Case study · May to October 2026
A one-person media business, run on Claude.
In five months, one person working around a full-time job built a science-education publisher that researches, scripts, packages and publishes largely through AI-run workflows.
- The business
- A science-education YouTube channel and online publisher in a health niche. My own business, so I describe it without naming it.
- The team
- One founder, part-time, with Claude as the operating layer.
- The challenge
- Publish twice a week at high factual accuracy, and run a website, an email list, digital products and a paid report service, with no staff and a full-time day job.
- What was built
- 12 custom AI skills, eight of them chained into one production line, a website that publishes itself, an email and sales funnel, quality checks between stages, a custom teleprompter app, and one source of truth with a tested disaster-recovery plan.
The production line
Each stage hands a finished file to the next. The owner only does what a person has to: record, edit, approve and press publish.
| Stage | What the AI does | What the owner does |
|---|---|---|
| Ideas | Scans the web and news, pulls keyword and competitor demand data, and filters out repeats of earlier videos | Picks the ideas |
| Packaging | Scores title options, writes the thumbnail image prompt and five hook options, then runs an AI review panel set up as a behavioural psychologist and a marketing academic | Approves the title and thumbnail |
| Script | Researches, checks every claim against its primary source, and writes the script in the owner's voice to a length model built from the owner's own past videos | Reviews the lines |
| Script checks | A blocking quality check: voice rules, a cold read of the opening by a separate AI reviewer, and a test that the script displays correctly in the teleprompter | Nothing until it passes |
| Record | A custom teleprompter web app follows the owner's voice and keeps the place | Records |
| Post-production | Builds data and diagram animations, proofs the captions against the audio and writes the image prompts | Edits and exports |
| Upload | Writes the description, sources, chapters, tags and pinned comment, with a compliance check on every link | Uploads and schedules |
| Website and email | From one sentence, writes the matching article, dates it for release and drafts the newsletter | Double-clicks one push script |
| After release | Pulls the analytics and diagnoses an underperforming video with up to three ranked fixes | Decides |
Results
Records as at 24 October 2026.
- Videos produced
- 89 since the first release on 30 May 2026
- Released
- 37, from 30 May to 23 October 2026
- Scheduled to release on their own
- 52, through to 4 April 2027
- Web articles
- 90 written; each publishes itself after its video
- Owner's effort to publish an article
- One double-click
- Errors caught before publishing
- Citation checks found a defect in each of nine videos in a row, including three wrong first authors
Hours saved weren't tracked, so this case study claims output and quality, not time. A before-and-after time study is the obvious next measure, and it is the first thing every client engagement does.
How it was approached
Designed like a process plant
Defined stages, one source of truth, checks between stages and a safe fallback for every failure.
AI review as a built-in step
Anything with a number in it, or anything a customer pays for, goes to a separate AI reviewer that never saw the draft. A second AI panel tests every title and thumbnail. The owner gives the final sign-off on everything.
Checks, not good intentions
Where a mistake is costly the check is mechanical and blocking: a link-compliance test, a fact check against the source, a layout test, a pre-publish checklist.
Every lesson becomes a rule
In August an AI-written clean-up script moved far more files than it should have. The backup and written restore plan recovered everything except some raw footage. Now every script that moves or deletes files runs as a test first, prints what it will touch and stops above a set limit. The same rules are built into every client system.
What this means for your business
The same pattern fits any business with repeatable knowledge work: map the process, automate the drafting and checking, keep people on the decisions, and put a check where mistakes are expensive. Tenders and quotes drafted from your past wins, reports from your own systems, a website that keeps itself current.