Write Your First System Prompt — Practical Prompt Engineering, Part 2

Contributor
5 days ago
3 min read

Updated: 2 hours ago

Practical Prompt Engineering · Part 2

A system prompt sets the model's behavior for the conversation. Done well, it dramatically improves output quality. This tutorial walks through writing one.

What You'll Build

A system prompt for a specific task that produces consistent, useful outputs.

Step 1: Pick a Specific Task (5 min)

Don't write a general-purpose prompt. Pick one task:

Summarize customer support tickets
Classify incoming emails
Generate product descriptions
Answer questions about your documentation

Specific tasks produce specific prompts.

Step 2: The Anatomy (5 min)

A working system prompt has:


1. Role: who the model is
2. Task: what it does
3. Inputs: what it'll receive
4. Output format: what it produces
5. Constraints: what it shouldn't do
6. Examples: 1-3 demonstrations (optional but powerful)

Step 3: Draft v1 (10 min)

For summarizing support tickets:


You are a customer support analyst. You summarize support tickets 
into a 2-sentence summary that captures: (1) the customer's issue 
and (2) any action items.

You will receive a ticket transcript. Output only the summary; no 
preamble or formatting.

Do not invent details. If the ticket is unclear, say so in the 
summary.

Simple. Specific. Bounded.

Step 4: Test It (10 min)

Try on 3-5 sample tickets:


ticket = """
Customer: My order hasn't arrived. Order number 12345. It's been 
2 weeks.
Agent: Let me look that up. I see it was shipped via standard 
shipping; ETA was 7-10 business days. I'll check the carrier.
Agent: Looks like the carrier had a delay. New ETA is tomorrow.
Customer: Okay thanks.
"""

response = client.messages.create(
    model="claude-sonnet-4-6",
    system=system_prompt,
    messages=[{"role": "user", "content": ticket}]
)

Read the output. Is it what you wanted?

Step 5: Iterate (15-30 min)

Common issues and fixes:

Output too long? Add to prompt: "Maximum 30 words."

Output too short? Add: "Include the key issue and resolution status."

Inconsistent format? Add explicit format: "Format: 'Issue: ... Action: ...'"

Hallucinations? Add: "Only use information from the ticket. Do not infer beyond what's stated."

Each iteration teaches you what works.

Step 6: Add Examples (15 min)

Few-shot examples dramatically improve consistency:


You are a customer support analyst. You summarize support tickets...

Here are examples:

Input: [example ticket 1]
Output: Issue: Order #X delayed beyond ETA. Action: Carrier issue 
identified; new ETA tomorrow.

Input: [example ticket 2]  
Output: Issue: Customer received wrong item. Action: Replacement 
shipped; return label provided.

Now summarize:
[actual ticket]

The model patterns its output after the examples. Powerful for consistency.

Step 7: Handle Edge Cases (15 min)

What if the ticket is empty? Malformed? Off-topic?

Add to the prompt:


Special cases:
- If the ticket is empty or has no customer issue: respond "No 
  customer issue identified."
- If the ticket contains content unrelated to a support issue: 
  respond "Out of scope."
- If you're uncertain about the action status: say so explicitly.

Edge cases handled in the prompt prevent surprises.

Step 8: Constrain Tone (10 min)

If tone matters:


Tone: professional and neutral. Do not use phrases like "Great 
question!" or "Thanks for reaching out."

Voice: third person; no first-person ("I see that...").

Generic prompts produce generic tones. Specific constraints produce specific tones.

Step 9: Test on Your Eval Set (15 min)

The eval cases from Part 1:


def evaluate_prompt(system_prompt, eval_cases):
    for case in eval_cases:
        output = client.messages.create(
            model="claude-sonnet-4-6",
            system=system_prompt,
            messages=[{"role": "user", "content": case["input"]}]
        ).content[0].text
        
        # Manual or automated check
        pass_check = case["expected_topic"] in output.lower()
        print(f"{'PASS' if pass_check else 'FAIL'}: {output[:100]}")

Without eval, you don't know if changes improve. With it, you do.

Step 10: Version and Store (10 min)

Save the working prompt:


prompts/
  ticket_summary_v1.txt   # First draft
  ticket_summary_v2.txt   # With examples
  ticket_summary_v3.txt   # Current production version

Each new version is a separate file. You can compare across versions, revert, A/B test.

What You Just Did

You wrote a system prompt that:

Has a clear role
Specifies the task
Bounds the output
Provides examples
Handles edge cases
Has a consistent tone

It produces consistent outputs that you can rely on.

Common Failure Modes

Vague role. "You are a helpful assistant." Doesn't steer toward your specific task.

No format specified. Output structure varies; downstream code breaks.

Asking too much. One prompt tries to do 5 things. Break into multiple.

No examples. For complex tasks, examples are the highest-leverage addition.

Untested. Prompt works on one example; fails on others.

Continue the Practical Prompt Engineering path

Previous — Part 1: Set Up Your Prompting Workspace
Next — Part 3: Few-Shot Examples That Work

Part of the Practical Prompt Engineering learning path.

ShiftQuality