Turn Claude Into Your Favorite New Teammate

A Practical Prompt-Engineering Playbook for Junior-College Graduates and Beyond

If you have just opened Claude for the first time, you may feel as if you are greeting a brand-new colleague who is brilliant yet knows nothing about your world. The nine short guides bundled with this article—straight from Anthropic’s own documentation—show how to turn that stranger into the most helpful teammate you have ever had. Below, every original idea, technical detail, and code snippet comes only from those nine files. Nothing has been added from the outside world.

Table of Contents

1. Why Prompt Engineering Beats “Wing-It” Questions
2. Give Claude a Name-Tag: Role Prompting
3. Write the Job Description: Clear, Contextual, Step-by-Step Prompts
4. Let Claude Think Out Loud: Chain-of-Thought Prompting
5. Feeding Long Documents: Long-Context Tips
6. Start Claude’s Sentence for Him: Prefill for Precision
7. Show, Don’t (Only) Tell: Few-Shot Prompting
8. Use XML Tags Like Folders: Structured Prompts
9. Case Study: Re-Write a SaaS Multitenancy Review in One Page
10. Your Next 30 Minutes with Claude

1. Why Prompt Engineering Beats “Wing-It” Questions

Anthropic’s own prompt-generator notebook is described as “a cure for blank-page paralysis.” Instead of staring at an empty chat box, you let Claude draft a first-version prompt that already follows best practices. You then refine that draft until it fits your exact task.
In short, prompt engineering is simply writing the kind of clear instructions you would give to an extremely capable intern who has no prior memory of your team culture.

2. Give Claude a Name-Tag: Role Prompting

Imagine you walk into a meeting and someone says, “You, give us a legal opinion.”
If that person is a lawyer, you will get jargon, citations, and risk warnings.
If that person is a marketer, you will get storytelling and bullet points.
Claude behaves the same way.
Set the role in the system parameter:

system: You are a senior SaaS architect with ten years of experience…

Before vs. After

3. Write the Job Description: Clear, Contextual, Step-by-Step Prompts

Treat Claude like an intern who just arrived from another planet.
Every prompt should read like a mini job description with four parts:

1. Context – Where will the answer be used? A pitch deck? An internal wiki?
2. Goal – What does success look like? One page? A JSON file?
3. Format – Plain text, Markdown table, or only code snippets?
4. Steps – Numbered list so Claude knows the order.

Bad vs. Good Prompt

4. Let Claude Think Out Loud: Chain-of-Thought Prompting

When the task is complex—research, math, or risk analysis—ask Claude to reason step-by-step. Anthropic lists three levels:

1. Basic: Append “Think step-by-step.”
2. Guided: Provide the exact steps to follow.
3. Structured: Wrap the thinking in <thinking> tags and the answer in <answer> tags so downstream code can split them.

5. Feeding Long Documents: Long-Context Tips

Claude 3 models accept up to 200,000 tokens, roughly 150,000 English words. To use that window well:

• Put the document first: Place your 20,000-word file at the top, then ask the question.
• Wrap each document: Use <document> with <source> and <document_content> sub-tags.
• Ask Claude to quote first: “Before you answer, quote the exact lines that support your point.” This reduces hallucination.

6. Start Claude’s Sentence for Him: Prefill for Precision

You may prefill Claude’s response by writing the first few words in the Assistant field. This is useful when you need strict formatting or want to skip small talk.

Example:

Human: …task description…

Assistant: ```json
{
  "risk_level": "

Claude will continue from the colon and keep the JSON valid.

7. Show, Don’t (Only) Tell: Few-Shot Prompting

Examples are faster than rules. Provide two or three input-output pairs inside <example> tags. Make sure they:

• Mirror your real task
• Cover edge cases
• Use consistent tags for easy parsing

<examples>
  <example>
    <input>Clause: Either party may terminate at any time.</input>
    <output>Risk: Unilateral termination leaves vendor exposed. Suggest 30-day written notice.</output>
  </example>
</examples>

8. Use XML Tags Like Folders: Structured Prompts

When a prompt mixes context, instructions, examples, and format requirements, XML tags act like folders:

<context>
  We are a B2B SaaS startup selling to manufacturing firms.
</context>

<instructions>
  1. Evaluate scalability…
</instructions>

<examples>
  …
</examples>

<output_format>
  Markdown table with columns: Dimension | Grade | Reason
</output_format>

9. Case Study: Re-Write a SaaS Multitenancy Review in One Page

Let us walk through a real task: condense the 1,200-word multitenancy strategy review (file 3) into a one-page CTO brief. We will apply every tip above.

9.1 System Prompt (Role)

system: You are a cloud-native architect with 12 years of experience who writes for CTOs.

9.2 Human Prompt (Clear Job Description)

Read the architecture review below. Deliver a one-page CTO brief that includes:
1. A score table (Scalability / Security / Cost)  
2. Top 3 risks  
3. A 4-week PoC plan  
Use Chinese (since the source is Chinese), wrap your thinking in <thinking>, and the final brief in <answer>.

9.3 Document Placement

Paste the entire file 3 content inside <document> tags at the very top.

9.4 Prefill

Assistant: <thinking>

Claude will then output:

<answer>
## CTO Brief: Multitenancy Strategy Review (2024-07-25)

| Dimension | Grade | One-Sentence Summary |
|-----------|-------|----------------------|
| Scalability | B+ | Database sharding is strong, but app tier lacks auto-scaling. |
| Security | B- | Data isolation is solid; encryption and IAM depth are missing. |
| Cost | B | Shared nodes save baseline costs, yet may over-provision at peak. |

### Top 3 Risks
1. **Peak Avalanche**: Shared app servers with no auto-scaling can crash during promotions.  
2. **Compliance Gap**: No encryption at rest blocks ISO 27001 approval.  
3. **Runaway Cost**: Per-tenant databases could triple monthly spend beyond 1,000 tenants.

### 4-Week PoC Plan
- Week 1: K6 load test 5k concurrent users to measure auto-scaling impact.  
- Week 2: Pilot AES-256 encryption + RBAC with three manufacturing clients.  
- Week 3: Cost benchmark Snowflake vs. sharding at 1 TB scale.  
- Week 4: Third-party pen-test focusing on tenant isolation.
</answer>

10. Your Next 30 Minutes with Claude

Enjoy working with your new teammate.