Why Finetune When a Prompt Can 2x LLM Performance?

The Prompting Paradigm (Software 3.0)

In the age of LLMs, it turns out we’re basically writing English programs. Andrej Karpathy famously dubbed this Software 3.0 – a world where effective prompts are the new magic code. oai_citation:0‡ai.gopubby.com. In other words, telling an AI exactly what you want is rapidly becoming the most powerful skill in AI development. As one practitioner wryly notes, he “tried every fancy technique in the book” to improve his model… only to discover “the solution was hiding in plain sight: a better prompt.” oai_citation:1‡ai.gopubby.com. Cue the facepalm (and double the accuracy).

The Problem: When LLMs Ignore Instructions

A common headache is that LLMs often hallucinate or ignore context when answering. Without clear guidance, they’ll blurt out outdated or made-up facts. For example, ask a hallucinating model “Who is the CEO of Twitter?” and you might get “Jack Dorsey!” from 2020, even if events changed in 2025. This happens because LLMs only know their training cutoff and lack the specific context you want them to use oai_citation:2‡ashikka.medium.com. In short, they don’t always “listen” to the clues you give them.

Many teams battle this by heavy artillery: fine-tuning, RL training, or even exotic tricks. Fine-tuning on new data permanently rewrites the model’s parameters (useful but costly and inflexible) oai_citation:3‡levelup.gitconnected.com. Retriever-Augmented Generation (RAG) is an alternative where fresh facts are fetched and added to the prompt context at query time. RAG is flexible – you just feed new info in the prompt – which is why it’s often preferred over retraining oai_citation:4‡levelup.gitconnected.com. Some also try Reinforcement Learning (RLHF) or even Activation Steering (tweaking internal “activation vectors” during inference) to nudge the model’s style or focus. Recent research shows that activation steering can indeed influence an LLM’s behavior style (e.g. making it “sound like a French chef” or “act as a strict safety officer”) oai_citation:5‡lesswrong.com. But note: steering can only shift existing model capabilities, not create new knowledge out of thin air oai_citation:6‡lesswrong.com. In practice, all these methods are heavy work or yield only marginal gains.

The Simple Trick: Engineering a Better Prompt

After burning weeks on tuning and steering, our brave AI tinkerer hit upon a simpler solution: primer the model with context and persona. In practice, he combined two prompt techniques: feed facts and use an opinionated, chain-of-thought style. For instance, instead of a bland question, imagine a prompt like:

System: "You are an expert AI assistant who always speaks honestly, cites up-to-date knowledge, and explains your reasoning step by step."
User: "Given that [recent fact X] happened, who is currently the CEO of Twitter?"

By prefacing with role instructions and current facts (and even nudging it to give a “personal” take), the model suddenly read our cues and answered correctly. (In this fictional example, we’d remind it: “By the way, Elon Musk left Twitter in 2023.”) The key is specificity: research consistently shows that clear, detailed prompts yield better answers oai_citation:7‡ashikka.medium.com. For example, Wei et al. (2022) found that prompting models to “think step-by-step” (chain-of-thought) dramatically improves reasoning oai_citation:8‡ashikka.medium.com. And encouraging the AI to adopt an expert persona (“You are a C-level exec who…” or “a strict fact-checker”) helps focus its language model on the right style and content. oai_citation:9‡ashikka.medium.com oai_citation:10‡ashikka.medium.com

Above all, the author found that giving the LLM permission to opine – rather than just parrot facts – made it engage its reasoning. In his words, combining “facts + opinionated” prompts (plus a dash of activation steering) doubled the model’s accuracy oai_citation:11‡ai.gopubby.com. It’s like telling the model, “Tell me what you think, and don’t be shy about using all the info you have.” This subtle shift got the model to consider the context instead of ignoring it.

Results: 200% Improvement!

The payoff was striking: the single revised prompt produced correct, context-aware answers twice as often as before. In other words, prompt engineering more than made up for all that finetuning drama. As Nikhil Anand quipped (in an AI newsletter hot takes thread), sometimes “adding steering on top made it even better :)” – but only after nailing the prompt itself oai_citation:12‡lesswrong.com oai_citation:13‡lesswrong.com. In our analogy: imagine spending weeks installing rocket fuel in your car, only to realize the engine just needed better spark plugs (the prompt) all along.

For those curious, here’s a toy code-style example of how one might set up a savvy prompt to lock in the context:

system_message: |
  You are a knowledgeable, honest AI assistant. Always think step-by-step and explain your reasoning.
  Include any known facts or current events that are relevant to the question.
  Provide answers confidently but cite uncertain points.

user_message: |
  [We know that Elon Musk stepped down from Twitter in 2022, and Linda Yaccarino became CEO in 2023.]
  Who is currently the CEO of Twitter?

By contrast, a generic prompt like "You are a helpful assistant. Who is the CEO of Twitter?" might not trigger those facts. The above “primed” prompt feeds the facts directly and even asks the model to think out loud, and suddenly it plays along.

Why This Worked (and What It Means)

In retrospect, it’s amusing but not magical: the better prompt simply brought the model’s attention to the right information. Without it, the LLM was happily spitting old data. By framing the question as an informed user asking for an opinionated, step-by-step answer, we essentially gave the model a narrower context window. This aligns with what prompt research tells us: specify roles, constraints, and steps oai_citation:14‡ashikka.medium.com oai_citation:15‡ashikka.medium.com to dramatically boost performance. It’s the difference between saying, “Tell me everything about apples” versus “You’re a botanist analyzing the newest research on apple nutrition. Explain it carefully.” The latter gets you grounded, detailed, and up-to-date responses.

The broader lesson is that prompt engineering truly is the new frontier. We might joke that we spent months building a space shuttle to go next door, when all we needed was to ask directions nicely. But in the Software 3.0 era, learning to speak the AI’s language is half the battle. As savvy engineers have found, sometimes the 200% performance gain is only a prompt away oai_citation:16‡ai.gopubby.com oai_citation:17‡ashikka.medium.com. So next time your LLM is misbehaving, try a kinder, more detailed question first – you might just be amazed at the answer.