AI bench / builders first
AI and LLM tools.
For the people building with language models and the people whose sites those models read: shape prompts and tool schemas, repair the JSON a model hands back, see how a document splits for retrieval, and decide which AI crawlers may read your pages. The last group runs small models inside your browser, so the photo or scan never leaves your machine.
ready now
Build with LLMs
Prompts, tool-calling schemas, and the structured output a model sends back.
System Prompt Builder
Build structured LLM system prompts with XML sections and formatted OpenAI/Claude tool-calling JSON schemas.
JSON Repair
Repair truncated, malformed, single-quoted, or unescaped JSON generated by LLMs without server round-trips.
JS Object to JSON
Convert unquoted JS object literals, single quotes, and trailing commas into strict JSON payloads.
JSON Schema Generator
Automatically infer and generate JSON Schema definitions from sample JSON payloads.
JSON Schema
Validate JSON data against JSON Schema declarations locally in your browser with clear error paths.
RAG Text Chunker
Visualize text splitting strategies, token/character chunk sizes, and overlaps for RAG vector indexing in your browser.
Make a site readable by AI
What AI crawlers may fetch, and what they find when they do.
AI Crawler Checker
Understand how a site's robots rules address known AI crawlers.
llms.txt
Inspect a site’s optional llms.txt structure, links, and source text.
Robots.txt Checker
Read crawler rules, sitemap declarations, and accessibility at a glance.
Schema Checker
Find JSON-LD blocks, schema types, parse errors, and basic warnings.
AI on your own device
Models that run in the browser tab, so the file is never uploaded.
Image to Text
Pull the text out of a photo or screenshot without uploading it.
PDF OCR
Read text out of a scanned PDF that has no selectable text.
Passport Photo
Make a passport or ID photo with the right size, a white, blue, or red background, automatic framing, and a 300 dpi print sheet.
Getting output you can parse
Most failures in an LLM feature are not wrong answers but unusable ones: a reply wrapped in prose, a trailing comma, a single-quoted key, an object cut off when the token limit ran out. Describe the shape you want as a JSON Schema — the JSON to JSON Schema generator writes one from a sample — and pass it to the model's structured-output or tool-calling feature rather than asking for JSON in prose.
When a reply still arrives broken, the LLM JSON repair tool fixes the mechanical faults, the JS object to JSON converter handles JavaScript-style literals, and the JSON Schema validator confirms the result matches before your code relies on it. The system prompt builder keeps the instructions and tool definitions around all of this in one consistent format.
Chunking for retrieval
A retrieval system answers from the chunks it finds, so the split decides what a model can see. Chunks that are too small lose the sentence that gives them meaning; chunks that are too large bury the relevant line among unrelated ones and cost more to embed and send. Overlap keeps a sentence that crosses a boundary findable from both sides, at the price of storing it twice.
The RAG text chunker shows exactly where a document breaks under a given size and overlap, which is quicker than debugging a poor answer back to the split that caused it.
Deciding what AI crawlers read
Training crawlers, AI search crawlers, and fetches made on a user's behalf use different User-agent tokens, so a site can refuse training while staying visible in AI answers. The AI crawler checker reads a site's robots.txt against ten of them and explains each; the robots.txt checker shows the whole file.
For the pages you do want read, an llms.txt file points models to your most useful documents, and structured data states plainly what a page is about. Neither is a guarantee, but both cost little.
Models that stay in the browser
The OCR and photo tools run their models inside the page. The first use downloads the model — a few megabytes — and after that the work happens on your own processor: a scanned contract, a passport photo, or a screenshot is read and processed without being uploaded anywhere.
The trade-off is speed on older devices and the limits of a small model. Image to text and scanned PDF to text handle printed text well and handwriting poorly; the passport photo maker finds the face and background automatically but still asks you to check the result against your country's rules.