Looking to hire me? Check out my introduction repository for a curated list!

@romandmitri/ai-patch

https://www.npmjs.com/package/@romandmitri/ai-patch

Problem

AI generation loops often rewrite an entire prompt or document for a small change. That consumes more tokens, costs more, takes longer, and increases the risk of hallucinations, context drift, and unintended rewrites.

Solution

@romandmitri/ai-patch bridges the gap between AI generation and diff-based updates. It lets an LLM return small, structured patches, then validates and applies them atomically while leaving unrelated content unchanged.

Install

npm install @romandmitri/ai-patch

Text is the currently supported content format. JSON coming... if needed.

Vercel AI SDK

Structured Output

import { Target } from "@romandmitri/ai-patch";
import { Output, generateText } from "ai";
import { z } from "zod";

const documentPatchlet = Patchlet.from("# Guide\n\nOld instructions.");
const summaryPatchlet = Patchlet.from("# Summary\n\nOld summary.\n");

const { output } = await generateText({
	model,
	instructions: "Update the document to something else.",
	messages: [
		{ role: "user", content: "The [document]:\n" + documentPatchlet.content },
		{ role: "user", content: "The [summary]:\n" + summaryPatchlet.content },
	],
	output: Output.object({
		schema: z.object({
			documentPatches: documentPatchlet.toSchema(),
			summaryPatches: summaryPatchlet.toSchema(),
		}),
	}),
});

const updatedDocument = documentPatchlet.patch(output.documentPatches);
const updatedSummary = summaryPatchlet.patch(output.summaryPatches);