DrFIRASS/faceless-reels

Faceless reels

This skill reproduces a production chain that was validated on a real reel ("This voice is not me", shipped November 2026). The goal: a 1080 x 1920, 30 fps reel of 60 to 120 s that holds attention to the end, because everything transforms in place, in rhythm with the voice.

Settings

Personal settings live in config.json at the root of this skill. If it does not exist, copy config.example.json to config.json and ask the user to fill in what is missing (at least elevenlabs_voice_id). Read it at the start of every run. <skill> below means the folder of this skill.

Key	Use
creator_name	How to refer to the user in reports
language	Default language of the reel (voice, captions, on-screen text). A request in another language wins
elevenlabs_voice_id, elevenlabs_model	The cloned voice
whisper_model	Local whisper.cpp model used for checks
projects_root	Where previous reel projects live (to reuse node_modules)
cta	The exact wording of the three closing moves, in the reel language
guides_page	The page linked in the user's bio, where the guides live
timezone, publish_hour, caps, telegram_env	Autonomous mode only

User rules

• Write plainly and directly. Never use em dashes or filler phrases, on screen or in messages.
• Typography of the reel language on screen. For French: non-breaking spaces before : ? ! and inside "5 réglages", guillemets « ».
• The voice is never cut. Only the animation is edited. No duration changes, no silence removed.
• No face in the video, except on the cover, which uses the user's photo.
• The cover is always the first frame of the video, and it is also delivered as a separate file.
• The ending always makes the three moves: subscribe to get the guide, click the link in bio, type the keyword. In the voice, in the captions and on screen (SUBSCRIBED button, LINK IN BIO button, a field where the keyword gets typed). The keyword is directly related to the guide and the topic, and Claude proposes it. Say "link in bio", never "link in the description": the guides page is in the user's bio.
• Everything said must be true: no invented numbers, no unsourced superlatives, no faked screenshots. A lived proof beats a promise.
• No token, password or key readable on screen.
• Caption of each reel: 5 hashtags at most, no links, on every network. It always ends by showing how to get the guide, for example: "The five prompts are in the guide, ready to copy and paste. To get it: subscribe, go to the link in bio and type the keyword FIVE."
• Faceless reel with a cloned voice: declare AI-generated content when publishing (for example with Blotato: isAiGenerated: true on TikTok, containsSyntheticMedia: true on YouTube).
• Music at 0.05 under the voice, never more.
• Announce the cost of paid generations. If the user already authorized Atlas Cloud or ElevenLabs in the request, announce and go on. Otherwise wait for their agreement.
• Go all the way without stopping, check your own work, say clearly what was verified and what was not. Never publish unless the user asks, except in autonomous mode.
• Small machines (8 GB of RAM): one Remotion bundle at a time, renders in the background, intermediate files cleaned at the end.

Steps

Plan everything at the start. Each step has its guide: read it when you reach the step.

1. Locate. Video folder, reference reel (file or link), topic, the user's cover image (cover.png in the video folder, else <skill>/assets/cover/cover-template.png and say so), free disk space (df -h /). Read the useful references/ and the project memory.
2. Analyze the reference, if there is one: extract the audio, transcribe it with Scribe, look at the video frame by frame, write _reel/reference/reel_analysis.md. See references/script.md.
3. Write the script and the promised guide: _reel/script.md, with the truth of each claim. See references/script.md.
4. Create the project: zsh <skill>/scripts/new_reel.sh "<video folder>", then fill in remotion/reel.json. See references/editing.md.
5. Free animatic: python3 <skill>/scripts/fake_words.py, then python3 scripts/build_reel.py, which checks reel.json, then a few stills. Does the script fit on screen?
6. Voice: 2 ElevenLabs takes, Scribe word timestamps on the chosen take, processing without cuts. See references/voice.md.
7. Atlas Cloud opening shot (optional, recommended): start image, end image, Kling, retiming on the voice. See references/opening-shot.md.
8. Scenes: src/reel/BRIEF.md beat by beat, then one agent per group of scenes in parallel and one checking agent. See references/editing.md and references/scene-rules.md.
9. Cover: click-worthy title and logos. See references/cover.md.
10. Check with stills, fix, then render: zsh scripts/render_final.sh in the background (10 to 15 min for 90 s).
11. Verify the final file and write exports/instagram.md. See references/checks.md.
12. Finish: project README, project memory, intermediate files deleted, files sent, summary with what is left to listen to or decide.

