A channel data prompt is only as good as the rows you feed it. Include too little and the model fills gaps with assumptions. Include everything and the real signal drowns in noise. The skill is choosing the middle.
Most teams err toward dumping every export into the prompt and hoping the model sorts it out. It rarely does. The answer ends up shaped by whichever rows happened to catch the model's attention.
Let the question decide the data. If you want to know why retention drops early, include opening lines, thumbnail text, and first-minute timestamps. View totals from three years ago will not help.
Work backward from the answer you need. Every field in the prompt should have a reason for being there, and you should be able to say what it would change.
That set covers most questions about why a video performed the way it did, and each item maps to something you can change.
A prompt is a small room. Only bring in what the conversation needs.
Cut rows the question does not touch and trim transcripts to the parts that matter. A full hour of speech usually has two or three lines that explain the outcome.
Keep enough context that the model can see the shape of the data. Stripping everything down to a single number removes the pattern along with the noise.
When in doubt, run the prompt with a lean set first and add a field only when the answer clearly needs it.
Feeling your way through a script works until it doesn't. Pattern extraction turns 'this channel feels good' into a repeatable, teachable formula.
One viral video teaches you almost nothing about a channel. Twenty ordinary videos reveal the structure that made it worth studying at all.
A blueprint you admire is useless until it becomes a limit you write inside. Here is how to convert observations into constraints that speed up drafting.
Analyze a channel, extract its structural DNA, and generate scripts engineered from patterns that actually work.
Analyze a Channel