How I Learned to Stop Saying 'Please' to My AI and Other Principles I Fail to Follow
I Recently got enlightened by a paper titled "Principled Instructions Are All You Need for Prompting Language Models" by Yang et al. (2023), that I've been prompting AI like I'm chatting with my overly sensitive Ego. Apparently, AI doesn't care how politely I ask—it just wants me to be clear, concise, and preferably, blunt. In honor of my newfound clarity (and ongoing incompetence), I've compiled an utterly scientific and Factual! table highlighting just how badly I've been messing up.
Behold vignettes showcasing traits of an LLM Whisperer:
| Category x (5) | Principle x (26) | Before Prompt (my usual embarrassing style) |
After Prompt (bossy and effective) |
Expected Outcome (Before) | Expected Outcome (After) |
|---|---|---|---|---|---|
| Prompt Structure & Clarity | Stop saying please (AI doesn't care) | Please explain photosynthesis, if you're not busy. | Explain photosynthesis. | AI rolls its eyes and rambles. | AI cuts to the chase. |
| Audience matters (stop confusing your AI) | Explain photosynthesis. | Explain photosynthesis like I'm five. | Explanation is either a Ph.D. thesis or utter gibberish. | A simplified explanation even I can understand. | |
| Use structured format (pretend you’re organized) | Explain cellular respiration. | ###Instruction### Explain cellular respiration. | AI spills a word salad. | AI neatly arranges a tidy response. | |
| Think step-by-step (like kindergarten) | Solve: 2 + 2 * 3 = ? | Solve: 2 + 2 * 3 = ? Think step by step. | "8, trust me bro." | Detailed calculation I can follow (and still question). | |
| Delimiters (use fancy brackets) | Summarize this: [text chaos] | Summarize this: [cleanly bracketed chaos] | AI wonders what it’s reading. | AI confidently summarizes like a seasoned librarian. | |
| Output primers (start it off right) | Write a poem about the ocean. | Write a poem about the ocean. The vast blue… | Randomly starts anywhere. | Elegantly picks up from "The vast blue..." | |
| Affirmative directives (stop confusing the AI with negatives) | Don't forget gravity examples. | Include gravity examples. | AI conveniently forgets. | AI obediently provides examples. | |
| Specificity & Information | Few-shot prompts (copy smarter people) | Capital of France? | Italy: Rome. Spain: Madrid. France? | Brief confusion followed by "Paris?" | Instant "Paris" because pattern recognition. |
| Different levels of explanations (pretend you're a child) | Explain quantum mechanics. | Explain quantum mechanics like I'm 11. | Quantum confusion overload. | Simplified quantum gibberish even kids get. | |
| Avoid stereotypes (don't get cancelled) | Describe a CEO. | Describe a CEO without stereotypes. | AI serves up cliché stew. | AI tactfully avoids lawsuits. | |
| Mimic provided style (AI does impressions now) | Write a cat story. | Cat story: fluffy and sunlit. Now dog story. | Completely unrelated writing style. | Dog gracefully leaps in sunlit mimicry. | |
| Specific beginnings (set AI boundaries) | Write a story. | Start: Once upon a time… | "It was a dark and stormy night" (no boundaries). | Perfectly begins "Once upon a time." | |
| Clear constraints (limit AI enthusiasm) | Summarize this book. | Summarize this book in three sentences. | AI writes a novella instead. | Succinct and to the point—miracle achieved. | |
| Task directives (get bossy with AI) | Summarize this article. | Your task: Summarize article. You MUST list 3 points. | Random wandering summary. | Precise 3-point summary delivered with a salute. | |
| User Interaction & Engagement | Let AI interrogate you (for better results) | Write a business plan. | Ask questions first, then write a business plan. | Business plan built on guesses. | Painful but accurate Q&A session, then solid plan. |
| Teach me with a test (fake enthusiasm for learning) | Explain Pythagoras. | Explain Pythagoras and quiz me. | Quick theorem dump and goodbye. | Interactive teaching moment with humiliating quizzes. | |
| Content & Language Style | Grammar correction (but keep it human) | Fix this sentance pleese. | Fix grammar but don’t ruin my style. | Robotic grammar enforcement. | Human-friendly grammar fixes preserving my charm. |
| Natural language (AI tries to be less creepy) | Weather today? | Human-like question: Weather today? | Robotic monotone response. | Casual, friendly weather chat. | |
| Assign a role (AI roleplay) | Explain relativity. | You're Einstein now, explain relativity. | Dry textbook explanation. | Authoritative Einstein impression (German accent optional). | |
| Repetition for emphasis (nagging works) | Summarize report. | Summarize report. Briefly summarize report. | Lengthy, unfocused summary. | Short, crisp summary due to nagging repetition. | |
| Tip incentives (bribe your AI) | Explain blockchain better. | Explain blockchain! $100 tip if better! | Unmotivated explanation. | Overachieving AI craving digital money. | |
| Complex Tasks & Coding Prompts | Break it down (like your therapist recommends) | Write complex economic essay. | First explain policy A, then B, then compare, then propose new. | Disorganized economic rant. | Structured masterpiece worthy of academic praise. |
| Multi-file coding (when one file isn't confusing enough) | Python program, main and module. | Generate script that auto-creates main and module files. | Scattered code fragments. | Organized, executable Python files auto-generated (witchcraft suspected). | |
| Combine CoT with few-shot (teach AI math, slowly) | 15 * 17? | 23=6, 54=20, now 15*17 step by step. | Direct answer possibly incorrect. | Painfully clear, step-by-step math demonstration. |
By all means, try these yourself—just don't expect your LLM/HLI (Human Language Interface) to thank you when you finally get it right.