Show HN: ZON-TS 50–65% fewer LLM tokens zero parse overhead better than TOON/CSV https://ift.tt/gPYq8l9

Show HN: ZON-TS 50–65% fewer LLM tokens zero parse overhead better than TOON/CSV hey HN — roni here, full-stack dev out of india (ex-gsoc @ internet archive). spent last weekend hacking ZON-TS because json was torching half my openai/claude budget on dumb redundant keys — hit that wall hard while prototyping agent chains. result: tiny TS lib (<2kb, 100% tests) that zips payloads ~50% smaller (692 tokens vs 1300 on gpt-5-nano benches) — fully human-readable, lossless, no parse tax. drop-in for openai sdk, langchain, claude, llama.cpp, zod validation, streaming... just added a full langchain chain example to the readme (encode prompt → llm call → decode+validate, saves real $$ on subagent loops). quick try: ```ts npm i zon-format import { encode, decode } from 'zon-format'; const zon = encode({foo: 'bar'}); console.log(decode(zon)); ``` github → https://github.com/ZON-Format/ZON-TS benches + site → https://zonformat.org YC’s fall rfs nailed it — writing effective agent prompts is brutal when every token adds up. if you’re in a batch grinding observability (helicone/lemma vibes) or hitting gemini limits like nessie did, what’s your biggest prompt bloat headache right now? paste a sample below and i’ll zon it live. feedback (harsh ok) very welcome cheaper tokens ftw https://zonformat.org December 9, 2025 at 10:21PM