You grab your morning cup of coffee as your mind races ahead to the day's writing tasks. The expected progress report. An overdue set of customer service guidelines to draft. And what about that humor article idea you've been thinking about? Your own human brainpower falters in the face of such a demanding workload. But increase your potential productivity by 10 or even 100 times by putting a state-of-the-art large language model (LLM) to work for you through skilled prompting.
"Prompting activates everything the model then produces," points out Scott Count, Ph.D., lead researcher at Anthropic, creators of the constitutional AI assistant Claude. "With the right prompting strategy, over 50% of prompts will reliably produce excellent LLM results." We break down the key facets of prompting to maximize results from mediocre to great.
The Power of Priming
Just as priming prepares houses for painting, prompts should prepare LLMs for upcoming tasks by putting them in the right frame of mind beforehand. Prefacing writing prompts with explicit instructions such as "Write helpful, informative content" or "Use simple vocabulary and provide background context for concepts before discussing them" loads the model with clear expectations before it generates a single word.
Priming prompts also allow you to adjust the model's tone as needed, from stern to silly. Starting a humorous question with "Answer in a playful, joking voice..." gives permission for silliness to follow. Meanwhile, serious questions often benefit from prior sincerity priming.
Priming establishes an output goalpost for LLMs, reducing the likelihood of initial failures requiring recalibration of results.
Phraseology Prudence
Have you ever sent a well-intentioned text message whose unintentional ambiguity sparked an unintended drama? The inherent uncertainty of language interpretation plagues LLMs as well. But prompt phraseology principles prevent misunderstandings:
- Simplify surrounding sentences to highlight central requests rather than burying them in lengthy contexts, minimizing confusion.
- Use active voice and positive framing so models understand exactly what to do, not just what to avoid, mitigating conflicting results.
- Specify any constraints, such as word limits or prohibited content types, up front, and provide appropriate creative guardrails.
- Balance specificity about expected quality, structure, and sources with open-ended language that leaves room for independent insights.
Such judicious wording allows LLMs to showcase their strengths, rather than stumbling over hidden hurdles in the prompts.
Lead by example
Few prompts prove to be completely self-contained. Providing illustrative examples remains elemental. Curating representative sample responses demonstrates response standards better than any descriptive guidelines could ever convey. Treat examples as cornerstone visual aids that anchor learning for LLMs.
Ideally, provide 2-4 example outputs per request type, enough to show a variety of strong results without overwhelming LLMs with excessive data. Be careful, however, that cherry-picked or simplistic examples don't misrepresent actual expectations, leading to recipient disillusionment when promised mastery proves lacking. Samples must illustrate achievable yet aspirational skills.
Most importantly, selected exemplars must be wisely withheld to serve instead as test standards for validating LLM competence on prompted tasks, thereby guarding against overfitting mimicry without robust comprehension. Only by confirming quality performance on completely unseen exemplars can prompt proficiency be legitimately claimed.
Advantages
Priming, thoughtful phrasing, and demonstration through sample responses make communicating desired behaviors to LLMs significantly more achievable. Together, these prompt ingredients provide access to amazing model-generated essays, code, music, designs, and more on demand. Even thorny tasks such as balanced policy analysis or dinner party toasts to long-lost relatives become reasonably accessible with the right prompting protocols.
Such success speaks to underlying LLM aptitudes just waiting to be unleashed through prompt mastery. "I'm still shocked when I see a well-crafted prompt reliably produce a Shakespeare-worthy sonnet on the first try - the creative potential that prompting brings under our control remains magical to me," reflects Anthropic researcher Dr. Elise Nix. The ability to prompt greatly expands what is possible with LLMs.
The downside
But giving LLMs the productive power to produce potentially harmful, misleading, or simply irrelevant content also raises societal concerns, especially given the tendency of models to inherit and reinforce biases. Crafting prompts well, then, is deeply intertwined with crafting them responsibly.
Accordingly, Anthropic researcher Dr. Thomas Mitchell advises prompt creators to "be aware of potential model harms-consider adding safeguards such as sincerity primes or warning classifiers as needed given prompt topics." Prompts also need to be vetted for harmful stereotypes or misinformation. In essence, fine-tuned prompting requires fine-tuned ethics.
Certain applications also have limitations. As Dr. Nix explains, "Highly open-ended, subjective tasks such as composing music or plotting novels currently frustrate standard prompting approaches. Exactly defining 'success' becomes a challenge. Creative fields warrant individualized prompting approaches, pending further research.
Follow these principles when prompting your own LLM and enjoy exponentially multiplied opportunities, safely and skillfully curated. The future of both prompting and LLMs themselves remains bright indeed, thanks to blossoming best practices.
Glossary
- Large Language Model (LLM): A complex AI system trained on vast datasets that is skilled at generating language and mimicking textual styles based on provided prompts and examples.
- Prompting: The method of giving instructions and context to a language model to define the desired objective and constraints on the text that the model should generate.
- Priming: Prefacing a prompt with some initial instructions or examples that 'prime' the model to be in the proper context or state of mind to handle the main prompted task that follows.
- Phraseology: The style, semantics, and syntax used when formulating prompts. Requires care to increase clarity while avoiding confusion.
- Active Voice: A prompt style using verbs in the active rather than passive voice, which provides clarity by emphasizing who or what is expected to perform an action in the generated text.
- Specificity: Providing details explicitly about expectations in a prompt regarding parameters like length, format, deadlines or banned content types. Reduces open-endedness.
- Open-Endedness: Allowing flexibility for the model to apply its own reasoning and creative choices, rather than overly restricting output options through narrow prompts or excessive detail.
- Exemplars: A set of sample responses provided along with a prompt to give concrete examples and standards for the language model to emulate with its own output.
- Overfitting: When a language model mirrors provided example responses too closely without demonstrating genuine comprehension, generalization and reasoning ability on new prompts. Harder to detect.
- Ethics: Careful consideration around potential harms in the application of language models and crafting of prompts to mitigate risks of generating toxic, untruthful or otherwise problematic content.