llm-json-lite

A tiny utility to convert JavaScript objects into a relaxed JSON format optimized for LLM (Large Language Model) prompts.

Instead of strict JSON with quotes and colons, llm-json-lite outputs a compact format using = for assignment and minimal noise — perfect for reducing token usage in LLM inputs.

💡 Save 8–16% tokens on average compared to standard JSON — meaning cheaper, faster, and more efficient LLM calls.

Features

• Compact format – saves tokens when sending data to LLMs.
• 8–16% fewer tokens – lower cost and faster processing.
• Human-readable – no quotes, no colons.
• Smart filtering – removes null, undefined, and empty arrays.
• Date handling – serializes Date objects to ISO YYYY-MM-DD.

Installation

pnpm add llm-json-lite
# or
npm install llm-json-lite
# or
yarn add llm-json-lite

Usage

import { optimize } from "llm-json-lite"

const data = {
  name: "Alice",
  age: 25,
  active: true,
  tags: ["dev", "ai"],
  birthday: new Date("2000-01-01"),
  empty: [],
  optional: null
}

console.log(optimize(data))

Output

You are given RELAXED-JSON where keys use = instead of : for assignment and items are comma-separated.
Rules:
- Objects: {k=v,k=v,...}
- Arrays: [v,v,...]
- Strings are unquoted and contain no spaces. Booleans are true/false. Dates are ISO (YYYY-MM-DD).
- Null/undefined/empty arrays are omitted.

DATA_START
{name=Alice,age=25,active=true,tags=[dev,ai],birthday=2000-01-01}
DATA_END

API

optimize(obj: Record<string, any>): string

Converts a JS object into a string that includes the instructions and the relaxed JSON data.

Use Cases

Preparing data for LLM prompts (ChatGPT, Claude, etc.)

Token-efficient serialization

Easy-to-read logs and debugging

Contributing

PRs are welcome! If you have suggestions or improvements, open an issue or submit a PR.

License

MIT © Pavel Surinin