Role — who the model is right now

1. What role is for

Role isn't decoration. It's the line that tells the model what behavior to bias toward.

A good role:

• Names a perspective the model wouldn't have by default.
• Constrains the answer space.
• Pairs with a verb the role would actually do.

Examples that work:

• You're a security reviewer. List only issues a security audit would flag.
• You're a junior teacher explaining to someone who's never seen this before. Use plain words and short sentences.

Examples that don't:

• You're a 10x senior engineer with 30 years of experience. (Adds nothing — the model still tries its best either way.)
• You're a helpful AI assistant. (Default. Decoration.)

2. When role helps, when it doesn't

Role helps when there's a clear biasing direction you want to push the answer toward. Strict reviewer. Skeptical critic. Plain-English teacher.

Role doesn't help when:

• You're asking for a factual or technical answer with one right answer.
• The role you picked doesn't match the verb. ("You're a poet. Write me a SQL query.")
• The rest of the prompt already constrains the answer. (If you specified format, criteria, and examples, the role isn't doing extra work.)

The honest test: delete the role line. If the output gets noticeably worse, the role was earning its place. If the output is the same, the role was decoration.

3. Decoration vs. work

Same task, two roles.

You're a 10x senior engineer with 20 years of experience.

Review this code.

You're a senior reviewer doing a pre-merge review for a high-traffic production service.

You ignore style nits. You flag only:
- Correctness bugs
- Race conditions
- Memory leaks
- Security issues

Review this code.

Both have a role. The first is window-dressing. The second is constraining the answer space — telling the model which kinds of comments to ignore and which to surface. That's role doing work.

4. Try it

Open your prompt template — the snippet you start most prompts with.