MarsDawn
An agent writes it. You still have to understand it.
An AI agent can draft a plan, a spec or a set of notes quickly. What it produces still has to be understood by the person who acts on it — not trusted because it reads fluently.
MarsDawn is built for that read: the rendered page next to the source, with Mermaid diagrams and KaTeX math drawn out instead of left as marks, so a document's structure is legible at a glance.
Fluent isn't the same as correct
Writing about AI-assisted coding, Simon Willison put it this way about code that will be worked on again, not thrown away: “the quality and understandability of the underlying code is crucial” (Vibe coding, 2025). The same holds for a document: an agent's draft that reads smoothly can still get the structure, the numbers or the logic wrong, and fluent prose doesn't announce which parts to check.
Inference, not compilation
Birgitta Böckeler, writing for Thoughtworks, draws the distinction plainly: “LLMs are NOT compilers, interpreters, transpilers or assemblers of natural language, they are inferrers” (I still care about the code). A compiler either accepts your input or reports an error; an agent can hand back something that runs, or reads, without being right. Somebody still has to check it.
What MarsDawn gives that reader
- The rendered page next to the source, updating as either side changes, so a claim in the text and the way it's structured are both in view at once.
- Mermaid diagrams drawn out: a flowchart an agent described in text becomes a shape you can actually follow.
- KaTeX math rendered, not left as a string of backslashes: a formula reads as a formula.
- Nothing runs on its own. MarsDawn doesn't grade, summarize or flag the document for you — it puts the structure in front of you so you can.
Next
- How that review stays cheap for the agent's own context: token-efficient review.
- Handing the reviewed document to someone else: sharing a PDF.
- What MarsDawn is, in one page: the home page.