A prompt can contain a task, a little world, several rules, and a contradiction hiding between two sentences. Paste one below and see what structure you can make visible.

Published 25 October 2026 · Updated 25 October 2026

Start with the words

The character counter updates as you type. It includes spaces and line breaks. A visible character can be made from several Unicode code points, so the expanded statistics explain the difference between characters, code points, UTF-16 units and bytes.

Choose Analyze when you want to inspect the prompt. Local mode uses rules in your browser. Optional AI · semantic mode uses a free model to interpret unusual language and build a more nuanced ontology. Review the named provider and its retention policy, then consent to that request. Free AI availability is limited; you can always return to local analysis.

Long prompts use source-grounded compression: repeated passages are deduplicated and exact excerpts are selected from across the original. The full prompt stays available for evidence. The AI proposes entities and relationships; the app anchors them to original source passages and preserves clear instructions and explicit storage details found locally. Coverage figures show how much was analysed and omitted, so a partial interpretation is visible. The editor accepts up to 200,000 Unicode code points; a diagram is a working overview, not a claim that every detail survived compression.

A map of what the prompt says

An ontology here means a working inventory of the people, things, concepts, actions, rules and relationships found in your text.

There are two layers to inspect. In “Write a kind bedtime story about a robot gardener”, writing a story is the task, kindness is a constraint, and the robot gardener belongs to the subject. Keeping those layers visible makes it easier to see what you are asking for and what you are asking about.

Select an item to inspect its evidence. An explicit item points back to words in the source. An inference records an interpretation. A correction you confirm is marked as information supplied by you. The original text remains available for comparison.

Choose a view that fits

The mind map groups the prompt into branches. The concept map keeps labelled connections between ideas. The knowledge graph connects instruction and subject entities across the prompt, with labelled relationships and traceable evidence. Inferred connections remain visibly distinct from explicit statements. A flowchart needs actions and some supported account of what happens next. A state diagram needs states and transitions. Data models need entities and relationships, then progressively more detail about attributes, identifiers and storage.

Conceptual, logical and physical describe levels of modelling detail. UML supplies a family of diagram notations. A story prompt may support a useful mind map while supplying too little information for a physical database model.

Prompt Atlas checks those prerequisites before recommending a view. A fit score describes how well a view matches the available structure. It is not a probability that an interpretation is correct.

Changing the selected view reuses the same interpretation. Correct an entity or relationship once, then inspect the consequences in the supported diagrams.

Leave room for uncertainty

“Make it exactly 100 words and no more than 50 words” contains two incompatible instructions. A useful inspection makes both visible.

“The moon invoices Tuesday” may be a fine beginning for something strange. Its diagram should preserve that strangeness. A drawing becomes less useful if it quietly supplies a billing system, customer records and a database that the sentence never specified.

Read unresolved questions as invitations to inspect the prompt. Confirm a likely interpretation, supply a missing detail, or keep the uncertainty.

Take the result with you

Export a diagram as SVG, save the structured interpretation as JSON, keep a readable Markdown outline, or export the ontology as a JSON-LD knowledge graph. Analysis mode, provider identity and compression coverage travel with the result. Check the source-text option before exporting: a prompt can contain information you would prefer to keep in the current session.

Prompt Atlas appears in both Viz and Apps. Both collections open this same tool.