This is what the JSO AI preset assistant will feel like. Type a description of your app, click Generate, and get a jso.config.json. Today the suggestions come from a deterministic rule-based mapping — matching keywords like "react", "node", "license", "balanced", "maximum". From 2026-Q3, the same UX is backed by a Claude-based assistant that reads the description with context and produces a much smarter config. Same shape, smarter backend.
Preview only. The mappings below are hand-written rules so you can feel the UX before the LLM-backed version ships. Real responses will be more nuanced — the LLM reads your description, asks clarifying questions when ambiguous, and surfaces trade-offs the rule-based version flattens.
Join the waitlist and you'll get an email when the real assistant goes live.
Describe your app
Generated jso.config.json
Detected signals
- (type a description and click Generate)
How this changes when JSO AI Basic ships
The rule-based mapping today catches obvious keywords and produces a reasonable starting config, but it does flatten nuance. Three concrete examples of what the LLM-backed version will do differently:
- "Protect the license-check function strongly." Rule-based version sets
FlatTransform: true globally. LLM version asks for the function name (or guesses it from the surrounding description) and recommends // @virtualize on just that function, leaving the rest of the bundle on balanced — smaller output, faster runtime.
- "Performance matters most." Rule-based version drops to a
standard preset. LLM version reads the rest of the description (is it a marketing widget or a payment flow?) and chooses between standard and balanced accordingly, naming the trade-off out loud.
- Ambiguous descriptions — "obfuscate my JavaScript." Rule-based version defaults to balanced. LLM version asks what frameworks you use, how performance-sensitive your hot paths are, and what code is high-value, then produces a config informed by the answers.
The promise of the AI version is honest reasoning about trade-offs, not just keyword matching. The rule-based preview is a UX prototype; the real feature is the conversation.