The Day I Stopped Writing “Pretty” Prompts

I spent more time polishing the prompt than reading the answer. I chose fancy words. Structured sentences with corporate precision. Added “kindly elaborate an executive synthesis regarding the aforementioned prerogatives” thinking that impressing the AI would yield better results.

I was wrong. And science just proved why.

A paper titled “Adam’s Law: Textual Frequency Law on Large Language Models,” published in the Proceedings of the 64th Annual Meeting of the ACL (July 2026, San Diego) by researchers from the Chinese University of Hong Kong, showed something completely counterintuitive: swapping a sophisticated prompt for a simple, direct command makes the AI considerably smarter.

+8% accuracy. Consistent across all models tested. Without changing the problem. Without changing the model. Just changing the words.

When I read it, my first reaction was “this can’t be right.” Then I tested it. It was right.

The Experiment: Sophistication vs. Simplicity

The methodology was elegant. Researchers took real math problems and rewrote them in two forms:

Hyper-polished version: rare vocabulary, formal, complex intellectual terms. The kind of language an academic paper uses.

Direct version (“dumbed-down”): the simplest, most common words possible. The kind of language someone would use on Reddit.

Crucial point: human reviewers verified that the core meaning and difficulty level were exactly the same. No problem simplification — only word choice. The resulting dataset (TFPD — Textual Frequency Paired Dataset) covers math reasoning, machine translation, commonsense reasoning, and agentic tool calling.

The result was uniform: when confronted with simpler language, models became ~8% more accurate instantly.

The Numbers

Gains were solid and consistent on the GSM8K benchmark (math reasoning):

DeepSeek-V3: from 63.55% (complex vocabulary) to 71.54% (simple vocabulary). +7.99%.

LLaMA 3.3 (70B-Instruct): from 80.49% to 88.75%. +8.26%.

GPT-4o-mini: from 60.70% to 68.70%. +8.00%.

The consistency (~8% across completely different models) is what makes the finding so robust. It’s not a statistical accident — it’s a structural pattern.

And the phenomenon extends beyond math. Translation tests across hundreds of languages showed the same dynamic. When researchers applied CTFT (Curriculum Textual Frequency Training) — ordering training data from least to most frequent — translation accuracy in some languages increased by nearly 30%.

Adam’s Law: Why This Happens

The explanation is elegant and connects directly to how LLMs work.

LLMs are trained on the internet’s ocean of data. In that ocean, common words appear millions of times more than academic or fancy terms. The model developed dense, robust representations for frequent language — and sparse, uncertain representations for rare language.

Common words → high-confidence zone. The model operates in dense latent space regions. Understands your intent without ambiguity. Dedicates computational capacity to the problem, not to vocabulary decoding.

Complex words → low-confidence zone. The model gets pushed into sparse latent space regions. Spends energy trying to decipher rare vocabulary. Less capacity remains for actual reasoning. Result: guessing with pretty grammar.

The metaphor that worked for me: it’s like asking a local for directions. If you speak their language with native accent, they understand instantly and focus on helping you. If you speak with a heavy accent and formal vocabulary, they spend half the conversation trying to understand you — and the directions get worse.

The Connection to Everything I’ve Written

This paper is the scientific validation of something I’ve been feeling in previous posts:

Context engineering. In “Beyond the Prompt,” I argued the future isn’t the perfect prompt — it’s the perfect context. Adam’s Law confirms: how you formulate the prompt matters less than you think. What matters is clarity, not sophistication.

Stanford’s harness study. The paper showed that context format (YAML, Markdown, JSON) had statistically insignificant effect. “Familiarity beats compression.” Adam’s Law says the same thing from another angle: familiarity (common words) beats sophistication (rare vocabulary).

Microsoft’s SkillOpt. The rules SkillOpt discovered automatically read like simple, direct instructions — “Workbook Forensics,” “Evidence Binding.” Not elaborate corporate prompts. The AI optimized for simplicity.

The Magic Effect. If people with less AI literacy are more receptive because they find the technology “magical,” perhaps part of it is that they write simpler prompts — and therefore get better responses. The “magic” may be partially explained by high-frequency language.

Practical Guide: How to Write Prompts Going Forward

Adam’s Law’s lesson is liberating. Next time you open the AI chat:

Avoid excessive corporate speak. Instead of “Kindly elaborate an executive synthesis regarding the aforementioned prerogatives,” try “Summarize the main points of the text above directly.”

Focus on the goal, not the appearance. Models respond better to logical clarity than to aesthetic word flourishes. Your prompt isn’t an email to the CEO — it’s an instruction for a machine that works better with common language.

Think Reddit. A good rule: write as if you’re explaining your question to a regular person in a forum. No unnecessary jargon. No excessive formality. Straight to the point.

Test the simple version first. If an elaborate prompt isn’t giving good results, try rewriting with the most common words you know. Data suggests +8% accuracy — for free.

The most valuable skill now isn’t impressing the algorithm with scholarly vocabulary — it’s translating complex problems into the cleanest, most concise, most common language possible.

Conclusion: Simplicity Is the New Sophistication

Prompt engineering didn’t die — but it changed focus. Value is no longer in elaborate formulas and impressive jargon. It’s in clarity.

And this reminds me of what I wrote about context engineering: the best prompt in the world can’t save bad context. And now we know a “pretty” prompt can actually be worse than a simple one — because it pushes the model into low-confidence zones in latent space.

AI isn’t impressed by your vocabulary. It’s confused by it.

Share if this simplified your life:

+8% accuracy. Without changing the model. Without changing the problem. Just using common words. Simplicity is the ultimate sophistication — and science just proved it.

Read Also