Layout Guide

Templafy is a document automation platform for PowerPoint, Word, Excel and Outlook. Admins, the brand and template teams, set up templates, slide libraries and brand assets. End users build documents from them every day.

For the last year the focus was AI-generated presentations. An end user prompts for a deck, and the agent writes the content and builds it from the company's own branded slides. For a large brand, generic slides are not enough. They need control over the layouts.

The end-user side. A prompt turns into a branded deck.

Other tools were shipping presentation generation at the same time, some as add-ins inside Microsoft. Generating slides was not the differentiator. Generating them on the customer's own layouts, every time, was.

Lead product designer on document agents, with two product managers, a discovery engineer and three engineering teams. I ran discovery, interaction and UI design, and the flow end to end for three parts: the custom document agent, the Layout Guide and the knowledge source inside the agent configuration.

How the parts connect. The Layout Guide sits upstream of everything the agent generates.

Layouts had no purpose, no placeholder rules and no content type. The agent had to make two calls with nothing to go on.

I interviewed three admins from private-preview customers and one solution partner that sets up templates, libraries and automations for brands.

The master can carry the data: name each placeholder correctly and the model can read it. That is a lot of manual work, layout by layout, on files that are rarely touched. Some masters also hold only a handful of layouts, so decks come out repetitive.

The slide master supports typed placeholders. Almost nobody uses them.

The discovery engineer tested a vision model that reads a whole deck, works out what each slide and each placeholder is for, and writes that metadata itself. The agent reads it while compiling a presentation, to select the slide and fill it.

Four places were on the table. The choice came down to who sets it up and what product it is sold with.

A placeholder needs a type, and a type is not specific enough on its own. I mapped a second level: the hint that tells the model what kind of content belongs there. Grounded in what the slide master supports, and extended where it needed to go.

The first iteration surfaced everything the model could extract, for the layout and for each placeholder. Rich, and busy. Did the admin need all of it?

The model could only process text and image placeholders. I stripped the metadata back to what drives two things: how the agent picks the layout, and how it fills the content. A layout keeps a name and a usage. A placeholder keeps usage instructions and, for text, a character count.

The count is calculated, not guessed: font, font size, placeholder size and the average letter width for that font. The admin can step it up or down. It stops the agent handing a paragraph to a slot built for one line.

Phase one shipped with a single uploaded deck. The admin uploads one branded presentation, each slide is processed on its own, and the list fills with a preview, suggested name and usage, and placeholder count. Opening a layout shows the model's draft, which the admin overwrites placeholder by placeholder. Saved layouts join the set the agent builds from.

Select a placeholder in the list and its box lights up in the preview. Select a box in the preview and its fields light up in the list. Without it, admins cannot tell which instruction they are editing, and reviewing a layout or troubleshooting a bad result gets much harder. This was the one decision I would not compromise on.

A whole deck at a time was cumbersome. Phase two let admins add one slide or many from the library, each landing in the guide as its own layout and reviewed in the same modal. Admins can filter the library by layout usage and remove a layout without deleting the slide. Nothing blocks: the admin adds and carries on, and gets a notification when the layout is ready.

Managed one layout at a time: added, reviewed, removable.

Categories stayed out of scope. They meant more work for the model on every import. Through the whole run no admin missed them.

• Three enterprise brands onboarded in private preview.
• Phase two shipped, and the whole-deck upload was phased out.
• The set of layouts can grow, instead of the handful a master carried.
• With usage instructions to read, the agent stopped falling back on the same few layouts.

The project was commercially driven. Working closely with sales, with demos and deals in mind, put the setup inside the AI Agents menu point.

I would start from library configuration, as the first step of a brand management overhaul. Tone of voice, corporate terminology and colours sit under AI assistant and productivity tools today. They belong in one place: a single menu point for brand management.