Precision Context Over Natural Language Only | Generated by AI

Home 2026.03 English 中文

---

Question: Is it incorrect to call it the “LLM’s language”? Is the more accurate idea that prompts should include concrete context (like CSS selectors, HTML, or code) rather than just natural English?

Answer:

Yes — saying “LLM’s language” is not accurate. Your corrected understanding is better.

LLMs do not have a special language. They accept any text:

• English
• Chinese
• code
• HTML
• CSS selectors
• logs
• stack traces
• JSON
• mixed content

What actually matters is context precision, not “LLM language”.

Your idea can be expressed more precisely like this:

Core Idea

LLMs work best when the prompt contains concrete, machine-level context, not vague natural language.

Bad prompt (vague):

Click that button and extract the text

Good prompt (precise context):

Click this element:
<button class="btn-primary submit-order" data-testid="checkout-submit">

CSS selector:
button[data-testid="checkout-submit"]

Then extract text from:
<div id="order-confirmation">

The second works better because:

• removes ambiguity
• reduces hallucination
• gives exact DOM structure
• matches what the model sees in code

---

Why English-only prompts are weaker

Natural language:

click the blue button on the right

Problems:

• multiple blue buttons
• layout may change
• model must guess

But this is deterministic:

document.querySelector('#checkout > button.primary')

LLM doesn’t guess — it executes reasoning directly.

---

So the real principle

Not “LLM language”, but:

LLMs prefer structured context over descriptive language

Best prompt = mix of:

• natural language (intent)
• code (precision)
• selectors (location)
• logs (state)
• errors (feedback)

Example ideal prompt:

Goal: click checkout button

HTML:
<button id="checkout-btn" class="primary">

Selector:
#checkout-btn

Error:
Element not clickable at point (532, 410)

What's wrong?

This is extremely strong.

---

Why Copilot / Claude Code work better with copied code

Because they are context-driven, not language-driven.

When you paste:

• VSCode code
• DOM HTML
• CSS selectors
• stack trace

You reduce search space for the model.

Less guessing → better output.

---

Your corrected statement (most accurate)

Your improved understanding is correct:

• LLMs don’t need special language
• any text works
• but prompt + context must be self-consistent
• copying real code/HTML/selectors improves results

This is exactly right.

---

One-sentence summary

LLMs don’t require a special language — they perform best when prompts include precise, real context (code, selectors, logs) instead of vague natural language.

---

Generated by AI. Curating and sharing still takes effort. If you find it useful, feel free to donate. WeChat: @lzwjava · X: @lzwjava · Say hi 👋

Back Donate