Shortify

Shortify — long-form → short-form (produce → caption → publish)

Take ONE long-form video and cut it into 3 angle-distinct vertical shorts, each with its own hook and b-roll, then write per-platform captions. The screen recording of the real product/result is the STAR; generated b-roll is a spice used only for conceptual openers.

Personalization

This skill works in your brand voice. Before running, load the brand profile at ~/.claude/brand-profile.md.

• If it doesn't exist, run the personalize skill first (or just tell me your name, niche, offer, and audience — I'll create it).
• If a field this skill needs is blank, I'll ask one or two quick questions and save the answers back to the profile so you only answer once.
• Never invent facts or results about the user — use only what's in the profile, or ask.

Fields this skill uses: Tone + Words-to-avoid (script voice), Proof/wins (the on-screen proof stat — never invent one), Offer + Default-CTA (the end-card + caption CTA), Main-platforms (where to publish).

Requirements

This skill orchestrates several media tools. It degrades gracefully — if one isn't available, I'll tell you exactly what to add or hand you the files to finish manually.

• ffmpeg — trimming, compositing, audio extraction (required).
• Whisper (local, no API key) — karaoke captions from the audio + finding caption-safe trigger words. Whisper mishears some product names (e.g. "Claude" → "cloud"/"quad"); correct these in the captions.
• A talking-head avatar tool (e.g. HeyGen) — optional; only if you want a synthetic presenter bottom-frame instead of your own footage.
• An image/video generator (e.g. Higgsfield, or Gemini image gen) — optional; for the conceptual cold-open and a few canonical b-roll scenes.
• A social scheduler (your platform's native scheduler, or any third-party scheduling tool) — optional; only needed for the auto-publish step. Without one, this skill still produces the finished cuts + captions and you upload them yourself. Never hardcode account IDs or keys — read them from your environment.

Never embed API keys. Read any keys the tools need from environment variables or a local .env.

Inputs (from the argument / message, else ask)

• transcript: the long-form timestamped .txt
• screencap: the long-form screen recording .mp4 (1080p ideal)
• angles (default proof, contrarian, transformation)
• slug: a short name for the output folder

Steps

1. SCRIPTS FIRST — review gate. Write just the 3 scripts (cheap, no render) and show the user to approve/edit BEFORE spending any avatar/render budget. For each angle check:

• The spoken hook matches a proven pattern (swap only the variables — tool, number, outcome — never invent a new hook shape).
• On-screen hook_title is 3-4 words (hard max 4 words per line; break longer titles onto a second line so the card never spans the full frame).
• Distinct, angle-specific openings; a concrete number in the first sentence; result before mechanism.
• Name the real tool/feature, not a generic "AI" or "a new mode".

Fix wording before rendering.

1. B-ROLL plan (the make-or-break step — build carefully and VERIFY). Per angle, align visuals beat-by-beat:

• The screen recording of the real result is the STAR — show it EARLY (right after the cold open) and cut through its panels as the script names them.
• Reserve generated art for the conceptual opener and a few shared canonical scenes (generate each once, reuse across angles via a cache).
• A LIST line ("watch every line / approve every step / catch every mistake") gets one rapid-fire word card per item — a WORD card (≤4 words), never a full sentence.
• Subagents can draft the b-roll plan, but do NOT trust it blindly — sanity-check that it's screencap-dominant with no sentence text-cards, and verify the actual frames.

1. RENDER each cut. Composite: presenter (your footage or an avatar) bottom, screen recording top, karaoke captions throughout, cold open, and a WIDE result payoff at the end. Trim silences/pauses so the presenter is present on frame one.

1. CAPTIONS — per platform. Write angle-distinct titles + first lines for each cut (never reuse the same caption across angles). Adapt the soft CTA by video type using the brand profile: tutorial → "Learn to build like this…"; showcase → "Get [your offer]…". Respect any platform hashtag limits (e.g. Instagram caps meaningful hashtags — keep it tight).

1. PUBLISH (optional). If a scheduler is configured, upload each cut and schedule the 3 angles into open future slots (stagger them several days apart, rotating post times). Publish to the user's Main-platforms. If no scheduler is set, output the finished files + captions and tell the user to upload them.

1. Verify + auto-repair. Extract frames (hook ~2s, the conceptual opener, a mid beat, the wide ending) from each cut and show the user. If any on-screen b-roll trigger didn't fire, transcribe that cut's audio with Whisper to see what was actually said, pick a plain adjacent word as the trigger (never a number or a dotted filename like CLAUDE.md, which Whisper writes as "claude .md"), and re-render just that cut.

Rules baked into the method (do not override)

• FORMATS (3 distinct shapes, never homogenized): proof = RESULT-LED (result + number, then proof line, what, how, stakes); contrarian = BELIEF-FLIP (the wrong belief FIRST, then the result as the mic-drop proof after it); transformation = BEFORE/AFTER (old-world cost, then new world + result).
• HOOKS (the most important part): always instantiated from a proven hook pattern — keep the shape, swap only the variables; distinct per angle; end on an intrigue-building payoff description; name the real tool/feature, never "AI". A contrarian hook may be a pattern-interrupt ("STOP DOING X"). On-screen hook_title: 3-4 words ideal, hard max 4 words per line.
• OPENINGS: angle-specific and distinct (proof = the result; contrarian = the belief it breaks; transformation = the shift); the FIRST SENTENCE carries the headline number; one PROOF stat line after the payoff (use a real number from the brand profile — never invent one); a conceptual generated b-roll opens contrarian/transformation before the screen recording.
• SCRIPTS: obey the brand-profile words-to-avoid; name the model + feature specifically; ANSWER WHY (the pain it removes), not just what; result before mechanism; natural thought-line pauses; no spoken hard CTA.
• B-ROLL: per-beat source (screen recording / infographic / illustration / presenter / library clip / word card); the REAL result is the STAR, shown EARLY and carrying most beats; a LIST line gets one rapid-fire word card per item (≤4 words — a sentence is NEVER a word card); shared canonical scenes generated once; end on a WIDE result payoff; never a plain-editor/notes frame when the line isn't about that; motion variety (no camera move more than twice in a row; pan infographics, zoom illustrations).
• CAPTIONS / TRANSCRIPTION: auto-correct product names Whisper mishears in the karaoke captions.
• CTA: use the brand-profile Offer + Default-CTA. Adapt by video type: tutorial → "Learn to build like this…"; showcase → "Get [offer]…". Prefer "link in bio" with the real link in the description. Per-angle end-card text is allowed.

Output

An output folder per run containing, per angle: the script, the b-roll plan, the rendered cut, and the caption set — plus a short summary of where each cut will publish (or a note to upload manually).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: duncan-buildroom
• Source: duncan-buildroom/profit-room-skills
• License: MIT

Install and usage instructions live in the source repository linked above.