Agent-native PowerPoint work usually gets awkward fast. Decks are one of those file types that look simple until you touch them, and then suddenly you’re dealing with zip files, XML, layout quirks, and the kind of tiny formatting breaks that waste an afternoon.

EveryInc’s hands-on-deck brings a more controlled way to work with .pptx files

You can explore the project here: https://github.com/EveryInc/hands-on-deck

What it does

hands-on-deck, packaged as an Agent Skill, gives AI agents a CLI called deck.py for inspecting, editing, creating, and verifying PowerPoint files through atomic JSON patches. That choice matters. Instead of asking an agent to hand-write OOXML, which is fragile and easy to break, it lets the agent describe changes declaratively, then lets the tool apply them safely.

That’s a cleaner operating model. Text keeps its formatting. Image swaps preserve aspect ratio. Duplicated elements inherit style. If a patch is invalid, nothing gets written. No half-finished deck sitting around like a paper cut on the desk.

Built around failure modes, not hype

What stands out is how the project is shaped around real mistakes agents make. Errors include the actual slide inventory, so an agent can correct itself without guessing. Every operation is pre-validated, and if a nine-step patch has nine problems, you get nine errors back at once.

The linter is just as practical. After each apply, it checks for issues like overflow, off-slide objects, overlaps, and near-miss alignment, then reports exact inch values and suggested fixes. It doesn’t pretend to know design intent. It just spots where the deck is drifting away from what was probably meant. That’s a useful distinction.

Why this feels different

There’s also a broader lesson here. Agents tend to do better with tools than with prompts, especially when the tool gives them a tight feedback loop. hands-on-deck leans into that. It even supports HTML-to-patch workflows, so free-form slide design can be measured and compiled into a deck instead of rebuilt from scratch.

The repo also ships with design guidance, benchmarks, and end-to-end tests, which is the part I trust most. Real systems need guardrails, not vibes.

For teams making decks all the time, or building agents that need to touch office files without wrecking them, this is worth a close look. It feels like a small CLI on the surface, but the operating idea is bigger, agent work gets better when the file format stops being a trap.