Claude Prompt Optimizer is a meta-prompting system that transforms vague, unstructured prompts into professionally optimized prompts following Anthropic's official best practices. Paste any raw prompt — from a one-liner to a detailed brief — and get back a structured, XML-tagged, model-specific prompt engineered for Claude Opus 4.6's full capabilities.
This is not a prompt collection. It's a systematic optimization engine with 10 codified rules, a 10-component prompt architecture, and complexity-based routing — all derived from Anthropic's documentation and real-world testing.
- Before & After
- Features
- How It Works
- Quick Start
- The 10 Optimization Rules
- Prompt Framework (10 Components)
- Examples
- Why Claude Opus 4.6 Specific?
- Repository Structure
- Contributing
- License
- Star This Repo
| ❌ Raw Prompt | ✅ Optimized Prompt |
|---|---|
|
<role>
You are a senior data engineer specializing
in anomaly detection and statistical analysis
with Python.
</role>
<context>
I have CSV datasets from production IoT
sensors. Files range from 10K–5M rows with
columns: timestamp, sensor_id, value, status.
Some contain missing values and irregular
time intervals.
</context>
<task>
Write a Python script that:
1. Loads one or more CSV files
2. Detects statistical anomalies using
z-score and IQR methods
3. Handles missing data gracefully
4. Outputs a summary report with flagged rows
</task>
<constraints>
- Use only pandas, numpy, scipy (no ML libs)
- Handle files up to 5M rows efficiently
- Include type hints and docstrings
- Add CLI argument parsing for file paths
</constraints>
<output_format>
Provide the complete script in a single Python
code block, followed by a usage example and
sample output showing detected anomalies.
</output_format> |
The optimizer analyzed the vague input, inferred the likely use case, added a domain-expert role, defined constraints for production quality, and structured everything with XML tags that Claude Opus 4.6 parses with precision.
- 10 Codified Optimization Rules — systematic, reproducible prompt improvements based on Anthropic's documentation
- 10-Component Prompt Framework — modular architecture covering role, context, task, constraints, examples, and more
- Complexity-Based Routing — automatically scales optimization depth (minimal → moderate → full) based on prompt complexity
- Claude Opus 4.6 Specific — tuned for 200K context window, adaptive thinking, literal instruction following, and no-prefill architecture
- XML Tag Structuring — transforms flat text into semantically tagged sections that Claude parses reliably
- Ambiguity Resolution — detects missing information and either infers it or prompts you for clarification
- Domain Detection — identifies the prompt's domain (coding, creative writing, analysis, etc.) and applies domain-specific patterns
- Quality Guardrails — built-in checks prevent common failure modes like over-optimization, hallucination triggers, and context overflow
- Language Preservation — responds in the user's language; optimized prompts stay in the original language unless you request otherwise
- Iterative Refinement — give feedback and the optimizer revises targeted sections without re-doing the full analysis
┌─────────────────┐ ┌──────────────────────────┐ ┌─────────────────────┐
│ Your Raw │ │ Prompt Optimizer │ │ Optimized Prompt │
│ Prompt │────▶│ │────▶│ Ready for Claude │
│ (any format) │ │ Analyze → Route → │ │ Opus 4.6 │
│ │ │ Structure → Optimize → │ │ │
│ │ │ Quality Check │ │ │
└─────────────────┘ └──────────────────────────┘ └─────────────────────┘
The 5-step workflow under the hood:
- Prompt Analysis — the optimizer detects intent, complexity level, domain, expected output type, and missing elements
- Complexity Routing — simple prompts get 3–4 components; moderate ones get 5–7; complex prompts get the full 10-component framework
- Rule Application — the relevant subset of 10 optimization rules fires (not all rules apply to every prompt)
- Quality Check — a built-in checklist ensures the task is unambiguous, XML tags are valid, examples match the desired behavior, and no contradictory instructions exist
- Structured Output — you receive the analysis, the optimized prompt in a copy-ready code block, and concise notes explaining what changed and why
Choose the method that fits your workflow:
- Open claude.ai and create a new Project
- Paste the contents of
CLAUDE.mdinto the Project Instructions field - Upload
GUIDE.mdas a knowledge file in the project - Start a conversation — paste any raw prompt and the optimizer handles the rest
- Copy the full contents of
CLAUDE.md - Paste it as the system prompt in any Claude interface (API Console, Workbench, etc.)
