2026

Schema-First Prompting

Claude / Cursor skill — Pydantic as the contract, the prompt for intent only.

LLM Prompt Engineering Open Source

Schema first, prompt second. The Pydantic model carries structure— names, types, nesting, and limits. The prompt carries what JSON cannot: tone, audience, and rhetorical rules. The skill documents how to keep those responsibilities separate so validators and the model never disagree.

What you get

Five principles (no duplication, reasoning order, don’t ask for known values, design for LLM behavior, start tight), then reference material: schema shape (unions, wrappers, bases), field design, provider strict modes and sanitizers, validation retry loops, and production habits—versioning, golden sets, logging validation failures.

Why it exists

Structured-output pipelines already treat the model as part of the prompt: field descriptions and JSON Schema shape what the model attends to. This skill makes that explicit—so teams stop restating max_length in prose, stop putting decision before scratchpad fields, and split “LLM-only” models from API or persistence types.

Install

npx skills add AgaMiko/schema-first-prompting

Works with Claude Code, Cursor, Copilot, Codex, Gemini CLI, and other agents via the Vercel skills ecosystem.

GitHub → Installation →

← All projects