AI How To: Advanced LLM Prompting
HTML-код
- Опубликовано: 5 фев 2025
- Use code RUclips to get an extra 20% off my full Cursor course here:
www.jointakeof...
In this video we use a real-world agentic workflow to explore several advanced LLM prompting techniques.
Repo Link: github.com/mck...
Follow me on...
Twitter/X: / mckaywrigley
GitHub: github.com/mck...
This is awesome man. I started following shortly after the original chat-gpt hype, and have been patiently waiting for this type of implementation. Can't wait to see where it goes as speed/accuracy/context length improve. Well done
you are my AI Engineering GOAT
I will look through your repo, lot's of good stuff in here.
This is awesome. Thanks for sharing. That spec & planning steps can be leveraged for other areas as well.
Chaining those 3 steps is super useful on anything that requires an LLM to do “reasoning”. Try it on math for example - great results!
Good job. Need to add openrouter support =)
Would be awesome if you could include examples of helpful PRs this setup has generated for you
Will probably do a follow up once our system is a little more polished!
Gonna save this, till I get to this level of knowledge where I leverage a framework like this. Currently I'm just trying to find an automated way to preprocess epub data to increase LLM reasoning and problem solving.
Only three step needed to get functional code output?
Depends on how complex your workflow is. Very case dependent, but it's a good place to start experimenting from.
why use xml tags over json output?
JSON output/function calling tends to result in lower performance for tasks that require higher levels of reasoning like code generation.
@@realmckaywrigley do you have any resources that go over this topic or is this based on your experience?
So for anyone who may read this. There are a couple of reasons why you might want to use XML tags over JSON output.
1. Well trained on XML tag's
Some models are trained on XML tags explicitly. Claude for example has a good understand of XML tags. Using XML in the input will result in better understanding of the prompt and result in better structured outputs.
2. No json mode or malformed json outputs
If a model doesn't have the ability to output in json or the outputs are sometimes malformed. Its often better to output using XML tags. In my testing LLM's have an easier time generating and formatting XML vs JSON.
3. Ability to "think" and output a structured response in one prompt
I use to use one model to generate the content and then another to format into JSON. It's cheaper and faster to use tags to generate the content and then provide an XML schema that you can parse all in one response.
Example output might look something like this:
Thoughts
Unique Name
A detailed description that I will parse later
Thanks for all your content! I have a business proposition for you. How can I get in touch?
Hi, Michael! Unfortunately right now I'm a bit swamped and not open to business proposals at the moment.
alright!