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.
The videos run about 60 seconds. The developer posts 3 to 5 a week across several platforms, mostly product demos and short tutorials, and 95 percent of the users arrive through that organic content rather than through anything paid. No expensive equipment, no editing skills, no agency. The 95 percent is the number worth being suspicious of, and the 12 hours a week behind it is the number worth doing arithmetic on.
Here is what a 60-second demo has to contain to be worth posting, why the weekly time budget is tighter than it looks, and where the organic numbers stop being believable.
A solo developer built most of their user acquisition on short-form video. The videos are around 60 seconds, the format is product demonstrations and concise tutorials, and the cadence is 3 to 5 a week across YouTube and TikTok. Most of the audience arrives organically.
That last claim is the one everyone repeats and nobody verifies.
A 95 percent organic share is not implausible for a product with no ad budget, because when the denominator contains no paid traffic at all, organic wins by default. What the number does not tell you is the size of the audience. A product with 200 users and a product with 200,000 can both report 95 percent organic, and the strategy that produced the first one is not evidence for anything. My guess, and it is only a guess, is that the audience here is somewhere in the low thousands, which would still be a real result and a very different one from what the percentage implies. Treat the ratio as a description of where this developer spent their effort, not as a benchmark to hit.
The same caution applies in the other direction. Brian Donatiello’s approach to AI-generated low-content books works on the same principle a good demo video works on: match what someone is already searching for instead of persuading someone who is not. Video and search look like different channels and behave like the same one.
Editing takes about 2 hours per 60-second video, including AI-assisted effects and transitions. The overall commitment is roughly 12 hours a week on video creation and platform management, and AI tools cut the scripting and editing time by about 30 percent.
Put those three numbers next to each other and the picture changes.
Four videos a week at 2 hours each is 8 hours of editing before anyone has written a script, answered a comment, or looked at an analytics dashboard, which leaves about 4 hours for everything else in a 12-hour budget. That is not a side activity. It is a part-time job attached to the product, and the 30 percent that AI tools give back is the difference between it being sustainable and it quietly stopping in week six.
This is where most indie video strategies die, and they do not die from bad content. They die because the founder had a good week of shipping and skipped the videos, then skipped them again. The failure never announces itself either, and it probably shows up as a three-week gap that nobody can account for afterwards.
Repurposing is the only real answer here. Roughly 60 percent of the video content gets reused across platforms, cutting production time by about 40 percent, and the reused clips hold engagement rates comparable to the originals. The important word is comparable. If repurposed content performed noticeably worse, the 40 percent saving would be an illusion, since you would need more posts to get the same result.
Cross-posting is not free either. Each platform wants its own aspect ratio, its own caption style, and its own first two seconds, so a genuinely reused clip is one that was shot with reuse in mind rather than one that was cropped afterwards.
An independent developer automated an entire publishing workflow using two AI skills, shiwen and ai-wechat-publisher, going from topic selection to a finished draft in under 3.5 hours. The person doing it was not a programmer.
Under 3.5 hours is a believable number precisely because it is unremarkable. It is not a claim that AI wrote the piece. It is a claim that the mechanical parts, the outlining and formatting and shuffling of assets between tools, stopped consuming an afternoon each.
Competitive research compresses the same way. WorkBuddy and BrowserAct together generate a complete competitor price list in about 5 minutes and a product opportunity report in about 7, which replaces the browser-tab afternoon that most solo developers either do badly or skip entirely.
Notice what none of these tools do. They do not decide what to build, which product to compare yourself against, or whether the demo you just recorded is worth a stranger’s 60 seconds. Judgment stays where it was, and the tools take the part around it. That is also why the AI product manager salary range of 300,000 to 800,000 exists at all: the job that survives automation is the one that decides what should be automated.
The revenue arithmetic behind a $1 million annual target is simple enough to do on a napkin. Selling 18 units of a $150 product every day gets you there, and 18 daily sales at a 2.5 percent landing-page conversion rate requires roughly 720 visitors a day.
720 daily visitors is the number that should make you pause before recording anything.
That is not a viral spike, it is a sustained floor, every single day including the weeks when a video underperforms. Short-form video is spiky by nature, so a channel producing an average of 720 daily clicks is one where a few posts carry the year and most of them carry nothing. Building a revenue plan on the average of a distribution that lumpy is how founders end up feeling that the strategy failed in a month where it simply reverted to the mean.
There is also a quieter finding in this case that deserves more attention than the headline number. AI tooling produced about a 20 percent increase in content output with no corresponding increase in monetization.
That gap is the whole lesson. More output is not more revenue, because the constraint was never the number of videos. If the demo does not land, making 20 percent more of it produces 20 percent more of nothing, and the founder is now busier while measuring the same revenue. I may be reading too much into a single case, but a 20 percent lift that moves no money at all looks less like a slow start and more like a signal that the bottleneck sits somewhere the tooling never touched, probably in what the first two seconds of the video promise.
The tools also cost something before they save anything. Around 8 hours went into learning them, and their output quality varies enough to need manual adjustment that nobody accounts for in the original estimate. The real cost of the 12-hour weekly commitment is not the 12 hours. It is what those hours would otherwise have produced, which for a solo developer is usually the product itself.
Prior video experience is not the barrier it looks like. Formal production training is not required for a 60-second screen recording of software doing something useful, and AI tools cover most of the gap that skill would have filled.
The things that do carry over are duller. Platform algorithms shift, so content strategy needs revisiting on the order of every few months rather than being set once. A/B testing formats moves engagement measurably. Answering comments quickly moves it as well, and both are habits rather than projects.
Publishing quality gates matter more than any of it. A structured gate that lifted article read rates to 60 percent did so by reducing algorithmic demotion, which is a reminder that platforms are not just distributing your content, they are grading it, and the grade compounds.
Skip this approach if you cannot commit 12 hours a week for several months, or if your product does not demonstrate visually. Both are load-bearing. A tool whose value only appears after a week of use has no 60-second version, and no amount of editing skill invents one.
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.