Autonomous mode

If the request comes from a scheduled task, or if the user asks to process the inspiration/ folder, follow references/autonomous-mode.md: no questions, cost caps, guide, scheduled posts, notification.

What makes the reel work

• Hook in under 3 s, with a proof or a promise, readable from the very first frame after the cover.
• A visible open loop: N numbered empty boxes, which become the progress bar.
• N elements of the same shape, growing in difficulty and value, the best one last, announced by a re-hook.
• One single shot: same background, same window. Never a hard cut.
• Every appearance lands on the spoken word, never more than 1.5 s without a change on screen.
• Lightning recap, then a punchline that changes how the viewer sees the topic.
• An ending made of gestures, shown and not only said: the cursor subscribes, clicks the link in bio, then the keyword gets typed.
• Loop: the last sentence leads back to the first one, and the last frame becomes the first one again.

Guides

Read	When
references/script.md	Analyze a reference, write the script, choose the keyword and the guide
references/voice.md	Generate the voice, get the words, process the sound without cuts
references/opening-shot.md	Prompts and retiming of the opening shot, costs
references/editing.md	Project layout, reel.json, agents, check stills, render
references/scene-rules.md	Rules to give to the agents who write the scenes
references/cover.md	Title, logos, formats, placing it as the first frame
references/checks.md	Checks of the final file, caption, what is left to a human
references/autonomous-mode.md	Process a file from the inspiration/ folder alone, from the reference to the scheduled posts

assets/reel-template/ is the Remotion project. assets/example-this-voice/ holds the whole validated reel: script, analysis, real scenes, brief. It is the quality reference: read its README.md, then at least Stakes.tsx and Setting2.tsx before writing scenes. assets/example-6-tools/ gives reusable scenes for a ranking or a tiered list.

Lessons already paid for

• With two takes, Scribe transcribed the second one: read source.url in the result and keep the take it actually transcribed.
• Scribe's words.json contains the tags ([softly]) and ":" as words: remove them, and keep the {"words": [...]} shape in scribe_words.json.
• A sentence before "Subscribe" left the window empty: CtaGuide.tsx now shows "Save" during that time.
• The chime and the sparkles covered the keyword and the punchline (voice at -1.9 dB on the keyword): volumes are lowered in the template. Still look at the tightest passages.
• Whisper drifts and makes mistakes: timings come from Scribe, text comes from the script (build_reel.py aligns the two).
• A Whisper prompt containing "Claude" made it read "clone" as "Claude". Check without a prompt, and with timestamps: -nt skips sentences.
• A sound effect on a key word of the hook masks the voice: place it after the end of the word (gap of at least 7 dB, measured).
• A scene that stops exactly at the end of its section cuts the final stamp: scenes overlap (STAGES in Reel.tsx), and a stamp is placed from the start of the sentence.
• An empty window while a sentence announces a list looks broken: place dotted slots from the start of the sentence.
• Numbers written as words break caption pages: build_reel.py keeps "5 settings" together.
• GPU rendering makes elements disappear one frame out of three: always GL=swangle CONCURRENCY=1 (already in render_final.sh).
• No 3D transform on a video, no mixBlendMode, unique React keys.
• chrome-headless-shell hangs: CHROME_MODE=chrome-for-testing (already in the scripts).
• Redoing a sentence after the render: keep the validated voice, generate the ending in one piece and splice it in a pause (references/voice.md).
• On the cover, a brand logo reduced to a symbol can be misread: the two bars of ElevenLabs looked like a pause button, so the name stays written under it.