Input: a source-video path and an optional reference-reel path.

Deliver the requested formats and the editable project. Default to 1080x1920; when requested, also compose 1920x1080 with its own framing and text placement. Verify each aspect ratio independently, including its opening and CTA. The edit must tell the speaker's story through images, timing, and sound. Virality is an aspiration, not a result that editing can guarantee. Work locally unless delegation is requested.

The first creative gate is: "In the first 3 seconds is someone convinced they need to watch until the end?" Establish a specific reason to stay and earn its payoff inside the video. Effects, asset volume, and production polish come after this test. Treat the answer as editorial judgment until audience data exists.

Read before planning. It holds Nate's measured reel style (pacing, layouts, captions, B-roll order, sound palette), the raw-reel, long-to-Short, and narrated-explainer variants, and the lessons that cost time. Use it as the default when the user gives no style; the user's brief wins.

When the user gives a note on a draft, fix the edit, then offer to write the note into the relevant skill or reference as a rule, check, or script change, so the next edit starts from it. Record the note and the resulting rule together.

User authorization persists. An instruction to make a finished video and iterate authorizes ordinary local editing and rendering. An explicit request for Kie B-roll authorizes those generations. Resolve review gates through the current authorization; ask only for missing approval, not again at every draft. Never post the reel or contact anyone without explicit authorization.

Write and a beat plan before HTML. Identify the hook, open loop, escalation, evidence, payoff, and CTA. Give every shot a reason to exist.

Read curiosity-and-entertainment.md. Create with each honest viewer question, planted object, setup time, reminders or partial answers, closure time, and the spoken evidence that earns the closure. Default to one main loop and at most one secondary loop open at once. Close the promised lesson before the CTA; a comment request must not be the only answer. A recurring object alone is not an open loop: establish what is missing or unresolved, then visibly change that state.

Build three materially different rough openings using the same truthful source speech. Compare the first three seconds for topic clarity, curiosity, muted comprehension, and follow-through. Choose and record the reason; do not select only the loudest effect. Render a low-resolution moving animatic with speech and rough sound before polishing materials. Fix predictability, missing context, or weak payoffs at this stage. Existing approved materials can be reused, but still review the new sequence before spending time on final rendering.

Vary rhythm by sequence: a short cluster of discoveries, a brief anticipation, then a payoff. Preserve word timing while varying animation speed and scale. The 1-2 second event guideline is not a metronome or permission to add filler. Allow the speaker's expression to carry a conviction beat. For every shot record the viewer's expectation and the new information or consequence it supplies.

Audit information duplication. Narration and captions carry the exact words; graphics should usually show the example, relationship, evidence, or consequence. Remove a third restatement as a headline unless deliberate emphasis earns it. Prefer a specific product action or source detail to generic typing footage. Do not invent a product result simply to make a more satisfying payoff.

For each opening record the viewer benefit, unresolved question, visible evidence by second three, and the exact payoff scene. Reject vague teasing and any promise that this footage cannot fulfill. An advertisement may resolve a concrete example or explain a mechanism before inviting the viewer to learn more; it must not promise a complete tutorial and deliver only a registration CTA.

Make the first frame immediately meaningful. A surprising action, contradiction, result glimpse, or revealing human expression can establish the tension. Choose an impact or transition only when it sharpens that idea. A quiet impossible-looking result can beat a loud montage. Avoid a mandatory stamp, flash, or 0.3-second effect.

Use a supplied sound reference when available. Remove recording lead-in, align the effect's audible impact with the visual landing within two frames, and shape its level and tail around the first words. Make the hit noticeable while keeping speech clear; briefly dip the music if helpful. Do not make loudness the only hook.

The opening still has to work muted. Establish the topic or tension immediately, then reveal another meaningful state within the first three seconds. Avoid an unrelated logo bumper, added silence, repeated flashes, or an effect that delays the argument. Match the chosen material style. Vary the device between reels; do not automatically stamp STOP on every video.

Save a first-three-seconds shot/sound plan and inspect the encoded opening frame by frame, including its first frame, impact, reveal, and return to the main edit. Check that the first spoken word survives the impact mix. Record this as a distinct opening-hook gate in the verification report.

For a minimal, Apple-style brief, read premium-motion-and-footage.md. Use restrained type, negative space, real object depth, directional light, and soft contact shadows. Minimal composition can still have rich, purposeful motion and sound. Do not mistake a small object on a blank slide for premium design.

Give the eye one clear attention target. Use depth, occlusion, light, and camera movement when they make the action easier to follow. Keep typography readable and use dimensional treatment only when the chosen material world calls for it. For a paper brief, inspect fibers, edges, folds, contact shadows, and perspective in the encoded output. These are conditional craft checks, not universal style rules.

Give every extended graphic beat an action with an outcome: a key seats and a project starts, criteria turn into checks, a card opens to reveal a result, or a stack assembles into a system. A card such as Define good can swivel or tilt so its edge and moving contact shadow remain visible as its criteria resolve. Avoid settle-and-freeze title cards. Subtle secondary motion supports the main action; identical idle bobbing applied to every object is not storytelling.

