How do I turn a video (or topic) + a headshot into 3-5 on-brand, high-CTR YouTube thumbnails the way a top agency would: study the channel's winning style, pick the sharpest hook, generate a cinematic background, cut out the real person, and composite crisp bold text + badges as true layers

the agent that answers this

Raw image to viral thumbnail. Turn a video (or topic) + a headshot into 3-5 on-brand, high-CTR YouTube thumbnails the way a top agency would: study the channel's winning style, pick the sharpest hook, generate a cinematic background, cut out the real person, and composite crisp bold text + badges as true layers.

• Step 1tool

  Preflight: ensure tooling is ready. Free CLIs (auto-install if missing): Remotion (npx), ffmpeg, rembg via `uvx --from "rembg[cpu,cli]" rembg ...` (downloads u2net model first run; the [cli] extra alone fails, use [cpu,cli]). Paid: Runway ML API key (dev.runwayml.com). NOTE: gen4_image promptText must be condensed (~500 chars) or it 400s; Impact font resolves on macOS Chromium for the text layer. If a paid key is missing, stop and tell the user what to get + the cheaper alternative.

  Your model fills this step

  • No integration? npm i remotion
  • No integration? uvx rembg[cpu,cli]
• Step 2decision

  Confirm the inputs: the source video (or topic) this thumbnail is for, the subject headshot to composite, and 2-3 of the channel's best existing thumbnails to match the house style.

  Decision step
• Step 3

  Study the channel's existing thumbnails + this video's transcript/topic. Extract the repeatable viral pattern (bold split/contrast layout, ONE huge number-or-word hook, the face cut-out, SOLD/check/cross badges, arrows, white+yellow heavy Impact text) AND the single sharpest curiosity-gap hook for THIS video.

  youtube-video-analyst — an installable skill for AI agents, published by shipshitdev/library.

  Forensic deconstruction of YouTube videos to extract viral formulas, hooks, and retention mechanics. Analyzes video transcripts across 11 systematic sections: hook architecture, structural blueprint, retention mechanics, emotional…

  Full skill: youtube-video-analyst
• Step 4decision

  Generate 3-5 distinct thumbnail concepts (each = layout + a 3-6 word hook + accent color), scored for click-through: curiosity gap, contrast/legibility at small size, emotional face, on-brand. Pick the top concepts to produce.

  Decision step
• Step 5decision

  PROMPT AGENT: for each chosen concept, write (a) the background image prompt — photoreal cinematic scene, NO text, NO faces, brand motif (e.g. compass), a clean/darker band reserved for the face composite, condensed to ~500 chars; and (b) the exact text + badge/arrow spec (strings, fonts, colors, positions, shadows).

  Decision step
• Step 6tool

  Generate each background image via the Runway image API (gen4_image, 1920x1080, no text/no face) and download locally.

  Your model fills this step

  • No integration? Nano Banana / gemini_2.5_flash via Runway
  • No integration? gen4_image_turbo with a brand reference image
  • No integration? hand-built gradient/stock background
• Step 7tool

  Cut the subject headshot to a clean transparent PNG (rembg) and cache it for reuse on future thumbnails.

  Asset preprocessing for HyperFrames compositions — text-to-speech narration (Kokoro), audio/video transcription (Whisper), and background removal for transpa...

  --- name: hyperframes-media description: Asset preprocessing for HyperFrames compositions — text-to-speech narration (Kokoro), audio/video transcription (Whisper), and background removal for transparent overlays (u2net). Use when…

  Full skill: Hyperframes Media

  • No integration? rembg via uvx --from "rembg[cpu,cli]"
  • No integration? remove.bg / Photoroom
  • No integration? ffmpeg chromakey on a solid backdrop
• Step 8tool

  FACE TOUCH-UP — always produce BOTH treatments and carry both forward: (a) FREE ffmpeg color-grade on the cut-out (contrast + warmth + saturation + sharpen, preserve alpha) — identity-perfect, $0; (b) Nano Banana (gemini_2.5_flash) image-to-image relight using the headshot as a reference image — studio relight + warm rim/edge light + clean skin retouch, KEEP the exact same face/identity/outfit — then rembg the relit result. The graded face is the safe default; the relit face often reads better/more 'pro' on YouTube.

  Your model fills this step

  • No integration? if Runway is unavailable or the relit likeness drifts, ship the free-graded face only
• Step 9tool

  Composite in Remotion (1280x720 still): background + face cut-out + bold Impact text (white + yellow, heavy drop shadow) + badges/arrows as crisp layers. Render BOTH a graded-face and a relit-face version of each chosen concept.

  Your model fills this step

  • No integration? Pillow/Canvas or HTML-to-screenshot if Remotion is unavailable
• Step 10decision

  Quality gate: critique each thumbnail as a domain expert, a skeptic, and the end user — legible at small size, on-brand, real likeness preserved (especially the relit face), text crisp. Fix clear problems in place.

  Decision step
• Step 11decision

  Present the variants side by side — each hook in BOTH the free-graded and the AI-relit face treatment — for the creator to pick and tweak (hook/colors/face pose/badges) before publishing. No auto-publish.

  Decision step

common questions

1. v4Aug 20generated

   reconciled active outcome capability input authority

2. v3Aug 20generated

   added immutable outcome-search identity

3. v2Jun 14manual

   added 1 step; rebound step 1 (skills.sh/firebase-basics -> gap), step 9 (decision -> gap)

4. v1Jun 14generated

   auto-generated from "Turn a video (or topic) + a headshot into 3-5 on-brand, high-CTR YouTube thumbna"

Agents are alive: every change is a version, and a run can propose improvements that get reviewed and applied.

related agents

• creator

  How do I read a video script, identify each shot's spoken content, and generate specific generic stock B-roll direction per shot — constrained for a HeyGen AI twin talking head that cannot move or act, with twin filling lower frame and B-roll playing in the top half or as background

  B-Roll Director for AI Twin Videos

• on-demandcreator

  How do I generate cinematic B-roll clips from a source video and a shot-direction brief using Runway?

  cinematic b-roll generator

• creator

  How do I an execution-ready visual production bundle that routes every transcript beat to exactly one of six production lanes and provides synchronized manifests for the Runway Cinematic Shot Generator and Visual Evidence & Remotion Compositor. Planning-only: it never sources, generates, downloads, composes, or renders assets

  Visual Treatment Planner: Runway + Remotion Production Bundle

• dailycreator

  How do I every day, run a multi-persona debate over the day's research to pick the best reel angle, then produce the reel brief and the IG caption, ready for your avatar render

  Daily IG reel: debate and brief