More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    DSPy: Program Language Models, Don't Prompt Them

    1

    Build LM systems declaratively with DSPy — signatures, modules, and optimizers that find good prompts from data instead of hand-tuning.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    DSPy: Program Language Models, Don't Prompt Them

    DSPy: Program Language Models, Don't Prompt Them

    Example session with this skill installed

    Build a small question-answering pipeline with DSPy: answer questions from a short context passage, with a metric and a held-out evaluation.

    • Read your context and instructions
    • Compiled the dspy

    QA pipeline built and measured. Signature: question, context -> answer, instruction 'answer concisely using only the context'. Module: ChainOfThought, chosen because the questions need multi-hop reasoning. Metric: exact match after normalization, hand-checked on five examples before optimizing. Optimization: 40 training examples, few-shot optimizer; training score 0.55 to 0.78. Evaluation: held-out set of 20 questions scored 0.70, up from a 0.50 unoptimized baseline. The gain is real and modest; the record names retrieval quality as the next bottleneck, not wording. Program, metric, optimizer, and both scores saved together. The metric was hand-checked on five examples before any optimization ran, so the optimizer was chasing the real goal rather than a proxy. The training set of 40 examples and the held-out set of 20 were kept strictly separate, and the held-out score is reported honestly: 0.70 is a real gain over the 0.50 baseline, not a training artifact. The bottleneck is named in the record — retrieval quality — so the next iteration adds a retrieval step instead of burning more optimization rounds on wording that is already good enough. The working log holds the signature, the metric with its five hand-checked examples, the optimizer settings, and both scores, so the next model swap starts from a record instead of a memory. Rubric score: 8/8.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Hand-written prompts are the assembly language of LM systems: brittle, unportable, and silently degrading every time the model changes. DSPy (Stanford NLP) replaces prompt-craft with programming — you declare each step's inputs and outputs as a signature, compose steps into modules, and let optimizers find good prompts from data. You engineer the structure; the data engineers the wording. This skill covers the full loop: configuring the LM, declaring signatures, composing Predict, ChainOfThought, and ReAct modules, defining the metric that decides success, optimizing on a small training set, and evaluating on held-out data before anything ships. It also covers when DSPy earns its keep and when it does not — a single prompt does not need a framework, and a pipeline without evaluation data cannot be optimized. A worked example builds a small question-answering pipeline from signature to optimized program, with the metric and the held-out score recorded.

    DSPy needs evaluation data — without a metric and a held-out set, there is nothing to optimize against.

    • Optimized prompts are tied to the LM they were optimized for; a model swap needs re-optimization.
    • A single one-off prompt does not need this framework; a plain API call is faster.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    No — it is the opposite. You stop writing prompts and start writing structure and metrics; the optimizer finds the wording. The value is that the structure survives a model change and the quality is measured, not vibes.

    Tens of examples is usually enough to start — 40 worked in the demo. More helps, but a good metric on 40 examples beats a sloppy metric on 400.

    For a single prompt, for work with no evaluation data, or where a plain API call already does the job. The framework earns its keep on pipelines, not on one-offs.

    New in LLM Ops