# Title & hook copywriting — research behind the Gemini prompt Why the prompt writes titles the way it does. Feedback from the streamers whose lives we clip was blunt: the titles were *summaries*. "Trova una moneta rara" describes the clip and gives nobody a reason to comment. What they asked for is the American clip-account register — playful, overstated, argument-starting ("Reda smetterà di fare live dopo aver trovato un cimelio da 10k") — which almost nobody does in Italy yet. This is what the research says that register can and cannot be. ## The policy boundary (this shapes every formula) - YouTube prohibits **malicious clickbait**: "using maliciously misleading titles, thumbnails, descriptions, or imagery to trick users into clicking." Canonical failure: a title promising a full match over a video that only has a clip. Enforcement is warning → strike → termination at 3/90 days. - The operative test is **delivery, not literalness**: click → does the viewer get what was promised? Hyperbole that resolves inside the clip passes; a promised event that never happens fails. - TikTok treats **engagement bait** as an integrity violation that makes the post **ineligible for the For You feed** — "Like this if you agree", "Comment your birth month and I'll…", "Follow me and I'll follow back", "Only 10% can answer this". - Meta demotes vote/comment bait but **explicitly exempts genuine asks for opinion, advice or help**. **Net rule encoded in the prompt:** provoke the comment *implicitly* — a claim people want to correct, a stake they want to judge. Never *mechanically*. The prompt used to REQUIRE the mechanical form ("ALWAYS include a CTA like 'Follow me and comment X and I'll send you the workflow'"); that instruction was shipping every clip into the demoted bucket and has been removed. ## What actually drives comments - **Cunningham's law** — people correct a wrong answer far faster than they answer a question. A debatable valuation or a judgement call outperforms an open question. Constraint: keep it a *judgement*, never a fabricated fact about a person, health, money or news. - **Open loops, identity, stakes** are the three retention drivers; reveal the end state and hide the process ("okay, but *how*?"). - The hook must land in **1–3s** (~45% leave before 3s), and **two** triggers beat one — three reads as spam. - **Prediction comments** are the highest-engagement comment class; specific prompts beat generic ones. - For streamer clips specifically: fast hook, a genuine emotional reaction, and a moment that stands alone without the stream. Titles should be moment-specific and searchable, not "momenti epici". ## What makes a moment clippable (the selection gate) Practitioner consensus from clipping operators (SERVIUOS, ~25k-clipper Discord; Nashifys; Joe Sbiti) — a moment earns a clip only if it hits one of four things, and the rest of the editing is secondary to that judgement: - **Unexpected turn** — plot twist / pattern interrupt; retention comes from the surprise, not the topic. - **Strong emotion or polarization** — the share test is "would this get sent to a friend / dropped in a group chat". A take half the audience wants to argue with beats one everyone agrees with, because arguing is a comment. - **Relatability** — "this is literally me", specific and everyday. - **New or useful info** — save-worthy; the save is itself a ranking signal. Two mechanics ride with it: the hook must land in the first 1–3s (a moment whose best beat is at 0:20 is a bad clip even if the beat is great), and dead air is the single biggest retention killer — the reaction beat after the reveal is *part* of the clip, the silence before it is not. Encoded in `GEMINI_PROMPT_TEMPLATE` as the `## IS THIS MOMENT EVEN WORTH CUTTING?` gate, deliberately placed **before** the rubric: the rubric only ranks moments that already cleared the gate. It pairs with the live monitor's `clip_selection: auto` floor — both exist so a weak segment yields fewer clips instead of padding the quota. Not adopted: those workflows also lean on manual per-clip music beds, sound effects and CapCut styling — out of scope for an automated pipeline. ## Italian register - Always `tu` / `voi`, never `lei`; the informal, near-improvised register is the norm and corporate phrasing reads as an ad. - `voi` is the pronoun that pulls replies ("voi l'avreste venduta?", "ditemi che sbaglio"). - CAPS on one or two words as emphasis, never the whole line; emoji as separators, not decoration. - Use the handle the audience actually uses, never a legal name. Lead with the name only when the streamer is the draw — otherwise the moment goes first. ## Where this lives in the code `pipeline/gemini_request.py` → `GEMINI_PROMPT_TEMPLATE`, section `## TITLE & CAPTION COPY`: the nine rules and the rotating pattern list are a direct encoding of the above. Two guards ride with it: - **Grounding rule** — exaggerate the framing, never invent the event (the delivery test). - **Speaker attribution** — the live monitor passes the channel owner as `CLIPPYME_CREATOR_NAME`, which lets a title name *whose stream this is*, but never puts a quote in their mouth: guests and co-streamers share the mic and the transcript carries no speaker identity.