Prompt Engineer

Overview

This skill covers systematic design and iteration of prompts for large language models (LLMs). It applies proven techniques — zero-shot, few-shot learning, chain-of-thought reasoning, role prompting, structured output constraints, system prompt design, and multi-step prompt chaining — to improve accuracy, consistency, and reliability of LLM outputs. The skill is applicable to any LLM API (OpenAI, Anthropic, Gemini, Mistral, open-source models) and covers both single-turn and multi-turn conversation design, as well as production-grade prompt templates with variable injection.

When to Use

• A prompt produces inconsistent, vague, or off-format outputs and needs iteration
• Designing prompts that must return structured JSON, XML, Markdown, or CSV output
• Building few-shot examples to guide classification, extraction, or transformation tasks
• Creating system prompts that establish persona, tone, constraints, or output rules
• Chaining multiple prompts together for complex multi-step reasoning tasks
• Reducing hallucination by adding grounding instructions, citations requirements, or self-checks
• Optimizing a prompt for a specific model (GPT-4, Claude, Llama, etc.) given its strengths
• Converting a vague user request into a precise, production-ready prompt template

When NOT to Use

• Fine-tuning or training a model on new data (use model training skills)
• Evaluating model quality across a benchmark suite (use eval-designer skill)
• Writing application code that calls the LLM API (use a coding skill)
• Comparing different LLMs for a use case (use model-comparator skill)
• RAG pipeline design (retrieval-augmented generation requires its own architecture skill)

Quick Reference

Instructions

1. Define the task precisely — Write a one-sentence task definition: what the model must do, what the input is, and what the output must look like. Vague tasks produce vague outputs. Example: "Extract all company names and their associated revenue figures from the following earnings call transcript and return them as a JSON array."
2. Choose the prompting technique — Select based on task complexity:
   • Zero-shot: Simple tasks with clear instructions. No examples needed.
   • Few-shot: Classification, formatting, or style tasks. Provide 2–5 labeled examples.
   • Chain-of-thought (CoT): Math, logic, multi-step reasoning. Add "Think step by step."
   • Role prompting: Tasks requiring expertise or persona. "You are a senior tax attorney…"
   • Self-consistency: Run the same CoT prompt N times and majority-vote the answer.
   • Prompt chaining: Decompose a complex task into sequential prompts where each feeds the next.
3. Write the system prompt — For APIs that support system prompts (OpenAI, Anthropic), put persistent instructions here: role, output format, constraints, what to do when uncertain. Keep it under 500 tokens unless the task genuinely requires more context.
4. Structure the user prompt — Use clear delimiters to separate instructions from input data. Use XML tags (<document>, <query>), triple backticks, or --- separators. Place instructions before the data, not after.
5. Specify output format explicitly — Tell the model exactly what format to use. If JSON, provide the schema or a filled example. If Markdown, show the heading structure. If a list, show how items should be formatted. Include a negative example if there is a common wrong format to avoid.
6. Add few-shot examples — For classification or extraction tasks, include 2–5 examples in the prompt. Format them identically to the real input/output pair. Choose examples that cover edge cases and are representative of the real distribution.
7. Iterate and test — Test on at least 10 representative inputs. Track: did the model follow the format? Did it hallucinate? Was the reasoning correct? Identify failure patterns and add instructions or examples to address them.
8. Version and document the prompt — Save prompts in a template file with variable placeholders ({{input}}). Document what model version it was tested on, what temperature, and what the expected pass rate is.
9. Optimize for the target model — Different models respond differently: Claude prefers XML tags and explicit role instructions; GPT-4 responds well to numbered instructions; open-source models often need more explicit format constraints. Test the same prompt on the target model even if it worked on another.
10. Add safety and fallback instructions — Include: what to do if the input is out of scope, how to handle ambiguous inputs, whether to ask for clarification or make a best-effort attempt, and how to indicate low confidence.

Examples

Example 1: Improving a Vague Prompt to Get Structured JSON Output

Input:

Original vague prompt: "Summarize this customer review and tell me if it's positive or negative."

Output:

Improved prompt (system + user):

SYSTEM:
You are a sentiment analysis assistant. You always respond with valid JSON only — no prose, no markdown fences. If you cannot determine sentiment, set "sentiment" to "unknown" and explain in "reason".

