Presentation & Slide Automation

Adoption made me happiest. The workflow landed well with the team and made deck creation noticeably easier. Before, almost everyone got something wrong: wrong theme, wrong format, body text dumped in as a wall. That doesn't happen anymore.

The problem

In a team running several brands, "make a deck" is not a single step. Which brand? Which Gamma theme belongs to it? 16:9 for the client, A4 for the document, portrait for social? And how do you split the text so it works as slides?

These decisions get made on every new deck, often by people who barely touch Gamma. Wrong theme picked, wrong format, text inserted with no slide breaks. The result doesn't look professional.

The approach

The workflow starts with a bilingual form. You pick everything at once: text, processing mode (generate, condense, or preserve structure), output format, target brand, language, slide count, optional extra instructions.

After submit it runs on its own. The workflow translates form inputs into Gamma API language, prepares the text for slides, calls the Gamma API, waits for completion, and sends the finished asset by email. Preview link and download link, no further interaction.

I'd picked up experience with the Gamma API on the content repurposing project and was surprised how well and how simply it works. The Gamma Creator followed naturally. Not for a single use case anymore, but as a flexible team tool.

From form to API

The form speaks the user's language: brand names, readable format labels, mode options in German and English. The Gamma API wants theme IDs, folder IDs, language codes, API-shaped parameters. The two worlds need translation.

I solved it with an Edit Fields node: form labels get scanned and mapped to the matching Gamma parameters via lookup. Each brand has a fixed mapping to theme ID and folder. New brands need a new mapping entry. Nothing else in the pipeline changes.

Folder discipline inside Gamma was also a thing. Without folder structure, everything piles up in one heap. Each brand got its own folder, recorded in the mapping.

The format decision

Deck (16:9) and document (A4) are direct paths. For social there's a second form step: pick the aspect ratio, 1:1, 4:5, or 9:16.

I went straight to the two-step form because it's the better UI. Someone building a client deck shouldn't be hit with social-format options they don't need.

The text prep

Gamma generates decks from text, but the input has to be split into slide-sized chunks. A LangChain agent with GPT-4.1-mini handles that. The system prompt is tightly scoped: preserve or impose structure, don't change content, no extra additions, mark slide breaks.

Behaviour shifts by mode. "Generate" expands a topic into a full slide deck. "Condense" trims existing content to the essentials. "Preserve" restructures the text without touching the substance.

GPT-4.1-mini was a cheap model at build time, and this step is pure sorting work. It works. Never change a running system, especially when it costs almost nothing.

The important thing: the AI must not destroy context. For users the workflow is a black box: text in, deck out. If the model starts rewriting content or inventing additions, no one catches it immediately. That's why the prompt is as restrictive as it gets.

The delivery

After the API call the workflow polls the job status on a regular interval. Simple polling, works reliably.

Early on I sent chat messages and emails in parallel, mainly to confirm runs were going through. The chat channel is off now. The email is the actual confirmation: HTML layout with preview link and download link.

The result

Fill in the form, submit, wait briefly. What comes out:

• Gamma asset in the selected format (16:9, A4, or social portrait)
• Correct brand theme with logo, brand colours, and font
• Text rendered as structured slides, not a dumped paragraph
• Preview link for browser editing
• Download link as PPTX or PDF
• All delivered by email

What I'd do differently today: fork the form more aggressively, guide the user more. And allow file and image uploads, so it's not text-only. The workflow itself works; the improvements live on the input side.