Verified against Claude · 2026-07-26
Diagnose why a specific video's retention graph drops where it does
Reads a described audience-retention curve alongside the actual script or transcript to name the likely cause of a specific drop, distinguishing a real content problem from a normal, harmless dip — instead of assuming every dip needs fixing.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are diagnosing a specific drop in a YouTube video's audience retention graph by cross-referencing the described curve against the actual script or transcript for that section. You are not giving generic retention advice — you are explaining what most likely caused this specific drop, at this specific timestamp, in this specific video, and whether it's actually worth fixing. RETENTION GRAPH DESCRIPTION Retention holds around 68% until 3:20, then drops sharply to 41% between 3:20 and 3:35, then flattens and declines only gradually for the rest of the video SCRIPT OR TRANSCRIPT FOR THE FLAGGED SECTION [3:05] So that's the basic setup. [3:22] Quick word from today's sponsor, NordVPN — [3:58] okay, back to the actual build... VIDEO CONTEXT A 12-minute DIY furniture build video; this channel's videos in this format usually average 55-60% overall retention WHAT ELSE IS KNOWN Three separate comments on this video mention skipping "the ad part" — no comments mention the actual build content negatively DIAGNOSTIC RULES First classify the drop by its shape, since different shapes mean different things: a sharp single-frame or single-second cliff usually means a specific moment — a jump cut, a sponsor segment starting, a scene change — actively pushed viewers to leave, and the transcript or edit at that exact timestamp should show a distinct trigger. A gradual decline over 20-60 seconds usually means the content itself lost the viewer's interest across that whole stretch, not at one instant — look for a slow setup, a repeated point, or a tangent spanning that range rather than hunting for one single line to blame. A drop that recovers shortly after — viewers leaving and then the curve flattening at a lower but stable level — usually means a specific segment (an ad read, an aside) filtered out viewers who weren't interested in it specifically, which is a different problem than a drop that keeps sloping downward and never stabilizes, which suggests the whole video lost the plot from that point on. Before recommending any fix, check whether the drop is actually large relative to the rest of the video's own curve — a small, ordinary dip that's in line with normal per-minute attrition elsewhere in the same video is not a problem to solve, and treating every wiggle in a retention graph as a crisis produces fixes for things that were never broken. Where the transcript shows a candidate cause, state your confidence honestly — a retention graph on its own can show where viewers left but cannot prove why, so name the mechanism you suspect and flag it as an inference, and note what would need to be true to confirm it, such as a comment thread mentioning the exact same section or a similar drop recurring across multiple videos at the same relative point. OUTPUT FORMAT 1. Drop classification (cliff, gradual decline, filtering dip, or ordinary variation) with the reasoning. 2. Most likely cause, cited to the specific line, cut, or segment in the provided transcript, with a stated confidence level. 3. Whether this drop is actually worth acting on, given its size relative to the rest of the video's curve. 4. If worth fixing, one concrete, specific change to that section — not generic advice like "make it more engaging."
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Classifying the drop by shape before assigning a cause reflects a real distinction in what a retention curve can actually tell you: a sharp, near-vertical drop concentrated in a few seconds is consistent with a specific triggering moment because viewers don't gradually decide to leave in under two seconds, they react to something — a jump cut, an ad read starting, a tone shift — whereas a slow decline spread over a full minute cannot be pinned to one instant because no single second explains a decay that's spread across sixty of them; treating both shapes with the same generic "find the boring part" instruction throws away information the graph itself is actually encoding. Separating a filtering dip that recovers and stabilizes from a decline that never recovers matters because they imply opposite responses — a stable plateau after a drop usually means a specific segment, like a sponsor read, cleanly lost viewers who weren't interested in it while the rest of the audience stayed, which is often an acceptable and even expected cost of that segment, while a curve that keeps sloping downward past the flagged point means whatever caused the initial drop kept bleeding viewers afterward, which is the pattern that actually warrants a structural fix to the video, not just the one segment. Requiring a check against the video's own baseline before recommending any fix prevents the common overcorrection of treating every visible wiggle in a retention graph as a crisis: some attrition every minute is completely normal, described in YouTube's own creator guidance as an expected pattern, and a dip that's proportionally similar to normal per-minute drop-off elsewhere in the same video is statistical noise, not a signal, so spending editing effort "fixing" it treats a normal curve shape as a bug. Explicitly requiring the model to state its confidence and name what would confirm the inference is the most important honesty constraint here: a retention graph shows precisely where in time viewers left but says nothing directly about why, so any causal story drawn from a transcript alone is an informed guess, and corroborating evidence like a recurring comment thread or the same relative drop appearing across multiple videos is what would actually elevate that guess to something closer to a confirmed diagnosis rather than a plausible-sounding story matched after the fact to a shape in a graph.
Verified against
Claude Sonnet 4.6 · 2026-07-26
ChatGPT GPT-5.1 · 2026-07-31
Changelog
- 2026-07-26 — Initial publish, verified against Claude (Sonnet 4.6) and ChatGPT (GPT-5.1).
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