Use the supplied human-edited reference to study how attention moves: wide context, isolated detail, interaction, visible consequence, then return to the speaker. Before delivery ask what makes each sequence entertaining beyond looking professional. A new heading or a new caption alone is not the answer.

Audit named people, products, and places in the transcript before storyboarding. Use a verified portrait when recognizing a real person matters, rather than defaulting to their name in text. Prefer their own site or an attributable source, save provenance, and inspect the actual image. Animate the photograph as a physical object when appropriate, with a meaningful callback. Never generate a likeness and present it as an authentic photo. A moving portrait remains a photographic layer and does not count as moving B-roll.

Review adjacent shots as silhouettes and actions, with headings hidden. If they still look like the same card reveal, redesign the weaker beat. Vary mechanisms: tear to reveal, unfold to show scale, press to create, connect to enable, process to produce a result. Preserve a coherent world and understandable cause and effect. Spend the strongest action on the most important spoken idea. A user's engagement score is feedback for iteration, not a metric a validator can certify.

For product stories, seek real source material: public product pages, official documentation, authentic screen recordings, and official logos. Record actual scrolls, selections, or visible interactions when useful. Present real logos on dimensional surfaces without redrawing an approximate mark or implying endorsement. Retain source URLs and distinguish authentic product footage from illustrative interfaces and generated B-roll. Never expose real credentials in captured UI. An animated source screenshot may be one layer in an explanatory composition, but it must not substitute for moving B-roll when the brief requires video.

Never invent, approximate, trace from memory, or generate an existing brand's logo. A radial star is not a substitute for the Claude mark. Use the exact user-supplied asset first; otherwise obtain the asset from the brand's official site or media kit. Search engines help locate sources, but do not establish authenticity. Save the source URL or supplied path, original file, and checksum. Inspect the artwork before use, preserve its shape and proportions, and animate the authentic asset or the surface carrying it. Do not alter a logo's geometry to create depth. Audit every logo occurrence in the final encoded reel against its source asset, including small marks and recurring motifs. If no verified asset is available, use plain brand-name text rather than fabricate a mark.

Choose the mix from the story and available evidence. There is no default graphics/B-roll percentage. Keep a duration ledger with authentic footage, generated illustration, designed graphics, and speaker-only time; explain the denominator and count split panels once. Captions over footage remain footage. A still image with a camera move is a photographic layer, not moving B-roll.

Use each distinct B-roll scene only once per reel. Another crop, filename, speed, reverse, or grade does not make it a new scene. Reusing the same selected scene across the two requested aspect ratios is expected. Record source scene identity, file hash, and selected interval in ; run the footage validator below. A graphic callback must develop the original idea into a new state and composition. It is not permission to replay footage.

Default to containing the full source frame when a tighter crop hides context. Keep the subject, action, and consequence visible throughout the selected interval, with captions outside that action. Inspect beginning, middle, and end in each ratio at phone size. Dense demo text must not be required to understand a fleeting insert; isolate a meaningful visible action or give it more time. Avoid duplicated, blurred background video as an automatic solution for vertical framing.

Prefer a specific action with a readable consequence. Compare adjacent shots with headings hidden: if the sequence repeats the same visual mechanism or could belong in any AI advertisement, replace the weakest insert. Generated metaphors can add surprise but cannot stand in for captured evidence. Prioritize authentic human reactions where they support the story.

Use quality-gates.md. The plan validator catches structural defects but cannot certify taste, motion, or the audio listening pass:

node .agents/skills/short-form-edit/scripts/validate-plan.mjs video-projects/<slug>
node .agents/skills/short-form-edit/scripts/validate-footage.mjs video-projects/<slug>

It reads , , and . See plan-schema.md. For another implementation, export that schema or adapt the validator explicitly; never silently skip checks because the schema differs.

Inspect encoded hero frames for every shot at full size and phone scale. Inspect contiguous strips across every cut and transition, plus the complete motion where supported. Check speech joins, fresh ASR from final audio, voice/bed balance, SFX timing, clipping, and the last syllable. Distinguish listening from signal analysis. Record any unavailable review modality honestly.

Run a separate uninterrupted entertainment review of the animatic and final where real-time video review is supported. Record attention-drop timestamps, predictable sequences, unresolved promises, and reading overload in . Fix the weakest sequences, then recheck them in context. Automated playback completion and contact sheets are technical/visual evidence, not a claim of uninterrupted perceptual viewing. If that modality is unavailable, use chronological sequence strips and explicit loop-state checks, and disclose the limit. Never manufacture retention scores or call internal hook comparisons an audience A/B test.

Fix findings, archive superseded versions, and repeat the checks affected by each change. Give the actual output hash, measured results, inspected artifacts, revisions, and unresolved limits. Do not claim passes based on code inspection, file existence, or an empty validator result.

Return a playable final MP4, its absolute path, and the skill path. Keep the composition, assets, source and retimed transcripts, EDL, shot/event plan, audio stems, reference analysis, and QA evidence in the project. Clearly identify the final version. Never label a draft final or claim guaranteed viral performance.