Comment by only-one1701 2 days ago
I see this "prompting is an art" stuff a lot. I gave Claude a list of 10 <Route> objects and asked it to make an adjustment to all of them. It gave me 9 back. When I asked it to try again it gave me 10 but one didn't work. What's "prompt engineering" there, telling it to try again until it gets it right? I'd rather just do it right the first time.
I think the point is you could just.... write the code yourself.
Then don't use it? Nobody is making you.
I am also barely using LLMs at the moment. Even 10% of the time would be generous.
What I was saying is that I have tried different ways of interacting with LLMs and was happy to discover that the way I describe stuff to another senior dev actually works quite fine with an LLM. So I stuck to that.
Again, if an LLM is not up to your task, don't waste your time with it. I am not advocating for "forget everything you knew and just go ask Mr. AI". I am advocating for enabling and productivity-boosting. Some tasks I hate, for some I lack the deeper expertise, others are just verbose and require a ton of typing. If you can prompt the LLM well and vet the code yourself after (something many commenters here deliberately omit so they can happily tear down their straw man) then the LLM will be a net positive.
It's one more tool in the box. That's all there is to it really. No idea why people get so polarizing.
Prompt engineering is just trying that task on a variety of models and prompt variations until you can better understand the syntax needed to get the desired outcome, if the desired outcome can be gotten.
Honestly you’re trying to prove AI is ineffective by telling us it didn’t work with your ineffective protocol. That is not a strong argument.
What should I have done there? Tell it to make sure that it gives me all 10 objects I give it back? Tell it to not put brackets in the wrong place? This is a real question --- what would you have done?
How long ago was this? I'd be surprised to see Claude 3.7 Sonnet make a mistake of this nature.
Either way, when a model starts making dumb mistakes like that these days I start a fresh conversation (to blow away all of the bad tokens in the current one), either with that model or another one.
I often switch from Claude 3.7 Sonnet to o3 or o4-mini these days. I paste in the most recent "good" version of the thing we're working on and prompt from there.
In no particular order:
* experiment with multiple models, preferably free high quality models like Gemini 2.5. Make sure you're using the right model, usually NOT one of the "mini" varieties even if its marketed for coding.
* experiment with different ways of delivering necessary context. I use repomix to compile a codebase to a text file and upload that file. I've found more integrated tooling like cursor, aider, or copilot, are less effective then dumping a text file into the prompt
* use multi-step workflows like the one described [1] to allow the llm to ask you questions to better understand the task
* similarly use a back-and-forth one-question-at-a-time conversation to have the llm draft the prompt for you
* for this prompt I would focus less on specifying 10 results and more about uploading all necessary modules (like with repomix) and then verifying all 10 were completed. Sometimes the act of over specifying results can corrupt the answer.
[1]: https://harper.blog/2025/02/16/my-llm-codegen-workflow-atm/
I'm a pretty vocal AI-hater, partly because I use it day to day and am more familiar with its shortfalls - and I hate the naive zealotry so many pro-AI people bring to AI discussions. BUTTT we can also be a bit more scientific in our assessments before discarding LLMs - or else we become just like those naive pro-AI-everything zealots.
We used to make fun of and look down on coders who mindlessly copy paste and mash the compile button until the code runs, for good reasons.