Where DSPy shows up in the material DSPy appears in the channel's captured material as a prompt-optimization library — a tool you point at an already-specified task to improve accuracy automatically, rather than hand-tuning the prompt yourself. It comes up twice in Michael Taylor's conversation about evals and prompt engineering: [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1042] [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1740] - As an optional accelerator for classifiers: after building a rubric of binary 0/1 evaluation criteria from a comparison of good vs. bad outputs, Taylor notes you can optionally push further with a prompt-optimization library like DSPy to improve classifier accuracy — which then improves the generation task too. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1042] - As a marker for well-specified tasks: for a well-specified task, you may not need prompt engineering at all — use DSPy. Taylor invokes DSPy precisely to contrast it with the unspecified world — you don't know what a good article is until you see it — where prompting skill still matters. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1740] [moment: 1740] How to use this distinction In the channel's framing, DSPy sits on one side of a divide: - Specifiable, verifiable tasks → DSPy-style optimization can find the prompt for you; hand prompt engineering adds little. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1740] - High-dimensional, inexplicit objectives (writing quality, taste, what makes a good article) → no optimizer substitutes for detailing your objective, which is what prompting is. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1638] The material is careful about scope: DSPy is mentioned only as a tool for improving classifier accuracy inside a hand-built eval workflow (Jupyter notebook, comparison-derived rubric, ~20 binary judges scored against known-good examples) — not as a framework that replaces that workflow. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1042] - DSPy is a prompt-optimization library — it improves prompts (and the classifiers built from them) automatically for well-specified tasks. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1042] - It is presented as optional: a way to push classifier accuracy further after you have already built a rubric of binary judges. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1042] - Its other role in the material is contrastive: if the task is well-specified, use DSPy instead of prompt engineering — but objectives in the real world are often too high-dimensional to specify, which is where prompting remains the skill. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@1740]