LLenny's Podcast
← All frameworks
ProductivityKevin Weil (CPO at OpenAI, ex-Instagram, Twitter)

Poor Man's Fine-Tuning (Few-Shot + Role Priming)

Steer a model with in-prompt examples and a role identity instead of a full fine-tune.

Difficulty
Starter
Time to result
~days to results
Steps
2
Confidence
90%

You don't need to be a prompt-engineering expert, but two lightweight techniques reliably improve output. First, include several example/good-answer pairs in your prompt — 'poor man's fine-tuning' — so the model learns the shape of what you want. Second, prime the model with a role identity ('you are the world's greatest brand marketer') to shift it into a productive mindset. Both mirror how you'd frame a task for a human.

Origin

Kevin Weil (CPO, OpenAI). The role-priming technique parallels examples popularized in the LLM community; Weil endorses both as practical, human-analogous techniques.

Core principles

  • 01In-prompt examples work much better than no examples, though not as well as a full fine-tune
  • 02Provide 'here's an example, here's a good answer' pairs, then pose the real problem
  • 03A role identity shifts the model into a specific, often higher-quality mindset
  • 04These mirror how framing and priming work on humans

How to run it

  1. 1

    Add example / good-answer pairs to the prompt

    Before your real request, include a few concrete example inputs each paired with a good answer, then ask the model to solve the actual problem.

    Pro tip People don't do this often enough — even 2-3 examples materially improve output.

    Watch out It won't match a true fine-tune, but it far outperforms giving no examples.

  2. 2

    Prime the model with a role identity

    Open the prompt by casting the model into an expert persona relevant to the task ('you are the world's greatest brand marketer; now here's a naming question').

    Pro tip Weil uses 'you're the world's best podcast interviewer' to generate questions he hadn't thought of.

In the wild

Role-primed interview questions

Weil says when preparing interview questions he types 'you're the world's best podcast interviewer' and describes his guest, which surfaces angles he hadn't considered. He confirms role priming ('I want you to be Einstein, now answer this physics problem') genuinely shifts the model into a more useful mindset.

Higher-quality, less generic outputs from the same underlying model.

Common mistakes

Skipping examples entirely

Giving no examples leaves easy quality on the table; in-prompt examples are the cheapest available improvement.

Believing you must be a prompt-engineering expert

Weil wants to kill the idea that deep prompt-engineering is required; a couple of simple, human-analogous techniques carry most of the value.

Is it for you?

Best for

Anyone using LLMs day-to-day who wants better outputs without training a model

Not ideal for

High-volume production use cases where a real fine-tune is justified and in-prompt examples add too much token cost

From the transcript

effectively poor man's fine-tuning by including examples in your prompt

1:27:30

here's an example and here's a good answer. Here's an example, here's a good answer. Now go solve this problem for me and the model…

1:28:00

You can also say things like I want you to be Einstein. Now answer this physics problem for me. Or you are the world's greatest…

1:28:30

From the episode

OpenAI’s CPO on how AI changes must-have skills, moats, coding, startup playbooks, more

Kevin Weil (CPO at OpenAI, ex-Instagram, Twitter)