tommy0103/prism

Prism is a tool for turning structured documents written by an AI agent into a single, good looking HTML page, without the agent needing to write any custom design or styling. Instead of writing plain markdown, an agent writes using a set of purpose built tags, called components, for things like decision records, code references, step by step flows, comparisons, and key metrics. Prism then builds all of that into one self contained HTML file with a consistent visual style baked in. The project ships 45 ready made components across eight families, covering decisions, reasoning, structure, source code display, data, process steps, containers, and small inline elements like badges and tags. Each component encodes not just a look but a writing convention: a conclusion belongs in a decision card, a piece of code belongs in a dedicated source block, and a comparison belongs in its own comparison component, rather than being described loosely in a paragraph. The idea is that this structure pushes an agent toward writing documents that are easier to scan, not just easier to read. To use it, an agent writes an HTML template using Prism's custom tags, then a small build script turns that template into a single finished HTML file with no server and no external dependencies needed to view it. Every generated document embeds a hash of the runtime code along with a security policy that blocks the page from making any network requests, and a separate verify command can confirm a given file was built with an unmodified copy of Prism. Prism installs through a skills command line tool and works with Claude Code, Codex, and other AI coding assistants that support a plugin or skill system. If you would rather write your own custom styling on top of a page, that is still possible since the components use plain, overridable class names. It is released under the MIT license.