Decks lets you edit a slide deck by writing comments on it

Decks is a new MIT-licensed slide editor built around a different interaction: a reader leaves a comment under a slide, a coding agent reads the comment, edits the deck and replies on the same slide. The repository went up on GitHub last weekend and reached 85 stars inside its first day. Each deck is a single HTML file with its own version history, exportable to PDF.

How a slide gets changed

A session has three roles. A reader looks at a slide and types a comment underneath it. The author presses "Ask Claude" to copy the comment and its context into Claude Code or Codex. The agent reads the deck, makes the change, writes a reply on the slide and saves a new version. The next reader sees both the edit and the explanation of what moved.

A single deck is one .html file, held in ~/slide-decks/ on the author's machine. Nothing streams through a hosted service by default, and no account is required. The server is Python 3.9+ using the standard library only, and Node 18+ is used for screenshots and PDFs through headless Chrome. The project works on macOS and Linux.

The checks it ships with

The project carries a few editorial checks of its own. A five-second test asks whether a reader gets the point of a slide at a glance. A ten-second test asks whether a reader can take the slide in during the time they will give it. Action titles are sentences that state the point of the slide instead of labelling it, with the example "Q3 revenue grew 22 percent, driven by enterprise expansion". Flow checks look at whether every slide carries the deck forward.

These are the rules the model is told to apply when revising a slide. The reader is not driving the software, the reader is checking the output and sending the next instruction back as a comment.

Who it is for

The audience the project describes is consultants, founders and anyone who writes a lot of decks on their own and wants a specific style enforced across them. The author positions it as different from editing a slide in a traditional presentation program: the deck is a text file an agent can read and rewrite in full, so a comment that says "action titles only, cut the two bullet slides, add a chart for Q3" is a change the agent can make across the whole document.

The export story is simple. A single PDF or a "client copy" HTML that strips notes and comments falls out of a command. Dual-licensing covers both halves of the project, with MIT for the code and CC BY 4.0 for the template content.

One machine, one deck, one HTML file

Decks is one day old on GitHub at the time of writing, with 85 stars and 7 forks, so this is early adoption rather than a shipping product with a user base to call. The server itself is small, which is both the appeal and the limit: there is no hosted service to buy, no team collaboration layer, and no versioned shared source. A consultant working on their own machine gets the whole flow. A team sharing a deck has to share the HTML file and run their own copies.

For anyone already running Claude Code or Codex every day on code, using the same tool on a slide deck is a small change, and the output format carries that same tool's rhythm: a reader comments, the agent writes.

Source

The code is at paulmil11/decks on GitHub.