A pipeline skill, sanitized

Sizzle-reel clip curator

This is the skill for turning a library of AI-generated scenes into the moments a trailer, a sizzle reel or a recap needs. It is mostly editorial judgment written as rules: what counts as a moment, how a set is balanced, what footage is allowed as a source, and what is inferred when a scene the story implies was never shot. The creator approves a proposal before anything is rendered. Project, scene and character names, paths and hosting steps are removed. Download it as SKILL.md.

---
name: sizzle-clip-curator
description: Curate and produce short, self-contained clips (5 seconds or less, 3 lines or less) from a production's existing library of generated footage for sizzle reels, trailers and recaps, or generate them for moments the story implies but never showed. Covers the curation rules (one mood tag each from a small taxonomy, balanced across tags, cast spread, finished footage only, continuity pulled from the footage itself), where footage and word-level transcripts live, the proposal the creator approves before anything is rendered, and the three production paths. Invoke when asked for sizzle-reel moments, a trailer cut list, recap clips, or a batch of promotional moments.
---

Sizzle-reel clip curator

A sanitized copy of one of the Claude Code skills from the AI production pipeline on stolenduck.com. The job: pull the best three-to-five-second moments out of a library of generated scenes, balance them into a set, and produce the few that need re-cutting or re-rendering, for a sizzle reel, a trailer or a recap. Project names, paths, scene and character names, hosting details and delivery steps are removed; the protagonist is "the protagonist," the principal characters are "the leads."

What a clip is

A self-contained moment: 5.0 seconds or less (aim for 3 to 4), 3 lines or less (a line is one character's turn), tagged with ONE mood from a small taxonomy of three or four tags (the examples below use Sweet, Fun and Drama). The clip plays start to end with nothing over it.

Workflow

A. Inventory   read the scene scripts and the clip transcripts; shortlist two to three times the target count
B. Proposal    write the proposal (format below) and summarise it to the creator. STOP.
C. Approval    the creator edits and approves rows. Nothing is rendered before this.
D. Production  per row: re-cut, re-render, or generate; polish in the edit; place the masters
E. Delivery    captions and whatever the consumer needs to find the clip

The proposal is the deliverable of a curation request. Do not skip to generation because a row "obviously" works; the creator re-tags and re-ranks rows every time.

Curation rules

• Length. 5.0 seconds hard, 3 to 4 typical. A moment that needs six seconds is two moments or a re-cut; trim pauses before trimming words. Check the transcript's word timings, not the clip length: speech often starts several seconds in.
• Lines. Three turns at most. Two beats land harder than three; a single killer line is fine.
• One tag each. Define each tag in a sentence (Sweet: tender, warm. Fun: comedy, banter. Drama: conflict, reveal, stakes). When a beat is two things, tag what the image sells and put the other in the notes.
• Balance. Equal counts per tag across a proposal. Spread the cast: no character above about 30% of rows, and protagonist-led and ensemble rows count toward spread, not toward a character.
• Finished footage only. A clip plays start to end with no interface, caption or overlay in or over it. That rules out as sources the idle loops and held frames that sit under other screens. Burned-in captions are an edit-time decision for the export, not part of the asset.
• Wardrobe and location continuity. If the scene (or the scene a clip is inferred from) already exists as footage, every character wears exactly what they wear in that footage and the place looks exactly like the footage's place. Pull both references from the delivered clip itself; do not re-describe an outfit or a set from the script when a frame of the real thing exists.
• Current footage only. A source clip must be one the production actually uses today. A resolver script answers "is this clip in use, and where." Libraries keep superseded renders beside current ones (an earlier pipeline's shots next to the current ones, with an earlier version of the protagonist's face), and a clip that merely exists on disk is not a source. Never cut from a clip the check reports as unused without saying so on the row.
• The canonical face. The protagonist is defined by one close-up and one full-body reference, and older footage can carry a previous face. Before re-cutting any shot with the protagonist in frame, compare their face to the canonical close-up; a shot with the old face cannot ship. Regenerate that shot with the current anchor or cut around the protagonist, and record the check on the row.
• Leads first. Clips should feature a lead. A clip with none in frame (a supporting character, a protagonist-only beat) is the lowest tier regardless of how good it is, and there should be few per batch.
• Re-render budget. One third-person re-render per clip per batch unless the creator raises it; list the runners-up under "Deferred re-renders" so the next batch starts from them.
• Inferred scenes are fair game when the script states or implies they happened: the ride home a line mentions, the briefing before a reveal, the toast after the event. Keep them consistent with what the audience later sees, and never invent an event that a later script contradicts.
• Marketing safety. Sizzle clips travel outside the product: no lines that only work with context the viewer lacks.

Where the footage is, and the transcript dump

Every clip has a word-level transcript sidecar (words with in and out times, grouped by speaker turn). Before shortlisting, dump every sidecar in a folder as one line per sentence with in and out seconds; those are the trim points for a re-cut. Speech-recognition drift is normal; match on meaning. An empty word list means the clip has no dialogue, not that the sidecar is missing.

The proposal

One file per batch. Header: target count, tag balance, render count. Then one row per clip, grouped by tag:

### 07 — Two Words                    LEAD-B · FUN · re-render
Source   <folder>/<clip> [in–out] + <clip> [in–out] (<script> #<label>)
Lines    LEAD-B: <line>  /  PROTAGONIST: <line>  /  LEAD-B: <line>
Clip     LeadB_TwoWords   ×1
Staging  Third person, handheld two-shot at the rack; he holds it out, she eyes it, he does not lower it.
Wardrobe LEAD-B from <clip> at <t>; PROTAGONIST: MISSING LOOK
Location <reference file> or <clip> at <t>
Why      The whole dynamic in five seconds: immovable, protective, funny.

