Context Engineering With DSPy

Every · · 1 hr 9 min

Mike Taylor, head of evals at Every, walks through DSPy with Drew and Isaac of Compound, the team behind it. They argue you should define tasks and examples, then let optimizers write the prompts instead of hand-tuning them.

Dan Shipper co-host MT Mike Taylor guest · 3 picks
Watch the episode 1 hr 9 min · YouTube

The gist

  1. Mike says DSPy signatures define a task by its typed inputs and outputs, so his two-year-old programs still work after swapping models and re-optimizing.
    DSPy
  2. Using a Pydantic object as the output type removed all his parsing and retry code when Mike generated 5,000 AI personas for a credit card company.
    Pydantic

What they recommend · 3 picks

DSPy Main topic · strong yes

A framework for algorithmically optimizing language model prompts and weights, especially for complex tasks...

"The magic words that you found, like they don't resonate with this new model. And so DSPy I think solved that problem for me, which is I can do all the fun stuff around the prompt, but I don't need to write the prompt anymore... My DSPy programs I've writte..."
Mike Taylor ▶ Play at 4:29

What Mike uses it for: Model-agnostic LLM programs, building CLIs, optimizing prompts, and getting structured outputs

Get an email when another creator talks about DSPy.

33:59
Optimizer-evaluator pattern AI tools

"One of the tricks I like to do in DSPy... it's called an optimizer evaluator pattern. If I have a subjective test like, 'Hey, does this rewritten text look like AI?' I will firs..."

Mike Taylor ·Mike's AI tools →
24:05
Pydantic Productivity

"I just made that output object in Pydantic and then passed that in... I'd written all this extra code of like, hey, if it doesn't come out with this, then try again. And none of..."

Mike Taylor ·Mike's productivity →

Also mentioned

26:40 Claude · "Now I don't hand in an HTML page or an email of like here's the work I did. I use Opus to actually create a..."
30:45 pandas · "I'm just going to load this into a data frame in Pandas."
38:12 ChatGPT · "This text is not AI, I wrote this. You know, this one is AI, like ChatGPT wrote this."
54:25 Gemini · "Then we can drop in Gemini. We can say, 'Hey, a new Gemini model came out.'... And then you can see here..."
1:00:22 LangChain · "I had kind of built DSPy in LangChain at the time. I had a really shitty one, and then I was like, 'Oh,..."
◎

Related episodes

Episodes covering similar topics

Writing automation workshop with Mike Taylor at Every Think Week
Every · Nov 04
prompt engineering
$100M ARR Founder: The Next Big AI Opportunity
Greg Isenberg · Oct 28
prompt engineering
The New Full-Stack Designer
Y Combinator · Oct 26
developer tools
I Tried Learning AI Video Editing in One Day
Ali Abdaal · Oct 26
prompt engineering
Every More from Every
The Inside Story of the 50 Who Built Modern AI With Kevin Roose
Nov 04, 2026
Writing automation workshop with Mike Taylor at Every Think Week
Nov 04, 2026
Every Think Week: Monologue HARDWARE?!
Nov 03, 2026
Human in the loop - 4 levels of automation
Oct 29, 2026
Can You Use AI To Make Good Art? — With Fred Graver
Oct 29, 2026
How a Frontier Team Can Help Your Company Keep Up With AI
Oct 21, 2026

Follow DSPy

Free email updates