Why this
As agents run longer, they dump more text. A fifty-line plan is fine. A thousand-line markdown spec is something you skim, then hand back to the model for another rewrite — without really inspecting it. The bottleneck is no longer whether the model can write. It is whether you stay in the loop.
Thariq Shihipar (Claude Code) wrote this down as the unreasonable effectiveness of HTML. Instead of a wall of markdown, ask for a self-contained page: a side-by-side of three approaches, an annotated diff, a clickable flow, a concept you can poke. You review by looking, not by scrolling a terminal.
Thariq's reason is staying in the loop. Markdown plans long enough to skim are plans you stop reading. HTML is how he stayed engaged with the agent's choices instead of handing them off.
His example gallery is the argument in working files — exploration, review, design, prototypes, decks, explainers, throwaway editors. You do not need a skill that wraps HTML in a template. You need to know you want a page, and then actually open it.
The page is still a local file. It has no URL, no name, no way for the next agent (or a colleague) to find it. Hosting a site for every explainer is the thing you were trying not to do. Chat artifacts disappear. Gists and drive links are someone else's product, with someone else's login.
Pustak is a store for those pages. One HTML file in, a URL out, served from the edge under your name. Public, unlisted, or private. An agent can write the next one over MCP — or you can upload from the dashboard if MCP is not installed yet.
That is the whole service. Thariq showed the format. Pustak is a shelf for the files.