decks: Claude Code and Codex edit slides based on comments beneath them

⚡ The gist in 5 seconds - A project called decks (repo paulmil11/decks) has been published on GitHub — presentations where edits are made by a coding agent in response to a comment under a slide. - Two agents are claimed to be supported: Claude Code and Codex; the agent's response appears right on the slide itself. - The description is limited to a single paragraph: it contains no system requirements, installation commands, or implementation details. ### 🔍 What we found A project called decks (repo paulmil11/decks) has appeared on GitHub, which the author describes as "agent-first slide decks" — presentations built around an AI agent rather than around an editor. The workflow is described as follows: you leave a comment under a specific slide, and a coding agent (Claude Code or Codex) makes the change and posts its response right there, on the slide. Judging by the description, the mechanics flip the usual order: the entry point is not the agent's chat or a separate window, but an anchor to a specific spot in the deck. The comment serves as both the task and the context, and the result of the edit is returned to the same slide — meaning the "request → change → report" loop is closed within the document itself. It's unclear from the description whether the agent runs locally, connects to a repository via CLI, or uses a backend — there are no public details about the implementation, dependencies, or license in the description. ### 💡 Why it matters Coding agents like Claude Code and Codex are already firmly embedded in the terminal and IDE, but outside of code their role usually boils down to generating text "in a chat." This project attempts to apply the same pattern to presentations: instead of manually rewriting a slide's wording or explaining the context to the agent in words, the edit is addressed to a specific slide via a comment. The practical value here lies precisely in the location anchoring — it eliminates the main pain point of working with agents on documents, where you have to re-explain which fragment you're even talking about. This is an early open-source experiment without proper documentation, so it should be evaluated as a proof of concept for an approach rather than a production-ready tool.