CharlieKey takeaways from your favorite podcasts
← 30 October 2026AI & I

Context Engineering With DSPy

DSPy lets you define tasks as typed signatures, then optimize prompts automatically against evals, so programs survive model changes and you stop hand-writing brittle prompts.

Watch on YouTube ↗

Charlie's summary

  1. Describing your own style fails twice: models ignore rules, and you can't see your own nuances (the “high school band problem”)

    Mike says writing rules for a style doesn't work because a rule doesn't mean the model follows it. Isaac adds a second failure mode: you will miss nuances you aren't aware of, leading to a game of telephone. Mike compares it to a high school band describing its sound while a friend says they sound like Modest Mouse. Examples and optimization beat self-description.

    ▶ 0:28
  2. The task stays the same while the implementation changes: signatures outlast every model swap

    Mike left prompt engineering because hand-written prompts break when a model changes, such as GPT 3.5 to 4, and the magic words stop working. Drew says DSPy's interface has barely changed since before ChatGPT, and his programs from two years ago still run. He just swaps the model and reoptimizes. Signatures describe the task, which doesn't care about the implementation underneath.

    ▶ 3:29
  3. Come for the optimizers, stay for the signatures: typed inputs and outputs remove parsing code

    A signature is a class with a docstring, typed input fields and output fields. Mike now thinks of LLM tasks this way. Because outputs are structured, he needed no regex, retry logic or parsing. He generated 5,000 credit-card survey personas against a messy 15-option Pydantic object, and Drew's team passes a giant Overture schema as the output type. Isaac's rule of thumb: if you are prompt engineering heavily inside the docstring, that's a red flag.

    ▶ 11:11
  4. Treat your data like self-driving footage: your job is collecting examples, not writing rules

    Mike says DSPy forces you to do your homework: gather good examples, whether handwritten, synthetic or from domain experts. Drew notes this isolates subject-matter experts' examples from your control flow, avoiding code conflicts. Mike compares it to Karpathy's self-driving shift from hand-written rules to feeding in driving footage. Always look at your data. You no longer need to read the code, but you do need to look at the data, and vibe-coded viewers make that fun.

    ▶ 18:17

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.