Skip navigation
Here’s a 17min deep dive on advanced prompting techniques for LLMs.
0:00 / 17:22

Here’s a 17min deep dive on advanced prompting techniques for LLMs.

Mckay Wrigley

@mckaywrigley230K followers

Visit on X
2,084 likesDownloadOriginal post

196,696 views

Here’s a 17min deep dive on advanced prompting techniques for LLMs. Fully demonstrated on a real-world, multi-step AI workflow. Watch for a complete breakdown.

10 Comments

Comments from the original post on X (Twitter)

  • Mckay Wrigley's profile picture

    @mckaywrigley

    The video covers: - prompt chaining - chain-of-thought with <scratchpad> tags - xml tags - system vs. user messages - output parsing - prefilling - information hierarchy - role prompting - goal prompting - recursive llm calls Lots of good stuff in there!

    63 likes

  • Mckay Wrigley's profile picture

    @mckaywrigley

    Link to the repo: Note that the prompts in the video are on the dev branch. Still more work to be done before we merge latest updates to main. Dev branch is experimental and may break when run.

    52 likes

  • Mckay Wrigley's profile picture

    @mckaywrigley

    One thing that I hope that this video makes clear is that learning to code allows you to do much more interesting, powerful things with LLMs! If you want to maximize your prompting skills it’s worth learning some basic programming :)

    66 likes

  • Mckay Wrigley's profile picture

    @mckaywrigley

    Put the full video on YouTube so that it’s a little easier to watch. I’ll probably do some more cross-post stuff with long video going forward.

    40 likes

  • Chirag's profile picture

    @ChiragCX

    how ppl are still vibe checking prompts is beyond me.. DSPy!

    7 likes

  • Mckay Wrigley's profile picture

    @mckaywrigley

    I’m not a fan of extrapolating away concepts under-the-hood - I’m notoriously against prompt “frameworks” (loosely defined). Practitioners should be as close to the pure prompt as possible. Combine that with good evals!

    15 likes

  • The Canaanite's profile picture

    @mysticaltech

    Man, this is mastery. Super valuable stuff, thanks. Also, Aider and Claude-Engineer are cool, but I don't think they are that cool. What you're building is something else, man; you've basically almost pseudo-solved system-2 for coding. And the UI/UX😘, very sweet. Feels way superior to GitHub Copilot Workspace.

    4 likes

  • Adam Knight's profile picture

    @BigAdamKnight

    I setup a nice little workflow hooking up a chat interface (C3.5S) with tools that run through n8n, I used tool use/function calling with Claude. Would you recommend the XML tag parsing over tool use?

    2 likes

  • Mckay Wrigley's profile picture

    @mckaywrigley

    I’ve found function calling to be worse tasks that require complex reasoning (like codegen on large codebases). I think it works really great on well-defined, more constrained tasks though.

    2 likes

  • Jameson Sharp ג’יימס's profile picture

    @JamesonSharpKc

    LLM’s 💡

    2 likes