Then I watched a someone familiar with the codebase ask Claude to build the thing, in precise terms matching their expertise and understanding of the code. It worked flawlessly the first time.
Neither of us "coded", but their skill with the underlying theory of the program allowed them to ask the right questions, infinitely more productive in this case.
Skill and understanding matter now more than ever! LLMs are pushing us rapidly away from specialized technicians to theory builders.
Good LLMing seems to be about isolating the right information and instructing it correctly from there. Both the context and the prompt make a tremendous difference.
I've been finding recently that I can get significantly better results with fewer tokens by paying mind to this more often.
I'm definitely a casual though. There are probably plenty of nuances and tricks I'm unaware of.
It makes sense considering that these practices could be thought of as "institutionalized skills."
It's a K-type curve. People that know things will benefit greatly. Everyone else will probably get worse. I am especially worried about all young minds that are probably going to have significant gaps in their ability to learn and reason based on how much exposure they've had with AI to solve the problems for them.
Of course, but how do you begin to understand the "stochastic parrot"?
Yesterday I used LLMs all day long and everything worked perfectly. Productivity was great and I was happy. I was ready to embrace the future.
Now, today, no matter what I try, everything LLMs have produced has been a complete dumpster fire and waste of my time. Not even Opus will follow basic instructions. My day is practically over now and I haven't accomplished anything other than pointlessly fighting LLMs. Yesterday's productivity gains are now gone, I'm frustrated, exhausted, and wonder why I didn't just do it myself.
This is a recurring theme for me. Every time I think I've finally cracked the code, next time it is like I'm back using an LLM for the first time in my life. What is the formal approach that finds consistency?
You also have to treat this as outsourcing labor to a savant with a very, very short memory, so:
1. Write every prompt like a government work contract in which you're required to select the lowest bidder, so put guardrails everywhere. Keep a text editor open with your work contract, edit the goal at the bottom, and then fire off your reply.
2. Instruct the model to keep a detailed log in a file and, after a context compaction, instruct it to read this again.
3. Use models from different companies to review one another's work. If you're using Opus-4.5 for code generation, then consider using GPT-5.2-Codex for review.
4. Build a mental model for which models are good at which tasks. Mine is:
3a. Mathematical Thinking (proofs, et al.): Gemini DeepThink
3b. Software Architectural Planning: GPT5-Pro (not 5.1 or 5.2)
3c. Web Search & Deep Research: Gemini 3-Pro
3d. Technical Writing: GPT-4.5
3e. Code Generation & Refactoring: Opus-4.5
3f. Image Generation: Nano Banana ProThat was using pay per token.
> Write every prompt like a government work contract in which you're required to select the lowest bidder, so put guardrails everywhere.
That is what I was doing yesterday. Worked fantastically. Today, I do the very same thing and... Nope. Can't even stick to the simplest instructions that have been perfectly fine in the past.
> If you're using Opus-4.5 for code generation, then consider using GPT-5.2-Codex for review.
As mentioned, I tried using Opus, but it didn't even get the point of producing anything worth reviewing. I've had great luck with it before, but not today.
> Instruct the model to keep a detailed log in a file and, after a context compaction
No chance of getting anywhere close to needing compaction today. I had to abort long before that.
> Build a mental model for which models are good at which tasks.
See, like I mentioned before, I thought I had this figured out, but now today it has all gone out the window.
It’s like I need a sticky disclaimer:
1. No, I didn’t form an outdated impression based on GPT-4 that I never updated, in fact I use these tools *constantly every single day*
2. Yes, I am using Opus 4.5
3. Yes, I am using a CLAUDE.md file that documents my expectations in detail
3a. No, it isn’t 20000 characters or anything
3b. Yes, thank you, I have in fact already heard about the “pink elephant problem”
4. Yes, I am routinely starting with fresh context
4a. No, I don’t expect every solution to be one-shotable
5. Yes, I am still using Opus fucking 4.5
6. At no point did I actually ask for Unsolicited LLM Tips 101.
Like, are people really suggesting they never, ever get a suboptimal or (god forbid) completely broken "solution" from Claude Code/Codex/etc?That doesn't mean these tools are useless! Or that I’m “afraid” or in denial or trying to hurt your feelings or something! I’m just trying to be objective about my own personal experience.
It’s just impossible to have an honest, productive discussion if the other person can always just lob responses like “actually you need to use the API not the 200/mo plan you pay for” or “Opus 4.5 unless you’re using it already in which case GPT 5.2 XHigh / or vice versa” to invalidate your experience on the basis of “you’re holding it wrong” with an endlessly slippery standard of “right”.
> to invalidate your experience on the basis of “you’re holding it wrong”
This was not my intent in replying to 9rx. I was just trying to help.
Its been 12 hours and all the image gen tools failed miserably. They are only good at producing surface level stuff, anything beyond that? Nah.
So sure, if what you do is surface level (and crap in my opinion) ofc you will see some kind of benefit. But if you have any taste (which I presume you dont) you would handily admit it is not all that great and the amount invested makes zero sense.
I write embedded software in C for a telecommunications research laboratory. Is this sufficiently deep for you?
FWIW, I don't use LLMs for this.
> But if you have any taste (which I presume you dont)
What value is there to you in an ad hominem attack here? Did you see any LLM evangelism in my post? I offered information based on my experience to help someone use a tool.
Somehow the more senior you are [in the field of use], the better results you get. You can run faster and get more done! If you're good, you get great results faster. If you're bad, you get bad results faster.
You still gotta understand what you're doing. GeLLMan Amnesia is real.