MIT licensed· open source· works with Claude, ChatGPT & any AI agent

hands on deck

The PowerPoint app, built for AI agents.

Give your AI agent a slide deck and it will edit it the way you would — open it, make the changes, look at the result, and fix what's off. Without wrecking your template, your brand, or the file itself.

Watch the demo
> /plugin marketplace add EveryInc/hands-on-deck
GitHub
P
PowerPoint
for humans
K
Keynote
for humans
S
Google Slides
for humans
>_
hands-on-deck
for AI agents

All four work on the same file. Your decks can live in PowerPoint, Keynote, or Google Slides — if it saves as .pptx (the world's most common deck format, and all three speak it), hands-on-deck works on it.

§ 01Why AI keeps wrecking decks

You see a slide. The AI sees
ten thousand lines of code.

Here's the secret every deck app hides from you: a .pptx file is really a zipped folder of hundreds of code files (a format called XML). PowerPoint reads that code and paints the slide. An AI agent doesn't get the painting — it gets the code. Every AI deck disaster you've seen starts right here.

What you see

one slide · a title, four bullets, a chart
ten seconds to take in

What the agent gets

the same slide · ~2,000 lines of this, per slide
a 20-slide deck ≈ tens of thousands of lines

Until now, an agent had two bad options.

Edit the code by hand fragile · slow

How most tools work today — including the AI apps' built-in deck ability: the agent unpacks all those files and types its changes directly into the code.

It's slow, it burns through tokens, and one typo can make PowerPoint refuse to open the file. Worst of all: an agent spending its whole attention on writing code has none left for how the slide actually looks.

Regenerate from scratch lossy

The other way: throw your file away and generate a fresh one from a description.

The result never matches. Your template, brand fonts, logo placement, image treatments — every decision your designer ever made lives in the original file, and regeneration erases all of it.

Use an app — like you do hands-on-deck

You would never edit a deck by rewriting its code. You'd open an application built for the job.

hands-on-deck is that application, built for agents: the agent says what should change, the tool changes it. The code stays the tool's problem — never the agent's, never yours.

§ 02How it works

Agents can't point and click.
They're brilliant at typed commands.

An AI agent can't drag a text box in PowerPoint — those apps need a mouse and eyes on a screen. But there's one kind of software agents operate masterfully: command-line tools — programs driven by typed instructions. It's the entire reason coding agents like Claude Code and Codex work so well. So we rebuilt the PowerPoint app as one. Here's a real edit, start to finish:

1 you ask

In plain English

You talk to your agent the way you'd brief a colleague.

"We're pitching Acme now — rebrand the deck, and swap the old chart on slide 4 for this new screenshot."
2 it writes instructions

Intent, not code

The agent writes a short, readable list of changes. This is the entire "program":

replace-text everywhere: Globex → Acme swap-image slide 4: new-chart.png
3 the tool edits

Safely, all at once

Every instruction is checked first — if anything's wrong, nothing is changed. New text keeps the old text's fonts; a swapped image keeps its exact frame.

validating 2 ops… applying… ✓ saved
4 it checks the result

With its own eyes

The tool hands back pictures of the finished slides plus a report of anything that looks off — and the agent fixes what it finds.

slide-4.jpg rendered ✓ no new layout issues

The agent never touches the file's code. It states its intent and verifies the outcome; the tool does the bytes. The agent thinks like a designer instead of a programmer — which makes it faster, dramatically cheaper in tokens, and leaves it attention to spare for what actually matters: how the slides look.

§ 03What it can do

Everything you'd do in the app —
as a command.

Every click and drag you use in PowerPoint has a typed equivalent here. A few of the ~30 operations, translated:

In the app, you would…
The agent runs…
Click into a text box and retype the title
set-text

New words, same look — fonts and formatting carry over automatically.

Find & replace the client's name across all 40 slides
replace-text

One command covers every slide — including the template itself.

Drag a new image over the old one and nudge it to fit
swap-image

New picture, exact same frame, position, and crop.

Change the brand colors and fonts for a new client
set-theme · replace-color

Re-theme the entire deck in a couple of commands.

