Field notes on AI coding agents: what they cost, where they break, and what shipped. Figures without a traceable source were cut.
Written with AI assistance. Figures without a traceable source were cut before publishing.
Two brothers ran a single short-video account until one piece of content took off, and only then copied it across several accounts, which is how the cumulative play count passed 58 million. The order is the entire lesson, and it is the part most people reverse.
Rob Hallam posts 3 to 5 videos a week and gets around 1,000 views on each one, which is nowhere near a viral number and was never meant to be. Across the 4 zero-budget cases in the source material, roughly 95% of users arrived through content rather than paid acquisition. Nobody in that group bought reach.
Here is the filter that decides whether this channel can work for your product at all, what a fixed posting schedule actually costs you, and the point where opening another account stops paying for itself.
The qualifying question is not whether your product is good. It is whether someone can watch it work in under a minute.
That filter is narrower than it sounds and it excludes real products. A SaaS tool with a visible before and after passes. A browser extension that removes an annoyance passes, because the annoyance is the setup and the removal is the payoff, and both fit in one shot. A physical product passes almost by default. What fails is anything whose value only becomes clear after the viewer understands a workflow they do not currently have, which describes most infrastructure, most B2B platforms with a long evaluation cycle, and most tools whose benefit is measured in a quarter rather than in a click.
If your product fails the filter, the honest move is to stop reading rather than to make longer videos. Long explanation videos are a different channel with different economics, and treating them as the same thing is how people conclude that short video does not work when what actually happened is that they brought the wrong product to it.
The thing that separates the working cases from the abandoned ones is not talent. It is that fixed frequency, a repeatable hook, and a written process replace inspiration and willpower, both of which are unreliable inputs that happen to feel like the important ones.
3 to 5 videos a week is a commitment measured in hours, and the hours are mostly not the ones people expect. Filming is fast. Editing is fast enough once the format is fixed. What consumes the week is deciding what the next video is about, which is exactly the decision that a fixed schedule is supposed to make cheap and which stays expensive as long as you are choosing topics one at a time.
That is why the material library matters more than the camera. One founder-led business treats a set of 2,000 curated customer questions as its script repository, on the reasoning that 2,000 real questions are already 2,000 video premises, and the same repository feeds the official account, the founder’s personal account, and the topic-specific sub-accounts. When the topic decision is already made, the marginal cost of the next video approaches the cost of recording it.
If you do not have 2,000 questions, you have a support inbox, and it is the same asset in worse packaging.
Cormac Hayden’s videos open with the problem, then the fix, then the ask, in that order. That structure is not a stylistic preference. It is a filter applied to the viewer in the first moment, since a viewer who does not have the problem should leave immediately, and the sooner they leave the less they cost you in completion rate.
Which brings back the brothers and the ordering question. They found content that worked on one account before replicating it, and the replication is what produced the play count, but the validation is what made the replication worth doing. A matrix multiplies whatever you already have. Four accounts running unproven content produce four times nothing, at four times the posting effort, and the effort is real even when the reach is not.
I think this is the most common failure in the whole approach, and it is attractive precisely because it looks like progress. Opening accounts is easy and measurable. Testing hooks is slow and mostly returns bad news. Given a choice between a task that feels productive and one that produces information, people reliably choose the first, then conclude that the matrix strategy is overrated.
The competitor teardown is the cheap version of testing. Pull apart accounts that already work in your space along the dimensions that actually vary, which is to say positioning, topic selection, title construction, body structure, the evidence they use, visual treatment, how they handle replies, and how they convert. Read the comments underneath, because the comments are where viewers describe the situation they are actually in, and the next topic is frequently sitting in a follow-up question that nobody answered.
Generating content with AI and posting it on a schedule is the obvious way to hit 3 to 5 videos a week without the week eating you. It is also the fastest way to get quietly downranked.
The mechanism is completion rate. Platforms do not need to detect that content is machine-generated, and they probably do not try, because they can measure whether people watch it to the end, and generic content fails that test on its own. One published pipeline reports moving completion rates from a typical 15% to 20% band up past 60%, using competitor teardowns to set the target, a distilled voice to keep the output consistent, a de-AI pass on the fingerprints, and a quality gate that blocks anything below the bar from going out at all.
The gate is the part worth stealing. Every other step improves the average, while the gate is the only one that prevents the specific bad outcome, which is a run of weak videos teaching the ranking system that your account is not worth showing. Averages recover. Rankings recover slowly.
Two cautions belong on the same page as the wins, because the wins are the ones that get quoted.
The first is that this channel is not universal even within a single product. The same self-publishing tool has to reach experienced authors, who sit in Facebook writing groups, and beginners, who are on YouTube and TikTok watching tutorials. One product, one price, 2 completely different places to be found. Deciding that short video is your acquisition strategy is really a bet about which segment you are going after, and it is worth making that bet explicitly rather than discovering it later.
The second is that a channel-driven product can peak. StoryShort matched, in 3 months, the revenue that an established B2B tool took 2 years to build, with Stripe records showing about $22,000 a month and cumulative revenue approaching $500,000. It was later listed for sale at $1.2 million on a 4.4 times annual revenue multiple, after a 30-day stretch where revenue came in 11% below the previous 30 days. Arguably that is a success story either way, and it is still a reminder that content-driven growth curves bend.
The pattern that holds across every case here is smaller than the headline numbers suggest. A product that can be shown quickly, a schedule that survives a bad week, a hook that was tested before it was multiplied, and a library that makes the next video cheap. Jordan’s Resellbot started as 30 lines of JavaScript written because his partner was manually reposting listings every day, and it now runs at roughly $63,000 a month, which probably says more about picking a visible problem than about any distribution tactic.
Start with one account and one format. Multiply after the numbers give you permission.
Also readable on Telegraph.
Read next
Part of ai-coding-field-notes — field notes on AI coding agents. Found something wrong, or shipped something similar? Open an issue — corrections are the point.