Agent skill

Slaide ships an agent skill for Claude Code, Codex, and Gemini. With it installed, you describe a deck in plain words and the agent builds it end to end. It writes the .slaide source, checks its own work by looking at the slides, and exports the result.

What makes it autonomous

The skill does two things that let the model finish without hand-holding.

• It gives the model the grammar and theme vocabulary up front. So the model uses real layout, slot, color, and gradient names instead of guessing.
• It lets the model see what it rendered. The model does not just trust that valid text looks right. It screenshots the deck and reads the image back.

The loop

1. You ask for a deck. For example: “a 6-slide overview of our Q3 roadmap.”
2. The agent writes the .slaide, routing Markdown into the theme’s slots.
3. It runs slaide validate --strict and fixes every issue.
4. It runs slaide shoot --montage to render all slides into one small image. Then it reads that image and catches clipped text, weak contrast, and overflow that validation cannot see. It edits and re-shoots until it is clean.
5. It exports the deck. slaide build for web, slaide export --pdf and --pptx.

Steps 2 to 4 repeat on their own until the deck holds together. You review a small text diff. Then take over by hand in the editor whenever you want. Same file, no drift.

Install

Run one command:

slaide install

It detects the AI coding CLIs you have, like Claude Code, Codex, and Gemini, and installs the skill plus the MCP server.

Useful options:

slaide install --list            # show which CLIs were detected
slaide install --scope project   # install into this project, not globally
slaide install --no-mcp          # install the skill only, skip the MCP server

Building a custom agent instead? Use the MCP server. Same engine, same results.