USER:
Analyze the following customer review. Return a JSON object with exactly these fields:
{
  "sentiment": "positive" | "negative" | "neutral" | "unknown",
  "confidence": 0.0–1.0,
  "key_themes": ["theme1", "theme2"],
  "summary": "one sentence summary",
  "reason": "brief explanation of sentiment classification"
}

Review:
<review>
{{customer_review_text}}
</review>

Why this works:

• System prompt enforces JSON-only output at the model level
• Explicit schema with types eliminates ambiguity
• XML tag delimiter prevents the model from treating the review text as instructions
• "unknown" fallback prevents hallucinated confident answers
• confidence field enables downstream filtering of low-quality classifications

Example 2: Few-Shot Chain-of-Thought Prompt for Legal Clause Classification

Input:

Task: Classify contract clauses as "liability-limiting", "indemnification", "IP-assignment", or "other".
The model keeps misclassifying indemnification clauses as liability-limiting.

Output:

Few-shot CoT system prompt:

SYSTEM:
You are a contract analysis assistant specializing in clause classification. Think through each clause step by step before classifying. Respond only in the format shown in the examples.

USER:
Classify each contract clause. Think step by step, then provide your classification.

---
EXAMPLE 1:
Clause: "Neither party shall be liable for indirect, incidental, or consequential damages arising out of this agreement."
Reasoning: This clause restricts what types of damages can be claimed. It limits liability exposure — it does not require one party to protect another. This is a liability cap, not an indemnification.
Classification: liability-limiting

---
EXAMPLE 2:
Clause: "Vendor shall indemnify, defend, and hold harmless Client from any third-party claims arising from Vendor's breach of this agreement."
Reasoning: This clause requires the Vendor to actively protect the Client from external claims. It creates an obligation to defend and compensate — this is indemnification, not merely limiting damages.
Classification: indemnification

---
EXAMPLE 3:
Clause: "All work product created by Contractor under this agreement is assigned to Client upon creation."
Reasoning: This clause transfers ownership of created works. It concerns who owns intellectual property, not liability or indemnification.
Classification: IP-assignment

---
Now classify this clause:
Clause: "{{clause_text}}"
Reasoning:

Why this works:

• Chain-of-thought examples show the model the reasoning pattern to distinguish similar categories
• The two easily-confused categories (liability vs indemnification) each get an explicit contrasting example
• Format template forces reasoning before classification, reducing snap-judgment errors
• Reasoning: at the end primes the model to complete the reasoning before the answer

Best Practices

• Always test prompts on at least 10 real inputs before declaring them production-ready
• Version prompts in code just like application code — breaking changes in prompts are real bugs
• Use the lowest effective temperature: 0.0 for deterministic extraction, 0.7–1.0 for creative tasks
• Prefer XML tags over triple backticks as delimiters — they're less likely to appear in real input
• Put the most important instructions at the beginning AND end of the prompt (primacy + recency effect)
• When using few-shot examples, ensure they cover the edge cases you care about most
• Keep system prompts focused — every sentence should earn its token budget
• Use "You must" and "Always" for hard constraints; use "prefer" or "try to" for soft preferences

Common Mistakes

• Giving vague instructions like "be helpful and accurate" without specifying what accuracy means
• Not specifying what to do when the model is uncertain — it will hallucinate a confident answer
• Using too many few-shot examples (>10) which can cause the model to pattern-match instead of reason
• Forgetting to test the prompt on the actual target model — prompts are not portable across models
• Putting instructions after the input data — models weight early context more heavily
• Asking multiple distinct tasks in one prompt — split into separate prompts or clearly numbered steps
• Assuming the same prompt works at different temperatures — always co-tune prompt and temperature

Tips & Tricks

• Add "Do not add any text before or after the JSON" to enforce clean parseable JSON output
• Use "If you are unsure, say 'I don't know' rather than guessing" to reduce hallucination
• For long documents, use "Here is the most relevant section:" before the content to focus attention
• Chain-of-thought is most effective in the middle of a response — put format instructions last
• "Let's think step by step" reliably improves math and logic; for simpler tasks it wastes tokens
• Test prompt robustness by deliberately injecting adversarial inputs (e.g., "ignore previous instructions")
• Use Anthropic's Constitutional AI principles or OpenAI's system prompt best practices as references

Related Skills

• dataset-curator
• eval-designer
• model-comparator