Copy that nicely-styled box, paste it, change the words
duplicate

Clone any shape with all its styling; just supply the new text.

Reorder slides, or pull three good ones in from last quarter's deck
slides · merge

Restructure within a deck, or across decks.

Design a brand-new slide on a blank canvas
html2patch

The agent designs the slide as a small webpage — the design language it knows best — and the tool converts it into real, editable PowerPoint shapes, placed inside your existing template.

Zoom in and squint — does everything actually line up?
render · lint

Slides come back as images, with a proofreading report. More below.

Tables, speaker notes, backgrounds, transitions, hidden slides, alt text — covered too. The full command reference lives in the repo; agents pull it up themselves with deck.py docs.

§ 04The safety net

It checks its own work —
like you would.

The difference between an AI that makes slides and an AI you can trust with your deck is what happens after the edit. Here, verification is part of the job — not an afterthought.

It looks at the slides

After every change, slides render to images the agent inspects with its own eyes — the same way you'd flip through a deck before hitting send.

A built-in proofreader

Text spilling out of its box. A shape hanging off the slide. Text hidden behind an image. A box a hair out of line with its neighbors. The tool measures all of it — in exact inches, with the exact fix — and flags only what the latest edit broke.

All-or-nothing edits

Every instruction list is verified upfront. One mistake among ten instructions? Zero changes are written, and the agent gets all ten answers at once. You will never open a half-edited deck.

Honest repairs

The auto-fixer re-measures after repairing. Anything it couldn't fix is reported, not papered over.

§ 05Make it yours

Your template stays.
Your rules go on top.

It works inside the decks you already have

Because hands-on-deck edits your file instead of regenerating it, everything your designer decided — master slides, layouts, brand fonts, logo placement — survives untouched. Ask for a new slide and it's built from your template's layouts, not a generic default.

That's the biggest difference from "AI deck generators": they make their deck with your words in it. hands-on-deck makes your deck.

The engine under your firm's own deck skill

Teams don't just need edits — they need edits their way. hands-on-deck is built to be the engine underneath a thin skill of your own: a page of your firm's rules that the agent follows while the engine does the mechanics.

Change your rules anytime without touching the engine — and keep receiving engine improvements automatically.

§ 06Install

Use it with the AI you already use.

hands-on-deck is packaged as an Agent Skill — a folder of instructions and tools any modern AI agent can pick up and use. No server, no account, no API key.

Claude Code / Codexagent CLIs > /plugin marketplace add EveryInc/hands-on-deck    > /plugin install hands-on-deck@hands-on-deck
turn on marketplace auto-update, and every improvement we ship arrives on its own
claude.ai / ChatGPTchat apps zip the skills/hands-on-deck folder → upload it as a skill in your app's settings
Any other agentit's just a CLI $ git clone https://github.com/EveryInc/hands-on-deck  → put deck.py docs in front of your agent

Our decks are made in Keynote / Google Slides — does this work for us?

Yes. Save or export as .pptx — both do it cleanly, and Google Slides opens .pptx directly — and hands-on-deck works on it. The file format is the standard; the apps are interchangeable.

Why is this better than my AI's built-in deck ability?

It makes the agent think like a designer instead of a programmer. That means faster edits, a fraction of the tokens — and enough spare attention to actually sweat the details on the slide.

Do I need to be technical to use it?

No. Install it once — two commands, or one zip upload — then just talk to your agent about your deck. Everything on this page is the agent's job, not yours.

What can't it do?

Honestly: native editable charts, shape animations, and embedded video are out of scope — the 5% where you still open PowerPoint yourself. It does the 95% real decks are made of, and says so when it can't.

Who's behind it, and what does it cost?

Built by Every Consulting, distilled from making hundreds of real client decks with AI. Free and open source (MIT) — fork it, extend it, ship it inside your own product.

I'm an engineer. Where's the real documentation?

The GitHub README has the architecture: the patch format, atomic validation, the linter's geometry checks, html2patch, and the eval suite we test against.

the close

Go give your agent its own PowerPoint.

github.com/EveryInc/hands-on-deck