Fields: Source (clip and in–out seconds, or generate plus the script line it is inferred from), Lines (final dialogue), Clip (name and render count; the name is the lead plus a two-word hook, never a take letter or shot number), Staging (third-person framing, required for re-renders and generates), Wardrobe and Location (which existing clip or reference each comes from, or MISSING), Why (one sentence). A proposal with any MISSING item also carries a "Missing looks and locations" section that asks, per item, generate without input or generate with input. Close with Alternates (runners-up by tag) and Open decisions.

Production paths

Resolution: 720p for anything rendered. The delivery pipeline transcodes to 720×1280 anyway; higher renders only cost credits and time. Re-cuts keep the master's resolution.

Re-cut existing footage. A script takes the approved list and, from the transcript sidecar, merges adjacent sentences, pads about a quarter second at the head and 0.4 at the tail, hard-cuts non-adjacent sentences on a neutral frame (never a crossfade), and re-encodes rather than stream-copies so cuts land on the exact frame. Drop pads before dropping words, and drop a sentence before splitting one.

Re-render a first-person scene in third person. The first-person format keeps the protagonist's body, voice and face off camera; the re-render puts all three back. Build the staging from the first-person stage direction (camera verbs become the protagonist's blocking) and give them the line the scene implies. Keep the lead's existing performance as the reference, same wardrobe, set, light and line reading as the first-person clip, passing that clip as a video reference where the generator allows. One setup, no cuts inside a five-second clip. A physical beat is action: give the generator the action and one line, not three.

Generate an inferred scene. Anchor it to the script line that implies it, the location reference, and the wardrobe the characters wear in the adjacent scene. Write three lines at most in each character's established register. Then keyframe and video per the scene pipeline, the OTS skill for two-shots and the blocking pipeline for three or more people.

Outfit and location references from the footage itself. The outfit reference for any re-render or inferred scene is a frame grabbed from the delivered clip of the scene it continues: one per character per look, full body where the footage allows, plus a close-up if the detail matters, named by scene, character and look and saved as the canonical reference so every later render pulls the same file. Pick the frame from a wide or mid shot with the character facing camera; entrances and establishing shots are best. Locations work the same way. Look beyond the main repository and the hand-off folders before calling anything missing.

Missing looks and locations: infer first. When no footage and no reference shows a character's outfit for a scene, do not quietly invent one. Mark the row MISSING, state in the proposal what the character would most plausibly be wearing and why, and ask the creator, per item, to generate that inference or to give input first. Infer from, in order: what the script says or implies (the location, the weather, the activity, what the character expects to do there, the lead's reaction lines, props added on site); what the production has already established for that character (its wardrobe categories); adjacent footage, if the story makes it the same day; and palette and silhouette discipline (do not compete with the lead's established look or the set, and keep it marketing-safe). Write the inference as one line: category, specific pieces, why.

QA checklist per clip

• 5.0 seconds or less, three turns or less, one tag
• Plays start to end with no interface or overlay in or over it; the source is not an idle loop or held frame
• Every source clip reports as in use, or the row says why not
• Re-render count within the batch budget
• Named by lead and hook, never by take or shot number
• Wardrobe and location match the existing footage; the frame-grab references are saved and listed on the row
• No MISSING item left unresolved; each generated reference saved with the creator's chosen mode recorded
• A re-render matches the first-person clip's lead wardrobe, set and light; the protagonist's look is pinned and recorded
• Every shot with the protagonist in frame shows the canonical face
• An inferred scene contradicts nothing in later scripts
• The transcript reads correctly; the clip plays from the console
• Marketing-safe

Lessons

• Identity drifts when the face is small. A mid-air leap shrank the protagonist's face and the keyframe came back with the wrong person despite both anchors. Stage the keyframe on the beat before the action with the face large and to lens; let the video do the action. Say "the exact face: same eyes, brows, nose, lips and jaw," and pass the close-up anchor first.
• A solo close-up with an off-screen partner reads as "talking to whom?" Frame an over-the-shoulder from the partner instead.
• Group frames drop people. State the count ("EXACTLY FIVE PEOPLE, all visible") and number them left to right; add a close-up anchor for whoever was dropped.
• Bare skin trips the image safety filter even with cover-up wording; a shoulders-up framing passed first try and played funnier. Reframe before switching models.
• Looks must be named in words, not only by reference. One look came back wearing another until the prompt said "coral, pink and mint floral crop top, NOT leopard."
• A transient server error is rerun with the same prompt, not a changed one. Five-second generations land a tenth over; the finalizer caps at 5.0.
• The wrong-face scare was a wrong-script problem. The delivered clips and the masters matched; the re-cut had been taken from a retired draft script whose shots were still deployed. Hence the in-use-only rule and the resolver.
• Contact sheets beat guessing. A tiled frame grid per source clip is the cheapest way to pick wardrobe frames and QA seams.
• Two renders of one master can be transcribed with different sentence splits, and the recognizer can reorder lines. Always list both before cutting by index; prefer text selections to indices.
• A library that reads as slot-filling (a character-by-tag grid filled with whatever existed) is the thing to avoid. Curate the moment first, then balance tags and characters across the set.
• A five-second clip from an entrance whose speech starts at 6.6 seconds is the reaction, not the walk.