I learned how engagement works on Instagram at Kwanda. Then I used an AI workflow to create posts that consistently get over 1,000 likes. It's one of the cleaner examples I've built of human taste plus automation, so I thought I'd share how it works.
How it works
The first part is very human. You need to figure out the formula for yourself. I learned over time that the news headline style format really works for Instagram, something like "Ghana just allowed free visas for all African countries."
For our audience, which is the African diaspora, the topics that resonate are novel developments on the continent, visa policy, and, for some reason, anything related to Senegal. The agent actually helped us identify granular patterns by reviewing our Instagram engagement data, but I had to do the taste work first.
Once I had the formula, I built an automated workflow around it. Now I can say to the agent, "Hey, Uganda just released a new visa policy. Research it, draft the post, push it to Notion for me to review, and once I approve it, create the Instagram slides."
… Idea to scheduled post in about 5 minutes.
Here's how to build it
- Figure out the formula first, the human way. Go through your Instagram feed and your top-performing posts to identify the pattern. For us, it was news headline style hooks.
- Get deeper insights by giving your engagement data to an AI agent. Use something like Claude. Ask it to find patterns you may have missed. Export your Instagram analytics or connect your account through the Meta API. It can reveal details you might not see, like which posts get more saves and which get more shares.
- Give the agent a few trusted news sources. For example, use outlets like Reuters and the Associated Press. Set up an agent that can log in to your Instagram account and review your feed. It should look for high-performing posts from other accounts that you can adapt. For example, if a post from OkayAfrica performs well, you can create your own version. Keep the input focused. Give the agent enough quality sources to find fresh ideas each week, but avoid adding too much noise.
- Train it on your exact post format. Show it five or ten of your best performers, broken down line by line: hook, context, stat, call to curiosity. The more specific the template, the less editing you do later.
- Now your agent is ready. Either turn news articles into posts or have it go onto the web and do research in order to create posts.
- Create a review step. I personally have the agent draft the post, then push it into my Notion database, which I set up as a library for all our Instagram posts. This is the checkpoint that keeps quality up, because AI will occasionally pitch something that looks right but feels off.
- Once approved, have it generate the carousel slides or a single image post. Tools like Bannerbear, Figma, Canva's API, or a custom HTML-to-image workflow all work. Pass the agent your brand template and let it fill in the copy.
- Handle the hook image yourself. AI still can't pick a striking photo that makes someone stop scrolling. Grab it manually, then drop it into the carousel or single image post.
- Close the loop with agentic scheduling. Instruct your agent to download the assets, assemble the carousel or post, and schedule it through Zernio (what I use), Buffer, Later, or the Instagram Graph API.
- Add a weekly learning loop. Set the agent to scan your Instagram page or pull your engagement metrics once a week, and note any post above a threshold (we use 600 likes) in a success library. Next time it drafts, it reads from that library first, so it's always working off your latest wins.
What I've learned running this
The taste step is not optional. I tried skipping it early on and asked the agent to figure out the formula from scratch. The posts it came up with were technically fine, but never hit. You have to do the pattern matching yourself first.
Keep the approval checkpoint. It's tempting to let the agent auto-publish once it's reliable, but the 30 seconds you spend reviewing the draft is what catches the tone-deaf ones before they go out. We post at 1,000+ likes consistently because every post still gets a human eye.
The success library compounds. Three months in, the agent is writing off a library of about 40 winning posts, which is a bigger sample than I'd ever give it manually. That's where the "improves over time" bit actually kicks in.
Tools
Claude Code for the agent. Notion for drafts and the success library. Canva for carousel assets. Buffer for scheduling.
If you've got a format that works on any social platform, you should try this. Most of the value is in having the discipline to codify what already works, not in the AI itself.