Every model ships an official manual for how it wants to be prompted — and the manuals disagree with each other. A prompt written "correctly" for one model is written incorrectly for another. Prompt X-Ray lints your prompt against the target model's own published guidelines and shows you a receipt for every finding. It is a miatz build-lab concept playable: run the lint here to feel why the same prompt scores differently across vendors, then learn to build the rule engine and the payload reveal yourself.
How to use it
Paste a prompt (or keep the deliberately flawed example), pick the model you are about to run it on, and press X-ray it. Findings arrive grouped as blockers, warnings and suggestions — each with a fix hint and a link to the official document that states the rule. Pick a second model in the compare dropdown to see the same prompt judged by two different rulebooks at once, with the axes where the vendors' own docs disagree highlighted. The ◆ reveal at the bottom shows the literal API request each provider would receive.
What this clears up (the fundamentals)
Prompting advice is not universal — it is per-model, and the differences are documented by the vendors themselves. Chain-of-thought phrasing helps some models and is explicitly discouraged for reasoning models. One vendor's docs prefer XML structure; another's prefer markdown headers and report JSON performing poorly for documents. Few-shot examples are strongly recommended by one guide and minimized by another. The score is fully explainable: every point lost traces to a specific rule, and every rule traces to an official source. If the official docs are silent on an axis, no rule exists — absence of guidance is shown honestly, never invented.
Where it's used
This playable is the judgment step of the prompting module: after Prompt Anatomy teaches you what a well-formed prompt contains, Prompt X-Ray grades real prompts against real vendor rules — before you spend tokens finding out the hard way. The parameter advice panel covers the API-side settings (temperature, reasoning effort, thinking toggles) that no amount of prompt text can replace.
FAQ
Where do the rules come from?
From a versioned catalog of official, first-party prompting guides — the vendors' own documentation, fetch-verified with the date recorded. Every finding links its source. If a claim is not in the catalog, the rule does not exist here.
Does my prompt get uploaded anywhere?
No. The lint runs entirely in your browser — no account, no AI call, no network request with your text. The share link encodes your prompt in the URL itself (first 1,200 characters), so only share it deliberately.
Why does the same prompt score differently on two models?
Because the vendors' official guidance genuinely differs — structure idiom, chain-of-thought etiquette, few-shot policy and document placement are all documented differently per model. The compare view highlights exactly where the two rulebooks conflict.
What can't a text lint catch?
Parameter-side guidance (temperature, thinking modes), multi-turn history hygiene, and tool-calling dialects don't show up in a single prompt's text. Those appear as parameter advice and usage notes with sources, not as findings.
Limits
The lint is deterministic and heuristic: it checks what is checkable in prompt text against documented rules, and it can miss context only a human has. A high score means the prompt follows the published guidance, not that the output will be good. Rules carry the catalog's verification date; when a vendor rewrites its guide, the profiles get re-verified and your saved scores may drift — that drift is shown, never hidden.
Related
The prompting track continues with Prompt Lab (the payload autopsy this reveal extends), The Dice Room (what sampling parameters actually do), and Model Arena (race the outputs once the prompts are fair).