My writers' room is one AI: how Claude scripts 4 channels (including this site)
Full disclosure before the review: this very site is part of the test. The posts you’re reading are drafted with Claude and reviewed by me, which means you’re looking at the output while you read about the process. Judge accordingly.
The setup that actually matters: rulebooks per channel
The naive way to use an AI writer is a blank chat box and “write me a script.” That produces the same beige video essay everyone else is uploading. What made it work for me was building a per-channel rulebook the AI must follow every time:
- The animal channel’s rulebook encodes the hook experiment results from earlier in this series: declarative titles required, metaphor titles banned, loop endings preferred, length window enforced.
- The space channel’s rulebook carries structure rules for 25-minute documentaries and an English narration style guide.
- The commerce channel’s rulebook carries legal phrasing requirements for affiliate content.
- Each rulebook is a living document: when an experiment produces a verdict, the verdict gets written into the rules, and every future script inherits it automatically.
That last line is the entire value. A human writers’ room forgets. A rulebook doesn’t.
The receipts
| Item | Number | Source |
|---|---|---|
| Channels scripted this way | 4 | my operation |
| Videos shipped through it | 85+ | channel dashboards |
| Experiment verdicts encoded into rulebooks | hook styles, length window, pacing caps | production database |
| Human review rate | 100% of scripts | my rule, not optional |
What it can’t do, honestly
- It can’t know what happened on my channels unless I feed it the numbers. The rulebooks are powered by my measurement habit. An AI with no data writes generic scripts, confidently.
- It doesn’t replace taste. I reject or rewrite openings routinely. The 100% review rate is not ceremony; unreviewed AI output is exactly the content platforms are learning to bury.
- Cost note: I use a subscription plan rather than paying per script, so my marginal script cost is effectively flat. If you pay per-use, your economics differ; measure before scaling.
What you can use today
- Before your next AI-written script, write ten rules from your own channel’s data (best hooks, banned phrases, target length). Paste them in front of every request. This one habit separates usable output from beige.
- When a video over- or under-performs, write one sentence about why into your rules file. You’re building your rulebook either way; the only question is whether it’s written down.
- Keep the review step. The point is a faster writers’ room, not an unsupervised one.
Here are the receipts.