I’m writing this post for creative strategists or anyone who wants to use AI to make more/better/cheaper ads this year.
This is how we started integrating AI into our creative workflow and how we’re upskilling our Creative Strategists into Creative Strategist Prompt Engineers.
We’ve broken this down to a 4-step process internally where everything is about building prompts and workflows that compound. Ie. I can build something once and it serves me (and my team) for a long time. It can be as simple as a single prompt or as complex as an agent.
Step 1 - Integrate tools into your existing processes
There is no “infinite ads machine” tool that works yet. A way more defensible approach to take with AI (today) is to analyze your existing processes and work out which AI tools can enhance them. For example:
Usually scrape Reddit manually? → Start using Reddit Answers (new feature where you can chat to Reddit as you would with ChatGPT)
Usually scrape the internet to understand your customer → Use DeepResearch to put together an in-depth analysis.
Usually find ads on ad libraries and recreate them? → Build a Poppy board that does this
Step 2 - Start training AI models
The biggest complaint I hear about AI from strategists is that the ad ideas aren’t good enough. That is a complete myth. If your hooks, headlines, scripts or whatever are not good enough, then that’s because you haven’t trained the model properly. 100% a skill issue.
Here’s the way I think about writing prompts (aka training AI models):
“If there were an intern starting tomorrow that I would never meet and never be able to communicate to, and I could only give them one set of instructions to complete this task every day for the rest of their life, would this change the instructions (prompt) I would give them?”
For most people, the answer is yes. I try to include most (if not all) of these core prompt elements in every prompt I make:
- Identity (persona, tone, brand voice)
- Task
- Good examples of the task (and why they’re good)
- Bad examples of the task (and why they’re bad)
- Context (knowledge, additional info)
- Output format
The most important elements are the task and good examples. Stacking as many successful outputs and training the model on why they’re successful is the 80/20.
Ideally, these success outputs are examples of headlines/hooks/scripts that have worked inside your ad account.
Step 3 - Build Prompt Libraries
So much executor time is lost in rewriting prompts that either you or someone on your team have already built. Prompt libraries are something that every business owner reading this should set up for their organization.
A prompt library is a database of all your prompts, categorised and tagged for ease of access. We have ours built in Notion, but you should build it in whatever makes it most frictionless for your team to add to it - as long as you can build databases in there (e.g. Sheets, Airtable…). Every time someone builds a new prompt, it gets added to the database.
For example, ‘extracting static headlines from customer reviews’ or ‘finding hook ideas by browsing Reddit’. Now every time a strategist wants to complete this task in future, they already have a pre-written prompt, that’s trained on what works for your brand, to paste in. We also version control it so that we can iterate over time.
Step 4 - Build AI workflows
Once you have models trained on your data and have started to share prompts across your team, it’s time to start building your own workflows. The builders we mainly use are Gumloop, n8n and Zapier.
Important sidenote: There is a difference between workflows and agents. An agent is autonomous, a workflow is a set of scripted steps. Everyone is using the term ‘agents’ because it’s sexier, but right now pretty much everything you see in our space is a workflow. Agents will come next and that’s when things will get very interesting.
Here are some workflows that we’ve built internally:
- Facebook Ad Comment Scraper - scrapes comments and conducts analysis
- TikTok Organic Scraper - scrolls TikTok, finds relevant videos and suggests them, then writes scripts
- Facebook Ad Library Scraper - scrapes competitors’ ad libraries and finds inspo
- On-brand Analysis (Content Review) - reviews content to check if it’s on brand
With the amount of nodes already built into these platforms, the possibilities really are endless. And again, we are not looking to reinvent the wheel here. We are just taking processes that are already conducted manually and turning them into repeatable workflows that can be shared across the team.
Anyway, I’ll wrap this up here as my flight is about to land. Curious to hear how y’all are approaching AI as creative teams and if you enjoy these longer format posts