- Send your raw prompt as the user message
import anthropic
client = anthropic.Anthropic()
# Load the optimizer system prompt
with open("CLAUDE.md", "r") as f:
system_prompt = f.read()
message = client.messages.create(
model="claude-opus-4-6-20250929",
max_tokens=8096,
system=system_prompt,
messages=[
{"role": "user", "content": "Your raw prompt here"}
]
)
print(message.content[0].text)Once running, you can type these commands in the conversation:
| Command | What It Does |
|---|---|
help |
Explains how the optimizer works |
rules |
Shows a compact summary of all 10 optimization rules |
example |
Displays a full before/after optimization demo |
framework |
Shows the 10-component framework overview |
tips |
Gives 5 practical tips for writing better input prompts |
| # | Rule | What It Does |
|---|---|---|
| 1 | Be Explicit & Detailed | Replaces vague instructions with specific, detailed ones. Claude 4.x follows instructions literally — precision pays off. |
| 2 | Provide Context & Motivation | Explains why an instruction matters, not just what to do. Claude performs better when it understands the purpose. |
| 3 | Use XML Tags for Structure | Wraps prompt sections in semantic XML tags (<role>, <task>, <constraints>, etc.) that Claude is trained to parse reliably. |
| 4 | Inject Few-Shot Examples | Adds 3–5 diverse input/output examples when the task involves classification, formatting, or ambiguous patterns. |
| 5 | Activate Chain-of-Thought | Adds explicit step-by-step reasoning triggers for complex tasks. Works with Opus 4.6's adaptive thinking to boost quality. |
| 6 | Assign an Expert Role | Gives Claude a specific persona with domain expertise, experience level, and communication style. |
| 7 | Define Output Format Explicitly | Prescribes the exact structure, length, and format of the response. Uses positive phrasing ("write prose paragraphs" not "don't use lists"). |
| 8 | Optimize for Long Context | Places long documents at the start of the prompt, tags them with indexed XML, and instructs Claude to extract before answering. |
| 9 | Steer Tool Use & Agentic Behavior | Clearly specifies whether Claude should act or advise, and encourages parallel tool calls for independent actions. |
| 10 | Prevent Over-Engineering | Adds explicit anti-overengineering clauses to keep Claude focused on what was actually requested. |
Not every rule fires for every prompt. A simple factual question won't get 10 XML tags — it'll get a clean, focused refinement. The optimizer scales proportionally.
The optimizer draws from 10 structural components when building the optimized prompt. Think of these as building blocks — the optimizer selects the right subset based on complexity routing.
┌─────────────────────────────────────────────────────┐
│ OPTIMIZED PROMPT │
├─────────────────────────────────────────────────────┤
│ 1. ROLE / PERSONA │
│ Who should Claude be? │
├─────────────────────────────────────────────────────┤
│ 2. TASK CONTEXT │
│ Why is this task being performed? │
├─────────────────────────────────────────────────────┤
│ 3. TONE CONTEXT │
│ What communication style? │
├─────────────────────────────────────────────────────┤
│ 4. BACKGROUND DATA / DOCUMENTS │
│ Relevant information (XML-tagged) │
├─────────────────────────────────────────────────────┤
│ 5. DETAILED TASK DESCRIPTION │
│ What exactly needs to be done? │
├─────────────────────────────────────────────────────┤
│ 6. RULES & CONSTRAINTS │
│ What is / isn't allowed? │
├─────────────────────────────────────────────────────┤
│ 7. EXAMPLES (Few-Shot) │
│ Input/output pairs in <examples> tags │
├─────────────────────────────────────────────────────┤
│ 8. OUTPUT FORMAT │
│ Structure, length, format of the response │
├─────────────────────────────────────────────────────┤
│ 9. THINKING INSTRUCTIONS │
│ Chain-of-Thought / Thinking Guidance │
├─────────────────────────────────────────────────────┤
│ 10. INPUT / VARIABLE │
│ {{USER_INPUT}} placeholder │
└─────────────────────────────────────────────────────┘
Component Details (click to expand)
| # | Component | XML Tag | When Used |
|---|---|---|---|
| 1 | Role / Persona | <role> |
Almost always — anchors expertise and tone |
| 2 | Task Context | <context> |
When the task needs background or motivation |
| 3 | Tone Context | (within <role> or <instructions>) |
When communication style matters |
| 4 | Background Data | <documents>, <input> |
When user provides data, code, or documents |
| 5 | Task Description | <task> |
Always — the core instruction |
| 6 | Rules & Constraints | <constraints>, <instructions> |
When boundaries, limits, or prohibitions apply |
| 7 | Examples | <examples><example> |
For classification, formatting, or ambiguous patterns |
| 8 | Output Format | <output_format> |
When format isn't self-evident |
| 9 | Thinking Instructions | <thinking> guidance |
For complex reasoning, math, multi-step logic |
| 10 | Input / Variable | {{VARIABLE_NAME}} |
When the prompt is a reusable template |
Full before/after examples for different domains are in the examples/ directory:
- Coding Prompts — script generation, API design, code review
- Creative Writing — fiction, blog posts, copywriting
- Business & Analysis — strategy, executive summaries, data analysis
| ❌ Raw | ✅ Optimized |
|---|---|
|
<role>
You are an experienced technology journalist
specializing in digital health and medical AI
applications. You write for an audience of
healthcare managers, physicians, and health-tech
investors.
