Prompt Engineering
Lil'Log
Lilian Weng dropped a deep-dive on prompt engineering — basically the art of talking to AI models to get what you want without retraining them. Turns out it's less science, more trial-and-error voodoo that changes model to model.
Based on reporting by Lil'Log — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Lilian Weng's latest post tackles a topic that sounds simple until you actually try to do it well: prompt engineering, or what she also calls in-context prompting. The idea is that you steer a large language model's behavior purely through the words you feed it, no gradient updates, no fine-tuning, no touching the weights at all. Just phrasing, examples, and structure doing the heavy lifting.
What's notable is how bluntly Weng frames this as an empirical science rather than a tidy set of rules. A prompting trick that works wonders on one model can flop on another, and there's no universal playbook that transfers cleanly across architectures. That means anyone serious about this has to run experiments, build heuristics, and accept a lot of trial and error along the way. It's less like writing code and more like coaxing a temperamental instrument into tune.
She also draws a clear boundary around scope. This isn't about Cloze-style fill-in-the-blank tasks, image generation, or multimodal setups — it's specifically about autoregressive language models, the kind that predict the next token in a sequence. Narrowing the focus that way lets her go deep on what actually matters for text-based LLMs instead of trying to cover every flavor of generative AI at once.
Underneath all the technique-swapping, Weng ties the whole subject back to something bigger: alignment and steerability. Prompt engineering, in her framing, isn't just a bag of clever tricks for getting better outputs. It's a stand-in for the harder problem of getting a model to actually do what you want, reliably, using nothing but instructions. That connects directly to her earlier work on controllable text generation, suggesting prompting is really just the lightweight, no-training version of a much older challenge in NLP.
My take — AI-written commentary, not fact-checked reporting
I like that Weng is upfront about prompt engineering being messy and model-specific rather than pretending it's a clean discipline — that honesty is rare in an industry obsessed with selling silver-bullet techniques. The real lesson here is that steerability without retraining is fundamentally a patchwork solution, and treating it as the primary alignment tool long-term feels like duct tape on a much bigger structural problem.
Read more about this at: Lil'Log