Markdown Studio

Markdown Studio sits in an unusual spot: it is a genuinely complete markdown editor that happens to have prompt-engineering tooling bolted on, rather than a thin prompt UI with an editor attached. The architecture is the selling point. Rendering, token counting, exports, and the LLM calls all execute client-side, and the vendor documents using cookieless Umami analytics for feature usage only. There is no authentication server and no user database, which means a breach of the vendor leaks nothing, because there is nothing stored to leak. You bring your own OpenAI, Claude, or Gemini keys and are billed by those providers directly. Where it earns its place in a workflow is iteration speed. Battle Mode runs one prompt across models in parallel and shows outputs, tokens, and cost together — the comparison you would otherwise do by hand across three browser tabs. The token counter tracks GPT-4, Claude, and Gemini context limits as you type, so you catch an over-length prompt before you burn a call. Smart variables with {{variable}} syntax and saved presets turn a one-off prompt into a reusable template, and the 24 templates cover code review, documentation, and bug analysis out of the box. Prompt tag autocomplete, triggered by typing '<', exposes 23 AI prompt tags including thinking, instructions, examples, and Claude 4 control tags, which cuts down on remembering syntax. The editor half is not an afterthought. GitHub sync writes to repos and gists with conflict detection and encrypted tokens, so your prompt library is versioned in the place your other work already lives. Multi-tab editing handles 10 files with independent undo history and auto-save. Live split preview highlights the line you're on. 40+ slash commands insert tables, code blocks, templates, and diagrams. Mermaid covers flowcharts and sequence diagrams; KaTeX handles LaTeX math; presentation mode turns a document into arrow-key-navigable slides with mobile swipe. Exports go to PDF, HTML, JSON, or Markdown. The enterprise story is narrower than it sounds. Deterministic execution (temperature=0, seed-based consistency, exact model version pinning), audit logging with severity levels and SIEM webhook delivery, client-side PII detection for SSNs, cards, emails, and phones, and SHA-256 snapshot certification with drift detection are real governance primitives, and doing them client-side is a legitimate differentiator. But they are governance primitives, not collaboration. There is still no shared workspace, no seats, no permissions model, and no real-time co-editing. A team that needs four people editing the same prompt library with role-based access will not find that here; they'll be looking at PromptLayer or Langfuse and paying for it. The practical caveats: you need your own API keys, and you'll be billed by OpenAI, Anthropic, or Google for every Battle Mode run — the tool is free, the inference is not. It works offline after first load, but the AI features obviously need