Prompt engineering
Prompt engineering for large language models is prompting debate over whether users should need specialized techniques to get good results from tools like ChatGPT. Commenters trade resources and tactics—such as detailed system prompts, structured outputs, and breaking tasks into steps—while questioning if this really counts as "engineering" or is just trial-and-error communication. Many expect these quirks to fade as models and interfaces improve, but note that learning to craft clearer prompts already exposes broader lessons about human communication, ambiguity, and user interface design.
Status of the OpenAI Guide & Alternatives
- Many note the OpenAI guide is not new; it’s just appearing on HN for the first time.
- Several commenters say there are better or more complete resources (other guides, courses, GitHub repos).
- Some see the guide as a good, simple intro; others say it only covers “basic demos” and omits powerful patterns like complex system prompts and structured outputs.
Is Prompt Engineering Necessary?
- One camp argues it shouldn’t exist: if language models truly understood natural language, plain speech plus good UX should suffice.
- Others reply that current LLMs are imperfect, and tailored prompting measurably improves results.
- Comparisons are made to how humans also need training to communicate clearly; ambiguity and missing context are common failure modes in both humans and LLMs.
- Some expect prompt engineering to fade as models and interfaces improve, similar to how low-level hardware constraints became less visible in computing.
Is It Really “Engineering”?
- A long subthread debates the term: some call it trial-and-error “craft,” not engineering.
- Others say iterative tweaking under constraints is exactly what many engineering disciplines look like.
- Alternative labels suggested: “prompt crafting,” “context composition.”
- Underlying tension: what level of theory/rigor is required before something merits the “engineering” label.
Techniques & Practices Discussed
- Common advice: be explicit, provide context and reference text, split complex tasks, and allow “thinking” steps.
- Heavy use of long, precise system prompts; every undesired behavior leads to another rule or example.
- Structured outputs via JSON schemas, function calling, and external tools are seen as powerful but complex.
- Tone and framing matter: strict commands, emotional language (“enjoy,” “embarrassed”), even threats can change behavior, which unsettles some commenters.
UX, Interfaces, and Future Evolution
- Comparisons to Google search: prompt engineering is likened to advanced search operators.
- Some expect a shift toward structured or hybrid query languages (SQL-like, CLI-like) and richer GUIs with filters and options.
- Others predict spoken, conversational interfaces will dominate as speech-to-text and LLMs converge.
Evaluation, Reliability & Safety Concerns
- Several criticize evaluations that ignore prompt engineering, arguing they understate model capability.
- Counterpoint: hand-tuning prompts per example is “cheating” and not viable in real systems.
- Prompt injection and the lack of a clear boundary between “instructions” and “data” are seen as fundamental issues that impair reliability and formal reasoning.