marze
01/04

Most prompts
fail quietly.

They don't throw. They return something plausible, you accept it, and the cost never shows up on a line item. Marzel reads the prompt before you send it, names each failure against the span that caused it, and hands back the corrected version.

02/04

Every finding is anchored to a span.

No scores, no vibes, no "consider adding more detail." A severity, the exact words that triggered it, and what the model would have done with them. Below is one run, unedited, translated from the original.

Specimen 01 · verbatim run
Input13 words

how do i use axios fetchAll() to get all the pages at once

The question reads as answerable. Sent as written, the model picks up the premise and writes you a working-looking wrapper around a method that has never existed.

Analysis1 finding · 1 high
False premise

Axios has no native fetchAll(). Fetching every page means implementing pagination yourself, either with repeated calls or concurrent requests through Promise.all.

REWRITE one finding at high severity
03/04

It fills what it can guess. It refuses to guess what it can't.

Every gap in a prompt is one of two kinds, and the two get opposite treatment. Getting that split wrong is what makes most rewriting tools worse than the original.

Gap type A

Presumable context

Language, framework, database, error handling, output shape. A common default answers it about as well as you would have.

Filled in, and every default it chose goes on the record so you can swap it.

Gap type B

Irreplaceable fact

Your actual code, the full traceback, the real schema, the query itself. No default exists. Only you have it.

Left as an empty bracketed slot. Never invented, not even plausibly.

Rule of thumb

If inventing it would send the answer the wrong way, it stays a slot. If a common default handles it well, it gets filled.

Before either

If the prompt is built on a method, an API or a behaviour that doesn't exist, that gets caught first and reported on its own. No rewrite can rescue a question with a wrong assumption inside it, and a model asked politely enough will answer it anyway.

04/04

You already pay for the frontier. The prompt decides what you get for it.

A working developer running an assistant, an agent and an editor is past US$100 a month before anything ships. That number is fixed and it isn't the problem. The problem is that the one input you fully control is the one nobody audits.

And the real cost isn't the extra round trip. A wrong first instruction sends the agent to the wrong files, runs the wrong searches and writes the wrong code, and all of it stays in the window for the rest of the session. Tokens are renewable. Context isn't. Every correction after that competes with the wrong output still sitting behind it, and once the window fills, compaction summarises your good instructions away alongside the garbage that displaced them.

Marzel doesn't make the model smarter. It makes sure the spend you've already committed is pointed at a prompt that can be answered. The audit is the work you would have done yourself, if you were ever going to.

Analysis enginefrontier-class, single pass
What's keptyour prompts and their analyses, for your history
Who can read ityou, from your own account
Deletionone at a time or all at once, immediate
Trainingnever, on anything you send
Repository accessnone, ever
End

Showing this to a small number of people first.

If you spend real money on models every month and you've had the feeling that a prompt came back fine but not right, you're who this is being built for. Leave an address and I'll send you the first working build.