Create Custom AI Workflows Using Make.com Automations

Full video transcript

Darby opens the session joined by Chris LaFond, a veteran software engineer and chief AI officer at Profit Labs, who has been building AI automations in Make.com. Chris has over 25 years of software engineering experience, including roles at General Instrument, Motorola, Hewlett-Packard, TransCore, Mitchell International, and Sharp Healthcare. He has also been an Amazon brand owner since 2018 with a best-selling brand in the 3D printing space, has been part of the Jasper community since 2021, and started Profit Labs in April 2023. He describes moving from AI “autos” scripting and prompt-stacking workflows into building automations with Make.

Asked to explain Make.com automations in simple terms, Chris says that historically you needed real technical background to interact with an API, but no-code tools like Make now let you drop in a module and fill in the properties an API call requires — and you can add an AI module into a normal automation to give it intelligence. He notes that for the demo he needed to build a custom version of Make’s Claude module from scratch because the built-in one has timeout issues.

Chris shares that he has also built an elaborate book-writing automation that got positive feedback and that he plans to put a front end on it as a product, with modules playing roles like editor and publisher as the user works through the process of writing a book. He is also exploring turning podcasts into transcripts and pulling out key points for newsletters, along with content creation and syndication workflows. For today’s demo, he sets out to replicate a “Rubber”-style front end: a web form triggers a Make automation and gets a web reply back, similar to how Rubber apps work.

Sharing his screen, Chris walks through Anthropic’s API documentation for creating a completion and starts building a scenario in Make. He explains that Make’s out-of-the-box Claude module doesn’t let you override its roughly one-minute timeout, so any request that needs more processing time fails. To get around this, he deletes that module and instead uses Make’s generic HTTP “Make an API Request” module, which does allow a custom timeout.

He sets up the scenario starting with a “Set multiple variables” module for inputs like product, segment, niche, and context, pulled from a buyer’s brief recipe he prepared. He explains that because he pasted the recipe text in from another document, escaped characters like line feeds and double quotes can cause problems, since Anthropic’s API “hates” unescaped characters — so he adds a cleanup step that converts line feeds to \n and swaps double quotes for single quotes before sending the text along.

Chris then shows the HTTP module configuration for the Claude call directly: headers for accept, anthropic-version, content-type, and the API key, plus the request body containing the prompt and a max_tokens value (he demonstrates setting it as high as 100,000, versus the default 20,000). He sets the module’s timeout to five minutes instead of the default one minute, and wraps the returned completion in an HTML document, which a webhook reply module sends back to a front-end form.

The form itself, built with a tool called pForm, was created by asking ChatGPT to build a form with a specific set of fields and an activity spinner, then hosting it and pointing its submission at the Make webhook. Darby and Chris run a live test with a “side hustle party game” example, watching the scenario execute while Chris explains each step.

Darby asks about adding a knowledge base or uploaded reference files, similar to Rubber’s app builder. Chris explains that would require upserting documents into a vector database, which is more complex, though Claude’s 100K-token context window means you can paste a lot of reference material directly into a prompt instead.

While the demo runs, Chris shows the underlying Claude HTTP module again, walking through exactly where the header and timeout values go, and confirms the response comes back as HTML which is wrapped for display in a browser.

Chris then shows his more elaborate book-writing scenario, which pulls from a Google Sheet containing the author’s role, book type, title, target audience, reading level, and description. The automation can generate either a description or outline, branches based on which one, stores intermediate results in Make’s built-in data store, and iterates through the outline chapter by chapter. Claude writes each chapter, then ChatGPT rewrites it because Chris considers ChatGPT the better writer, and the process also generates a preface based on the author’s information before converting the finished book to Markdown and HTML and saving it to Dropbox. The full book runs to roughly 30,000 words currently, though Chris wants to expand chapters and sections to make future books longer.

Chris also demonstrates an RSS-based social media automation: a raw news RSS feed (3D printing news in this case) is pulled in, the source URL is shortened, Claude is prompted to act as a 3D-printing expert and write a viral tweet under 30 words with emojis, hashtags, and a first-person voice, and the result posts to Twitter/X and a Facebook page. A related branch evaluates whether a story is newsletter-worthy and saves qualifying stories to a Google Sheet. Chris mentions he has started building a newsletter automation that would bundle RSS feeds from news, product pages, and blog posts, alongside manually written stories, to assemble a newsletter draft.

In response to audience questions, Chris discusses whether the OpenAI API can access URLs like Bitly links (possible via function calls, though more complex to automate), and whether scraping could power niche, localized content sites — for example scraping real estate listings for a city and using them to populate a blog. He suggests scraping the raw data first, storing it in Airtable or Google Sheets, and then generating content from that stored data. He also shows a module in his scenario library that posts directly into WordPress via a plugin that accepts actions from Make, including creating posts and categories and monitoring comments.

Asked what he hasn’t tried yet, Chris mentions wanting to experiment with “tree of thought” prompting frameworks, where multiple AI personas or perspectives interact and build a shared context or history of reasoning to solve problems more creively and iterate toward consensus. He’d like to apply an approach like this to the book-writing automation once it has a front end. He estimates having a testable version of the book-writing product ready by the beginning of December and invites community members to help test it.

An audience member (Bruce) describes wanting an app that generates full product sales-page copy across 15 sections, presenting three versions per section for the user to choose from or combine before moving to the next section, and asks if Make could support that. Chris says yes, though it would likely require breaking the automation into multiple runs or steps to accommodate the user’s approval process in the middle rather than a single end-to-end run.

The conversation closes with a discussion of upcoming multimodal AI capabilities — including vision features that can interpret uploaded images such as flowcharts or hand-drawn diagrams and turn them into working software or web interfaces — and speculation about robotics use cases for this kind of visual reasoning. Darby thanks Chris for the walkthrough and invites viewers to reach out to Chris directly or through the Profit Labs community to learn more about building their own automations.