- Author
- Posted by AlexC04, with replies from many commenters
- Published
- 19 August 2025
- Kind
- Hacker News discussion thread of a Medium essay
- Original
- news.ycombinator.com/item?id=44951287
- Subject
- Harnesses and agent loops
What it is
The thread discusses a Medium essay by alexchesser, who joins as AlexC04. The essay argues that prompt structure shapes where a model's attention goes.
What it says
- Commenters noted the essay shows prompts but no outputs. The poster cited a production limit.
- Critics said the attention analogy lacks proof and may confuse two meanings of attention.
- The commenter trjordan agreed structure and position matter but disputed several technical details.
- Commenter jwilber said almost no API-served model uses standard attention, so the analogy is wrong.
- The commenter energy123 vouched for the advice, whatever the cause.
- Others suggested asking an LLM to rewrite prompts, or giving it a JSON skeleton.
- A side thread debated whether prompt engineering deserves the word engineering.
What evidence it gives
The thread is opinion and anecdote. AlexC04 links a personal branch of the Hugging Face transformers repository as example output. AlexC04 also relays praise from colleagues. No commenter posts a controlled test.
Interests and limits
The poster is the essay's author and says the title was chosen for clicks. Comments cannot be verified, and technical claims in them are unchecked. The Medium essay itself was not read. Only the thread, its replies and the linked claims were read.