LLMs reward expertise

TL;DR Highlight

Short, specific, expert-style prompts like Terence Tao's get far more out of the same LLM.

Who Should Read

A backend or full-stack developer frustrated that LLM coding results are bland or head in the wrong direction.

Core Mechanics

LLMs turned everyone into generalists
- In the 2010s, a CSS gap meant leaning on a colleague or hoping the internet had your answer.
- Now anyone can delegate to an LLM and get sort-of-okay CSS.
Prompting skill is domain expertise
- The author argues the most important prompting skill is expertise in the domain you prompt about.
- He rejects the idea that skilled prompters and first-timers get the same results.
Tao's messages are short and to the point
- Terence Tao doesn't answer the model point by point, only the gist.
- The author says signalling expertise shifts the model into a talking-to-mathematicians mode with more concise outputs.
Push back with unease, not contradiction
- When a response looks wrong, Tao says things like "this looks more complex than I was hoping for".
- He rarely takes the model's advice on where to go next and makes his own leaps.

Community Reactions

A novice got stuck because she lacked the vocabulary
- A commenter's non-programmer friend circled in feature brainstorming with an AI for 1.5 hours and they stopped the experiment.
- The commenter says one message like "write an html page that does X, Y, Z" would have worked.
Stating your background changes the output
- A commenter says declaring deep biblical scholarship and asking for no translation of Greek or Latin changes results significantly.
- Mentioning 20+ years of C experience works similarly, they say.
More token burn, worse output
- A commenter who manages engineers says API stats show a strong inverse correlation between tokens burned and output produced.
- People who burn more tokens are less likely to get a good outcome.
Counterpoint: expertise isn't required
- A commenter says an Anthropic mathematician's prompt was basically "suppose you must resolve this conjecture, think really hard".
- Another notes Tao's chat was for building intuition, not solving the problem outright.

How to Apply

Open prompts with your background and expected level
- In today's coding prompt, start with "I've run this codebase for N years; skip the basics".
- Run the same question with and without it and compare answer length and depth.
Push back with unease statements
- When an answer seems off, say "this looks more complex than I hoped" or "couldn't this be simpler?".
- Cite codebase facts like "don't we already do X?".
Ask one specific question at a time
- Don't reply point by point to long answers; ask one question like "given X and Y, why A?".
- Decide the next direction yourself instead of waiting for the model's suggestion.
Use the model in learning mode for unfamiliar areas
- For an unfamiliar codebase or field, start with "explain at my level" before assigning work.
- Once you understand it, switch to the expert-mode prompt.

Try It

copy & run

# Paste into Claude/ChatGPT (a lightweight prompt version of the article's observations)

I have worked on {{domain, e.g. this payments backend}} for {{N}} years. Skip the basics and answer concisely, as to a fellow engineer.

Context: {{current code/problem description}}
Question: {{Given condition X, why does A happen? Just one thing}}

Rules:
- Don't respond point by point; focus on the gist.
- If I say "this looks more complex than I hoped", propose a simpler alternative.
- I decide what to do next. Keep any suggestion to one line.

Try next: run the same question again without the background line and compare answer length and depth.

Terminology

Jacobian ConjectureA famous open math problem; the article says a counterexample was recently found. It appears here only as the subject of Tao's ChatGPT conversation.

Domain expertiseDeep knowledge of a specific area such as a codebase or math. Like a driver who knows the map and gives a navigator a precise destination.

Vibe codingLetting an LLM write code without reviewing it closely. Comments mention it in the context of its limits without expertise.