I don't think it's worth it.

If you are asking about "how to use LLMs for work", first of all I would make sure I have clearance from my company, manager, or whoever would be best to rubber-stamp using LLMs for work-related stuff. Self-hosted LLMs are generally safe even if incredibly wonky, while public ones work more reliably, but are notorious for collecting - and then leaking - customer data. For this reason, I would not use public LLMs for work without approval, and ideally I would do so from a company account set up for me by the company itself, to not risk being accused of leaking confidential info.

Then, if you're good to go with using LLMs, I don't have a strict set of rules for that but the general advice I would give anyone would be as follows:
- Think again if you REALLY need to do it. More often than not LLMs are a nuisance.
- If you think you need to, don't trust your LLM. It is going to hallucinate, get confused, and produce factually incorrect outputs, so always fact-check your LLM.
- Your LLM is going to get confused by too broad scope of the task, and too vague context. Always break down the task into small steps, and give it one assignment at a time, one ask at a time
- Make sure that as you add more and more requirements, your LLM won't lose the previous ones from its context.
- Sometimes it's good to wipe model's context (i.e. start a new conversation) because over time it can really get messy.
- LLMs can be somewhat trustworthy when building stuff with fairly defined and known boundaries, and struggles when the boundaries depend on a zillion factors. For instance, in my experience it generates reasonably good Bash scripts (I spend some time refactoring and honing them, but not rewriting from scratch), but the outputs it generates in Java is plain garbage - it invents libraries and APIs that simply don't exist, or if they exist I wouldn't really use them because they come from dubious sources.

Now, if you're asking about how to run LLMs by yourself for your work-related stuff, it's a different story. I would start with Ollama as it has wide community support and I haven't honestly heard about alternatives. I am too lazy to compile and run it myself, so I went with Ollama in a pre-built Docker image, I can live with the excessiveness of doing so. I have an AMD GPU so I picked ROCm variant of the setup, for NVidia the setup differs a little but it's well laid out anyway. Then you need a model that will run locally reasonably well (both in terms of quality and tokens per second) - I recommend trying Qwen3 in quantized variants, q8, q6 or perhaps q4 to cram more parameters into limited memory. Then for IDE integration you'll need an IDE or IDE plugin that supports local Ollama integration - I tested VSCode + Cline and was mostly satisfied, though the agent tended to get stuck in reasoning loops - "but the user wanted X... I could try Y... or maybe Z.... but the user wanted X.... etc". I also wanted to check if IntelliJ CE can be integrated with Ollama, but didn't have time nor energy recently.