The Lab Behind Our Instagram

Introduction

This is How We Grow, where Lambert and I write up what we're doing to grow Kai. We recorded #018 in the room on Friday 23 October, just the two of us, with David away. The topic is Lambert's. For about 3 weeks he's been building what he calls a laboratory for our Instagram, and I wanted all of it on the record: how it started, what Claude does in it, and whether any of it is turning into business.

Where things stand: @buildingkai_ had 299 followers on 23 September and 983 on the day we recorded. Lambert now publishes about 17 pieces a day on Instagram and 10 on TikTok, plus YouTube, weekends included. The goal is 10,000 followers by the end of the year.

Prefer to watch or listen? Here's the full episode

We filmed for 8 months before the editing caught up

Cluely showed us what behind-the-scenes content can do

Lambert's reference, back in February and March, was Cluely: one of the first companies to stop talking about its product and post behind-the-scenes skits instead. People got to know the team through the skits, and when they later needed a meeting transcription tool, they tried the one they already knew. Morgen, our first product, never did this, and Kai felt like the moment to try. So we started filming everything at our first hackathon in February, kept at it through all 3 hackathons this year, and now someone picks up a camera every week.

Editing one video by hand took Lambert 2 to 3 hours

By the second hackathon, in May, we had a decent bank and Lambert started posting. Then the editing killed it. One video took him 2 to 3 hours, with no way of knowing whether it would work. He kept it up for 2 or 3 weeks and parked it. The one-off pieces were worse: the video introducing David took 2 weeks of script and a month of editing.

Lo filled the bank, and Jonathan gave it a voice

In July, at our hackathon here in Lisbon, we hired Lo, a videographer, for 2 or 3 weeks to film everything while we worked. The bank got much bigger, and her posts gave us our first video past 100,000 views. What we still lacked was someone from the startup world to talk to the camera and tell the story, and over August and September we couldn't find them.

Lambert found him by accident. Doomscrolling one night, he saw Jonathan post about his first day at business school in Lisbon and messaged him on the spot. Jonathan started in mid-September, and in Lambert's telling his energy rubbed off on the whole team. David pushed the same way (take the camera every time, talk about what we're doing), and having the CEO ask for it helped a lot.

Then 2 things landed at once: the momentum of our Product Hunt launch, and Opus 5.5, which turned out to be very good at editing video.

Claude tags the clips and edits the reels

Claude tags every second of 800 clips

We have close to 800 clips of ourselves now, shots and B-roll, and we're adding about 100 a week. Nobody had time to review, name and sort them, so Lambert had Claude tag every second of every shot. A shot runs from 20 seconds to a minute or so, and for each second Claude notes who's in frame, what they're doing (cooking in the kitchen, at the computer, in a conversation), the place and the atmosphere. The tags live in our knowledge base, so Claude knows what's in every file on the Drive and can match a shot to what Lambert asks for.

In #009, in July, the plan was still to script every video so Claude could pick clips on its own. The tagging never needed Opus 5.5, Lambert says. It only became worth doing once Claude could edit.

One clip, as it sits in our knowledge base.

Claude edits in Lambert's Final Cut Pro setup

Opus 5.5 came out while we were making the product films for our website (that process is in #013). Watching Claude edit those, Lambert figured it could edit the clip bank too. He gave it an old video of his to improve, and the result would have taken him a long while by hand.

Next he gave it a reel he'd saved on Instagram, one of the thousands he'd bookmarked to remake later, and asked it to cook something from our clips. The edit was impressive, but the shots didn't fit the idea, which is what pushed him to build the tags. With the tags in place, he gave Claude his Final Cut Pro setup (his filters, effects and subtitles) and added motion design, which Claude handles well.

Every wrong shot becomes a note

When Claude picks the wrong shot, Lambert tells it so, Claude writes it down, and it stops offering that shot there. It also notes, without being asked, when a reel underperforms with a given music, hook or clip.

Then Instagram's data comes back, and Claude reads the retention curve for where people stop watching. My example on the episode was a reel that loses 5% of its viewers at second 4: Claude swaps that frame and cuts another version. Before this, we had no way to see which frame was losing people.

Every idea goes out 3 ways before it goes anywhere else

Trial reels show a new reel to people who don't follow us

A trial reel goes first to people who don't follow you, and it's built for posting several versions of one video with a different hook, opening shot or clip. Lambert hadn't realised the account had them. Back when we had under 200 followers, it didn't. With the bank tagged and Claude editing, he can make 10 versions of one concept, each with a different first shot, hook or music, and let Instagram pick.

He built a dashboard to run it, the Shorts Lab: a calendar with 2 to 4 trials going out a day, each read at 72 hours. His own verdict is that it isn't mature yet.

The Shorts Lab's calendar, the week Where's the office went out.

The same video got 2,364 views or 46,347, depending on the hook

On Sunday 18 October Lambert posted a trial called Where's the office, in 3 versions. Nobody talks to the camera in it. It's shots of us working, pulled from the Drive and cut to a trend Lambert gave Claude, and the versions differ only in the hook: the title and the shot each one opens on.

At 72 hours, A had 3,752 views and C had 2,364. B had 46,347, and brought 23 new followers.

Where's the office in the Shorts Lab, 6 days in.

The counts on the card are from 24 October. The 72-hour read is the one above.

