The hidden cost of AI’s polished prose #
Ever notice how AI-generated text often feels a bit… fluffy? LLMs frequently add pleasantries, sprinkle in synonyms, and craft elaborate sentences. While seemingly polished, this verbosity slows readers down, making content harder to absorb.
When we give vague instructions like “write clearly,” the model gains too much freedom. It might choose longer sentences, switch technical terms unnecessarily, or adopt a tone that doesn't quite fit our needs. This "elegant variation" can obscure meaning and make critical information less accessible.
Instead, think about what you want your AI to do. Our goal isn't just "clear" text; it's text that's readable and easy to review. This approach doesn't guarantee factual correctness, of course—you still need to check the facts! But it does create an "output contract" that gives you more control over the AI's writing style.
This is exactly what Andrej Karpathy, an OpenAI founding member, explored. He found that by prompting LLMs to write like an aircraft manual, you can dramatically improve the utility of their output. The secret lies in a specific technical standard designed for clarity above all else.
Borrow the rulebook built for aircraft crews #
Borrowing a rulebook built for aircraft crews might sound extreme, but it's a proven solution. Enter ASD-ASD-STE100, or Simplified Technical English, a standard developed in the 1970s for international aircraft maintenance crews. This standard ensures that critical repair instructions are clear and unambiguous, regardless of the reader's native language.
ASD-ASD-STE100 uses a controlled vocabulary of roughly 900 approved general words. Each word has a single, defined meaning and part of speech. For instance, you "use" a tool, you do not "utilize" it. This eliminates synonyms and ensures consistent terminology, preventing confusion.
Beyond vocabulary, ASD-ASD-STE100 enforces strict grammatical rules. Sentences must be short, with procedural instructions commonly limited to 20 words. Each sentence contains only one instruction, eliminating complex phrasing and the potential for misinterpretation.
The standard also bans semicolons and contradictory phrasing, further simplifying communication. By applying these rules, aircraft maintenance manuals become incredibly precise, minimizing errors and maximizing safety. This disciplined approach is exactly what Andrej Karpathy and others leverage to make AI output equally clear.
The prompt that turns rules into an output contract #
Ready for a reusable prompt? Here’s the core instruction Karpathy uses to make AI outputs precise: “Write about 80% of the way to ASD-ASD-STE100 Simplified Technical English. Use short sentences, common words, consistent terms, and one action per instruction. Keep the tone natural.” This prompt transforms vague requests into a concrete output contract.
Naming a recognized standard like ASD-ASD-STE100 gives the LLM far more specific constraints than simply asking it to “be concise” or “explain clearly.” LLMs are deeply familiar with ASD-ASD-STE100 from their training data, allowing them to apply its rules on sentence length (20 words max for instructions), controlled vocabulary (900 approved words), and the one-action-per-instruction principle.
The "80% of the way" caveat is crucial. While full ASD-ASD-STE100 is highly effective for critical documentation, it can feel overly rigid for everyday explanations. This adjustment helps the AI maintain readability without sacrificing clarity.
To further refine the output, add details about your audience, the task, and the desired format. You can also invite the model to flag terms or details it cannot simplify without altering their meaning. For more details on the standard, visit the ASD-ASD-STE100 Simplified Technical English Maintenance Group.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Why 80% beats strict compliance #
Andrej Karpathy makes a good point: aiming for 80% of ASD-ASD-STE100 gives you the best of both worlds. Full adherence, while incredibly precise, can feel a bit stiff or unnatural for general use. The 80% target preserves the core clarity and conciseness without sacrificing readability or flow.
This calibrated approach shines in specific contexts. Use it for:
-
Technical explanations
-
Procedural guides
-
Internal documentation
-
Initial AI-generated drafts However, recognize its limits. Applying this method to creative writing, nuanced discussions, or content requiring a strong, distinct voice might flatten the output, removing important stylistic elements.
Always treat the prompt as an output contract, not an accuracy check. After the AI generates its response, critically compare it to the original input. Verify all facts and instructions. If the output misses the mark, adjust your prompt or provide more context rather than expecting the standard to magically correct factual errors.
Frequently Asked Questions #
What is ASD-STE100?
ASD-STE100, or Simplified Technical English, is a controlled language standard developed for clear technical documentation, including aircraft maintenance manuals.
How do I ask an AI to use ASD-STE100?
Ask it to write about 80% of the way to ASD-STE100, then specify your audience and whether you want an explanation or instructions.
Why does Andrej Karpathy recommend using only 80% of the standard?
Full compliance can make everyday writing sound rigid. An approximate version aims to keep the clarity while allowing more natural flow.
Does using ASD-STE100 make AI answers more accurate?
No. It can make wording clearer and less ambiguous, but it does not verify facts or prevent errors.