this exact structure:

## 🔍 Prompt Analysis

**Overall score: {X}/7 dimensions strong**

| Dimension | Score | Notes |
|------|------|------|
| Clarity | ✅/⚠️/❌ | {brief note} |
| Specificity | ✅/⚠️/❌ | {brief note} |
| Context | ✅/⚠️/❌ | {brief note} |
| Structure | ✅/⚠️/❌ | {brief note} |
| Examples | ✅/⚠️/❌ | {brief note} |
| Constraints | ✅/⚠️/❌ | {brief note} |
| Completeness | ✅/⚠️/❌ | {brief note} |

### Strengths
{strength points}

### What can be improved
{for each ⚠️/❌ dimension: the specific problem + a specific fix}

## ✨ Optimized Prompt

{the full rewritten version with all improvements}

---
**What changed:**
{a numbered list of what was changed and why}

After showing the analysis, ask the user: use original / use optimized / edit manually?

English-Learning Integration

After optimizing, automatically trigger the english-learner skill's English study:

1. Check the user's original Chinese phrasing for grammar mistakes, typos, or awkward usage — if any, show the corrections first
2. Extract 2-3 high-value English words/expressions from the optimized prompt
3. Compare the user's original phrasing with the optimized phrasing, and point out the word-choice improvements
4. Automatically save all new vocabulary to the word bank via batch_save (no need to ask the user)

Format:

🌐 **English Learning**

**Chinese corrections:** (only shown when issues are found)
| Original | Correction | Note |
|------|------|------|
| {mistake} | {fix} | {reason} |

**Prompt vocabulary highlights:**
| Your phrasing | Optimized phrasing | Why it's better |
|----------|----------|-----------|
| {original word} | {improved} | {brief note} |

💾 Saved to the word bank automatically

Rules:

• Only trigger when the prompt contains Chinese or the user is a native Chinese speaker
• Correct the Chinese issues first, then show the English improvements
• Focus on the learning value of "how to describe intent more precisely in English"
• Save directly, no user confirmation needed

Prompt-Improvement Patterns

Patterns to draw on when improving weak dimensions:

1. Vague → Specific

• Before: "Write good code"
• After: "Write TypeScript code with explicit return types, no any, and no function longer than 20 lines"

2. Missing context → With context

• Before: "Translate this"
• After: "Translate this marketing copy from Chinese to English, keeping a casual tone, aimed at a Gen-Z audience"

3. No examples → With examples

• Before: "Format the data"
• After: "Format the data as: input: {raw} → output: | col1 | col2 |"

4. No structure → With structure

• Before: A wall of text mixing multiple concerns
• After: Ordered steps with clear numbering and grouping

5. No constraints → With boundaries

• Before: "Write an article"
• After: "Write a 500-word article, no jargon, 8th-grade reading level, with 3 examples"

6. Missing role → With role anchoring

• Before: "Explain X"
• After: "Explain X as a senior backend engineer explaining it to a junior developer"

Error Handling

Execution Checklist

Confirm before responding:

• All 7 dimensions analyzed
• Scores assigned with brief notes
• Strengths acknowledged (not just criticism)
• Each weak dimension has a specific, actionable fix
• The full optimized prompt is provided
• The changes are explicitly listed
• The user is given a choice: accept / reject / edit

Hooks

This skill registers an IDE hook so that the pre-flight analysis triggers deterministically — the AI does not need to remember to invoke it.

Scope

Global — installed to ~/.claude/settings.json and ~/.trae/hooks.json (no project-level state).

Events

Trigger Strategy (three-tier mode)

The hook now triggers on all prose input (skipping only code, file paths, shell commands, and very short input ≤7 characters):

1. Explicit request (explicit) — the message contains optimize / improve / review / rewrite / check / refine my prompt, make this prompt more X, or the Chinese equivalents 优化提示词 / 改进提示词 / 重写提示词. Output: a full 7-dimension analysis + an Optimized Prompt block.
2. Structured prompt shape (structured) — the message is ≥400 characters and ≥4 lines, and contains ≥2 markers such as you are, your task is, act as, instructions:, constraints:, output format:. Output: a full 7-dimension analysis + an Optimized Prompt block.
3. Light mode (light) — all other prose input (short exchanges, chit-chat questions). Output: a single line — pick the weakest of the 7 dimensions and give one concrete rewrite suggestion; if all 7 pass, render the fixed string "✨ Prompt-opt: already clear, no rewrite needed." and continue the task.

The event log (~/.learnwy/prompt-optimizer/events.jsonl) records the trigger type (explicit / structured / light); use cli.cjs trends to view the distribution.

Install

node scripts/cli.cjs install --scope global --target both

Uninstall

node scripts/cli.cjs uninstall --scope global --target both