Claude reads each trial at 72 hours and proposes the next one

At 72 hours Claude reads every trial and tells Lambert what it found. Some ideas get parked. Some held people until a point the retention curve shows, so Claude flags them (you posted this 3 days ago, retention held until here, let's try another version) and writes the prompt for the next cut. Lambert runs it and argues with it, because he wants human eyes on everything we post. He doesn't think AI reads the emotion in a piece of content well yet.

A clear winner gets shared to our followers too, and they reshare it, so it keeps travelling. Anything that works very well gets more versions, like the Runaway series from #017, which Lambert keeps coming back to. Where's the office B was flagged as a winner, so it went to TikTok and YouTube, which don't have trials of their own. That's the rule now: test on Instagram, then move the winner.

The same lab now clips this podcast

When Lambert said the cool part was still coming, I asked whether everything before it had been the boring part. The cool part is this show.

We used to record How We Grow on Riverside. It struggles to keep the tracks in sync, the camera quality isn't always good, and it eats CPU. It's fine for remote recordings. Now Lambert has a skill in which Claude edits the episode, motion design included, since it knows our design and our product from the knowledge base. We record in the room when we can (this one was shot on 3 cameras) and on our laptops when we can't.

When an episode is done, another prompt picks its highlights, the way Riverside's highlights feature does. Once Lambert approves them, Claude builds a set for each one: an A/B/C trial of educational cuts, a motivation video and a few fun best-ofs. #017 gave us 11 highlights, and you can see what came out of them on our Instagram. On Twitch and YouTube, clipping long videos is a job people do by hand, and as I said on the episode, until now you needed an editor for it.

#017's highlights in the Shorts Lab.

Lambert's reference here is Cluely again. He remembers its CEO saying in an interview that he paid people to clip his long-form videos and post them on X, TikTok, YouTube and Instagram, because when clips of you show up every day from different angles, people start to feel they see you everywhere. We're doing our own version of that with Claude, with a person checking every piece before it goes out.

How to copy it with a phone and Claude

Near the end of the episode I tried to boil it down, because I think anyone can do this:

1. Film a lot. Put your phone behind you while you work, and film your meetings with your colleagues. That's your B-roll bank.
2. Put it in a Drive and have Claude tag it. Every clip, every second: who's in it, whether they're smiling, whether it's sunny or night.
3. Give Claude a reference. "I saw this reel and liked it, try to reproduce it with our clips." That's your first piece.
4. Post every piece as a trial, at least 3 times. Sometimes 6, sometimes 9. A different caption or a different clip can be the whole difference. For a while we posted 1 version, got 1,000 views and were happy with it.
5. Keep the analytics next to the bank. Claude learns from every post when the numbers sit in the same place as the clips and the edits.

We have 570,113 views and haven't started converting them

On air, Lambert guessed more than 1.3 million views for the past 2 weeks and I guessed half a million. The warehouse says 570,113 across Instagram, YouTube, TikTok and X from 9 to 23 October. On Instagram, @buildingkai_ had 299 followers on 23 September, 347 on 3 October, 375 on 9 October and 983 on 23 October, so 608 of them came between 9 and 23 October.

@buildingkai_ followers, 23 September to 23 October.

I pushed Lambert on what those numbers are worth, because views are easy: you can post brain rot and pile them up without building a company. His answer is that the views are doing what top-of-funnel content does: new followers, visits to the website, some sign-ups, and a lot of people asking what the product is. What they like is the mood, real people cooking and working together. But most of what we post is skits, like our "how many tokens did you spend this week" competition, and Lambert hasn't done a single thing to convert anyone yet. No stories, no bottom-of-funnel content, no DMs.

He doesn't want to trade away what works. We don't use AI-generated images and he says we won't, because people follow humans, which is where we started in #001. Part of the audience just wants to follow the journey, and some of them might need Kai in a year, the way people who knew Cluely from its skits tried it once they needed it. The people who need it now are mostly founders, who follow us for the startup life and may not know what we make. The work is to reach them, and make sure they know Kai helps with the bottleneck in their week, without losing the people who just follow the journey.

Lambert's example of the long game is Foundable, which built a big community, with a Discord full of AI tips, before its product was out. His point is that a launch then has plenty of people to try it and repost it, even if only a sliver of them sign up and fewer pay. (Foundable has since opened its product beyond early access.)

Where I land: we need sign-ups and revenue to be sustainable, and the account still isn't going to become a marketing page. It's a cultural investment. Community first, then the right content in front of the people it fits.

What isn't working yet

• Followers aren't customers. Lambert hasn't run any conversion yet, and the sign-ups from social are small next to the views.
• It costs hours. This has been Lambert's main focus for the past 2 weeks, weekends included. There are no cheap wins in it.
• Claude still picks the wrong shot. The notes cut down the repeats, and Lambert still checks every piece before it goes out.
• TikTok shows us to the wrong audience. The account isn't localized to the US, and a lot of the people seeing our videos there speak Portuguese, which Lambert says has us a bit stuck.
• The lab is young. It's under 3 weeks old, and "not mature yet" is Lambert's own description.

What's next

The next few weeks are about conversion. Lambert wants to work out what in the strategy is wrong and what to keep, then try the first bottom-of-funnel moves and some awareness aimed at founders who'd actually use Kai. Jonathan's talking-head videos carry the trust-building side. Lambert is also pushing harder on YouTube, where he's preparing a new concept for David. On the day we recorded, @buildingkai_ was 17 followers short of 1,000, and Lambert wants to come back to this topic in 2 months to see how the lab has changed.

If you want to follow along, the whole series lives at How We Grow, and the changelog shows what we ship week to week.