Prompting AI to Audit a Channel: A Working Method

An effective channel audit prompt gives the model a defined dataset, a specific set of questions, and a format for the answer. Vague instructions like 'analyze my channel' produce paragraphs that read well and tell you nothing you can act on. The model does not know your channel at all.

What the model knows is patterns in the text you paste and whatever it absorbed during training. Your job is to decide which material it sees and what you want back. Everything else in a good audit follows from those two decisions.

Give the model a dataset, not a vibe

Paste the actual material. Titles, view counts, upload dates, retention notes, top comments, and a few transcripts will do. The model can only reason about what sits in the prompt or its training data, and its training data stops well before your last fifty uploads.

Label the columns and keep the units honest. A table that mixes total views with average views in one column pushes the model toward answers that sound confident and are quietly wrong. Five minutes cleaning the input saves an hour untangling nonsense later.

Ask for findings, not summaries

A summary restates what you pasted. A finding names a pattern and points to the evidence behind it. Ask for the second one and require the model to cite which rows support each claim.

Force specificity by asking for a small number of observations, say three, ranked by how much they matter. When the model has to choose, the filler drops away and the real signal has room to show.

Constraints that push past the obvious

Constraints do more work than encouragement. Telling the model what to ignore removes the safe, generic observations it reaches for first when it has no direction.

• Name the comparison: 'contrast my last ten long videos with my last ten shorts on retention, not on raw views.'
• Set a boundary: 'do not mention upload frequency unless a row shows it actually moved performance.'
• Request the counter-case: 'list one pattern that would disprove your main finding.'
• Ask for a next step: 'give me one test I can run next month to check the top finding.'

Ask a model to summarize and you get a mirror. Ask it to choose and defend, and you start to get analysis.

An output shape worth reusing

State the role, the data, the questions, and the output format in that order. A prompt that opens with 'You are a retention analyst' and then hands over a table gives the model a frame before it reads a single row.

End with the output contract. 'Return five bullets, each with a claim, the rows that support it, and a confidence level of low, medium, or high.' The contract is what makes each answer checkable rather than merely plausible.

Key takeaways

Run the same prompt after your next upload and compare the results. An audit earns its keep when you repeat it, not when you run it once.

• Paste real data and label it clearly, because the model cannot see your channel.
• Ask for ranked findings with evidence instead of broad summaries.
• Use constraints and counter-cases to strip out generic filler.
• Always request a fixed output format so you can verify each claim.