Prompt Optimizer

Transform rough, ambiguous natural language into production-grade prompts through diagnostic analysis and structural reconstruction.

Workflow

Phase 1: Diagnose

Analyze the user's input against the 7 diagnostic dimensions in techniques.md. Identify which dimensions are weak or missing. Produce a concise diagnosis:

**缺失诊断**：
- 角色：未指定 / 模糊
- 上下文：缺少背景信息
- 约束：未设定边界
- 输出格式：未指定
- 示例：无
- 推理引导：无

Keep this diagnosis brief — no more than one line per dimension. Only list dimensions that are genuinely deficient.

Phase 2: Reconstruct

Based on the diagnosis, rebuild the prompt. Follow these rules:

Fill every gap: For each missing dimension, infer reasonable defaults from the user's intent. Never leave a dimension blank unless truly inapplicable.
Match the tier: Choose the structural tier based on task complexity:
- Simple Q&A / one-shot tasks → Tier 1 (Role + Task + Format)
- Analysis / creation / multi-step tasks → Tier 2 (add Context + Constraints)
- Complex workflows / high-stakes output → Tier 3 (add Examples + Reasoning Guide)
Infer, don't ask: Make reasonable assumptions for missing details rather than leaving placeholders. Infer from the domain, task type, and common practice. Only flag truly ambiguous intent for user clarification.
Be specific, not generic: Replace "You are an expert" with concrete domain-seniority descriptors. Replace "write well" with measurable quality criteria.
Preserve intent: Never alter the user's core goal. Only add structure, clarity, and guardrails.
Restraint over over-engineering (克制优先) — This is the most important rule:
- Give the model principles to follow, not templates to fill. Trust its ability to organize output based on actual content.
- Constraints should be minimal but sharp: 3-5 core principles hit harder than 15 checklist items.
- When in doubt, err on the side of fewer words. A 10-line prompt that nails the essence beats an 80-line prompt that micromanages.
- Never write pre-fabricated table schemas, fixed section orders, or rigid word counts unless the user explicitly demands them.
- Use phrases like "自行组织"、"怎么讲清楚就怎么讲"、"根据内容决定" instead of "按以下格式输出".
- The optimized prompt should feel like talking to a smart colleague, not filling out a bureaucratic form.
- Never prescribe the model's thinking process (e.g., "先通读全文，再逐项核实，最后审视"). The model knows how to reason — your job is to define WHAT to evaluate, not HOW to think.

Phase 3: Present — Three-Tier Output

Always output 3 versions of the optimized prompt, from minimal to comprehensive. Let the user choose the level that fits their needs.

## 诊断

[Concise diagnosis — one line per dimension, only deficient ones]

---

### 极简版

[1-3 sentences: role + core task. Ready to copy-paste for quick use.]

### 规范版

[Add context + key constraints on top of the minimal version. For formal work scenarios.]

### 发布级版

[Full dimensional coverage: add reasoning guide, examples (if helpful), adaptive structure hints. For high-stakes tasks or prompts shared with others.]

---

## 改进说明

[3-5 bullet points: what was changed and why. Be specific.]

Version guidelines:

极简版: ≤ 3 sentences. Role + task + one essential constraint. If the original prompt is already this simple, acknowledge it and skip to 规范版.
规范版: ≤ 10 sentences. The recommended default for most use cases. Covers role, context, task, and 3-5 sharp constraints.
发布级版: ≤ 20 sentences. Add reasoning hints and examples only if they genuinely help — never pad for length. Do NOT prescribe thinking flow or analysis steps; trust the model to reason based on content. Focus on: sharper role definition, broader context, and stronger quality criteria.

Each version should be self-contained and independently usable. Never write "see above" or "add the following to the previous version."

Quality Checklist

Before outputting, verify all three versions:

Each version can be executed independently without additional clarification
Every sentence serves a purpose (no filler)
The role is specific (domain + seniority + perspective)
Constraints include both format AND quality criteria
Restraint is applied: each version is as short as it can be while remaining effective
The user's original intent is fully preserved in all versions
No placeholder text like "[insert here]" or "TBD"

Handling Edge Cases

Already good prompt: If the input is already well-structured, acknowledge it and offer targeted enhancements (e.g., "Your prompt is already strong. Suggested micro-improvements: ...")
Too vague to infer: If the task itself is ambiguous (not just poorly structured), ask ONE clarifying question before optimizing. Do not optimize a prompt whose purpose is unclear.
Multiple prompts: If the user provides several prompts, optimize each separately with its own diagnosis.
Non-prompt input: If the user provides content that is not a prompt (e.g., a story, code, data), clarify: "This appears to be content rather than a prompt. Are you asking me to create a prompt that would generate content like this, or did you mean something else?"

Reference

For detailed prompt engineering techniques, patterns, and anti-patterns, read techniques.md when:

The task requires advanced techniques (Tier 3)
The user's domain is specialized and needs domain-specific patterns
You need to refresh on specific technique implementation details

prompt-optimizer