</role>
<task>
Write a well-researched blog article about the
current state and future of AI in healthcare.
</task>
<instructions>
- The article should be 1,500–2,000 words
- Structure it with a compelling intro, 4–5
content sections, and a forward-looking
conclusion
- Include concrete real-world examples from
diagnostics, drug development, patient care,
and administration
- Address challenges honestly: data privacy
(especially GDPR), liability, bias in
training data
- Back claims with specific, verifiable facts
and studies where possible
- Write in a professional but accessible tone
— no marketing speak, no exaggerated promises
</instructions>
<output_format>
Continuous prose with subheadings (##). No
bullet point lists in the body text. Format
for direct CMS insertion.
Structure:
1. Title (attention-grabbing, not clickbait)
2. Introduction (hook + relevance)
3. Body (4–5 sections with subheadings)
4. Conclusion with outlook
</output_format>
<constraints>
- Do not present unsubstantiated claims or
speculation as facts
- No marketing language or buzzword overload
- If uncertain about a fact, flag it
transparently
</constraints> |
This is the actual example from the optimizer's meta-prompt — a real transformation it produces.
This optimizer isn't model-agnostic. It's specifically tuned for Claude Opus 4.6's architecture and behaviors:
| Feature | How the Optimizer Uses It |
|---|---|
| 200K Context Window | Generates comprehensive prompts with full examples and context without worrying about token limits. Uses long-context optimization (Rule 8) to place documents correctly. |
| Adaptive Thinking | Strategically triggers extended thinking for complex reasoning tasks. Avoids the word "think" when extended thinking may be disabled (uses "analyze", "evaluate" instead). |
| XML Tag Parsing | Claude Opus 4.6 parses XML tags as semantic structure — the optimizer exploits this with consistent, well-nested tags. |
| Literal Instruction Following | Claude 4.x follows instructions more literally than predecessors. Examples are adopted exactly. The optimizer writes precise, unambiguous instructions. |
| No Prefill Support | Opus 4.6 returns a 400 error on assistant prefills. The optimizer never generates prefill patterns — it uses explicit format instructions instead. |
| Over-Engineering Tendency | Opus models tend to add unrequested features. Rule 10 adds explicit anti-overengineering clauses when appropriate. |
| Positive Phrasing | Claude 4.x responds better to "do X" than "don't do Y". The optimizer formulates all instructions positively. |
| Action vs. Advice Distinction | Claude 4.x distinguishes sharply between "suggest changes" and "implement changes". The optimizer makes this explicit. |
claude-prompt-optimizer/
├── README.md # You are here
├── CLAUDE.md # Core meta-prompt (the optimizer engine)
├── GUIDE.md # User guide: how to write input prompts
├── PROJECT_INSTRUCTIONS.md # System behavior & configuration docs
├── CONTRIBUTING.md # How to contribute
├── LICENSE # MIT License
├── .github/
│ └── ISSUE_TEMPLATE/
│ ├── bug_report.md # Report optimization problems
│ └── feature_request.md # Suggest new rules or features
└── examples/
├── README.md # Example category overview
├── coding-prompts.md # Before/after: coding tasks
├── creative-writing.md # Before/after: creative tasks
└── business-analysis.md # Before/after: business tasks
We welcome contributions! Whether it's a new optimization example, a rule refinement, or a bug report — see CONTRIBUTING.md for guidelines.
Key ways to contribute:
- Submit before/after optimization examples from your real-world usage
- Propose new optimization rules (with evidence from Anthropic's documentation)
- Report cases where the optimizer produces suboptimal results
- Improve or translate documentation
This project is licensed under the MIT License — see LICENSE for details.
You're free to use this optimizer personally, commercially, in your own projects, or as part of a larger system.
If this optimizer saves you time or improves your Claude results, please consider starring the repo — it helps others discover it and motivates continued development.
⭐ Star this repo | 🐛 Report an issue | 💡 Request a feature
Made with 🖤 by cheswick.dev · Not affiliated with Anthropic
