gh-jamesrochabrun-skills-all-skills/skills/anthropic-prompt-engineer/references/common_mistakes.md

Common Prompt Engineering Mistakes

Learn from these common pitfalls to create better prompts.

Mistake 1: Being Too Vague

❌ Problem:

Write something about AI

Why it fails:

• No clear scope or direction
• Undefined output format
• Unknown target audience
• No length specification

✅ Solution:

Write a 500-word introduction to large language models for software developers.

Include:
- What LLMs are and how they work (high-level)
- 3 practical use cases in software development
- 1 code example of API usage
- Limitations to be aware of

Tone: Technical but accessible
Format: Markdown with headers

---

Mistake 2: Instruction Confusion

❌ Problem:

Here is some text about Python. It talks about lists and dictionaries.
Can you make it better and also add examples? Use clear language.

Why it fails:

• Instructions mixed with content
• Unclear what "better" means
• No structural separation

✅ Solution:

<instructions>
Improve the following text by:
1. Adding 2 code examples (one for lists, one for dictionaries)
2. Simplifying complex sentences
3. Adding subheadings for each topic
</instructions>

<text_to_improve>
[original text here]
</text_to_improve>

<target_audience>
Beginner Python developers
</target_audience>

---

Mistake 3: No Examples (When Needed)

❌ Problem:

Extract entities from customer support tickets in a structured format.

Why it fails:

• "Structured format" is ambiguous
• No clarity on what entities matter
• Different interpretations possible

✅ Solution:

<instructions>
Extract entities from customer support tickets.
</instructions>

<examples>
<example>
<input>
My order #12345 never arrived. I contacted support@example.com but no response.
</input>
<output>
{
  "order_id": "12345",
  "issue_type": "delivery",
  "contact_method": "email",
  "email": "support@example.com"
}
</output>
</example>
</examples>

<input>
[new ticket]
</input>

---

Mistake 4: Prompt Too Long

❌ Problem:

[5000 words of background information]
[500 words of instructions]
[1000 words of examples]

Now answer this simple question: What is the capital of France?

Why it fails:

• Wastes tokens on irrelevant context
• Dilutes important information
• Slower and more expensive
• May confuse the model

✅ Solution:

What is the capital of France?

Principle: Use minimum necessary context

---

Mistake 5: Not Controlling Output Format

❌ Problem:

Give me the user's information from this text.

Result: Unpredictable format - might be prose, bullet points, or inconsistent structure

✅ Solution:

<instructions>
Extract user information and return as JSON.
</instructions>

<input>
[text here]
</input>

<output_format>
{
  "name": string,
  "email": string | null,
  "phone": string | null
}
</output_format>