Tutorial 2: Prompt for Code

Contributor
5 days ago
2 min read

Vague prompt → vague code. Specific input gets specific output.

Step 1: State the Goal (10 min)

Bad: "Write a sorting function."
Good: "Write a TypeScript function that sorts users by lastName,
       then firstName, both case-insensitive."

The model can't read your mind. Specify:

Language
Inputs
Outputs
Constraints

Step 2: Include Context (10 min)

class User(BaseModel):
    id: int
    email: EmailStr

Tell the model: "We use Pydantic models for validation. Write a similar model for Order with fields: id (int), user_id (int), amount (Decimal, > 0), status (enum: pending/paid/cancelled)."

Models match local conventions when shown them. Less rework.

Step 3: Be Explicit About Constraints (10 min)

Function must:
- Run in < 100ms for 10k items
- Handle empty list gracefully
- Not mutate the input
- Use only stdlib

Constraints up front. Model designs to them.

Without: model picks freely; you fix later.

Step 4: Show, Don't Just Tell (10 min)

Input:
  [{name: 'Alice', age: 30}, {name: 'Bob', age: 25}]

Expected output:
  [{name: 'Bob', age: 25}, {name: 'Alice', age: 30}]

Function: sort by age ascending.

Example output anchors the implementation. Fewer "almost right" cases.

Step 5: Iterate, Don't Restart (10 min)

After first response:

That's close. Change:
1. Use type hints (TypeScript)
2. Handle null inputs
3. Make case-insensitive

Don't rewrite the prompt. Iterate.

The model carries context. Use it.

Step 6: Ask for Tests (10 min)

Write the function and 5 test cases covering:
- Empty input
- Single item
- Duplicates
- Pre-sorted
- Reverse-sorted

Tests describe behavior. Often: writing tests via AI reveals what the implementation should do.

Then you verify the tests yourself.

Step 7: Specify the Output Format (5 min)

Return:
1. The function code
2. Tests as a separate code block
3. Brief explanation

Without: messy mixed output.

For structured tasks, even more specific: "Output as YAML matching this schema: ..."

Step 8: Use the Right Persona (5 min)

Act as a senior Python engineer who prioritizes simplicity over
cleverness.

Personas nudge style. Mileage varies.

Often: "expert in X" with adjectives matters more than the persona itself.

Step 9: Don't Skip the Why (10 min)

I'm building a high-traffic API; this function runs on every
request. Need it as fast as possible. Memory matters less.

Context → better implementation choices.

Without: model defaults to readability over performance.

Step 10: Verify (10 min)

The output looks right. Test it:

Run the code
Check edge cases
Read it carefully

LLMs make confident wrong code. Always verify.

Especially security-sensitive code: read every line.

What You Just Did

Prompting for code: goal, context, constraints, examples, iterate, tests, format, persona, why, verify. Quality prompts.

Common Failure Modes

Vague prompt. Vague code.

No context. Code that ignores conventions.

Don't iterate; rewrite. Lose carried context.

Skip verification. Confident-wrong code ships.

Over-specified prompt. Brittle to small changes.

ShiftQuality