AI short-video and Reels workflows

The original Gemini HTML lists 10 GitHub projects; their table order and A/B/C production scenarios are retained here. Capabilities, setup paths and limits follow corrections against official README, skill-file and code snapshots fetched on 2026-11-04. The original HTML is saved separately for comparison.

Entry numbers follow the original report's order; they do not rank quality or popularity.

Melvynx account baselines and style files

/outliers defaults to filtering outliers at 3 times a competitor account's own median views, or 2 times for your own account, excluding sponsored videos; it marks query windows with fewer than 10 videos as insufficient and does not rank them. Results go to video-agent/outliers/, with 5 adapted scripts generated by default in video-agent/scripts/. These research filters do not establish that a particular editing method caused view growth.

/steal examines cuts, layouts, captions, sound effects and volume in a reference video, then creates video-agent/styles/<name>/style.json and STYLE.md. Generate a test video with /short-editing <take> style <name>, compare it with the reference and decide whether to adopt the style. Styles must fit the engine's supported layouts; inputs should include videos you have rights to use, your own footage and visual elements to keep or exclude.

Instagram research skill thresholds

The general viral-outliers threshold is at least 700,000 views and at least 5 times the account's follower count. For accounts above 500,000 followers, the skill instead uses 5 times the account's own median views. These are the project's methods, not official Instagram recommendation rules.

Research cards should retain the source URL, retrieval time, view count and follower count, checked against the original video. The skill requires reviewing at least 6 frames, increasing to 10 to 12 for fast cuts, and verifying the opening transcript. A first line labeled [ 0.0s] does not by itself prove the speech has no missing words; listen to the original opening.

raydenai Mode B and Mode C

Mode B in the viral-video-creator skill describes frame extraction, OCR, Whisper, vision-model analysis and prompt assembly. Mode C describes variants that hold the subject fixed while changing the opening or other elements. Check which capabilities the program actually provides before running it; steps in the skill may require separately prepared tools.

The current lab/replicate.py extracts frames at 3 fps, sends at most 12 frames to Claude Vision and writes stories/replications/<source-id>/shotlist.yaml for review and revision. It does not implement OCR or Whisper transcription. Extract and verify opening speech and on-screen text separately when needed; sampled frames cannot establish the full script.

Melvynx's default edit removes pauses longer than 0.3 seconds, uses 1.15 times speed and targets -14 LUFS for voice. These are default style settings. Preview before adopting them when instructional steps, accents or music requirements differ.

Run these documented commands in your own production project after reading the corresponding skill, installer and dependency versions. This review did not execute these installation or generation steps.

Creator editing on a local machine

Install the skills in the project folder, then run /studio-setup in Claude Code. Required tools include Node 20+, Git and ffmpeg. Download/research workflows also require yt-dlp and Deno; initial setup prepares Chrome/Chromium and whisper.cpp.

Place recorded footage in video-agent/takes/ and run /short-editing. Inspect video-agent/edits/short-<slug>/output/short-<slug>.mp4 and the SRT; Instagram/TikTok research and automatic publishing require a separate Treg setup.

Instagram research cards

Prepare Agent-Reach, OpenCLI, ffmpeg, faster-whisper and Chrome signed into Instagram and connected to the OpenCLI extension. Edit .claude/skills/viral-outliers/references/brand-profile.md in the installed copy with your topic, audience, tone and filming constraints. opencli doctor should show that the extension is connected.

After a research run, check that the Markdown cards include an accessible original video, the complete opening transcript and your adapted script. This skill is limited to reading; it does not follow, like or publish for you.

Higgsfield experiment lab

Install the Higgsfield CLI separately and sign into your own account; visual analysis and pattern extraction also use an Anthropic API key, while reference analysis needs yt-dlp and ffmpeg/ffprobe. --dry-run checks price quotes and should not produce a paid video; before rendering, check the model, duration, quality and credits. Lab output goes to gallery/; scheduled publishing is outside the current scope.

Remotion template preview

These commands render the repository's included demo without substituting your product. The template is configured for 1920×1080 at 30 fps; adjust aspect ratio and layout for Reels instead of assuming a vertical or 60 fps default. Confirm the demo renders before replacing product assets and text according to the shot cards.

Seedance prompts

These commands copy only the opening and faceless-content skills; restart Claude Code before using them. Supply the video duration, topic, product, aspect ratio and visual constraints, then check the output prompt's timing and asset requirements before handing it to Higgsfield manually. Test one segment first to check the shots and visual content.

MoneyPrinterTurbo Agent helper

Prepare uv and an Agent with terminal, file and long-running execution support. The official skill's helper is at docs/skill/mpt_agent.py; run from that directory:

Read docs/skill/SKILL.md first; this entry point supports macOS and Windows and defaults to a Chinese vertical video, Pexels assets and Edge TTS. Supply missing credentials as indicated by the output; when a paid video provider requests cost confirmation, confirm before rerunning. Success output includes MPT_RESULT, VIDEO_FILE and TASK_DIR; open the MP4 to check captions, voiceover and whether the assets fit the topic.

The proposed combination is raydenai/viral-video-creation and rediumvex/ai-video-generator-claude, with MoneyPrinterTurbo added when needed. Hand off files or text at each step, checking formats on both sides first.

1. Choose a reference video you have rights to use and record its opening, shot purposes, text and sound effects. When using rayden's analysis tools or Mode B guidance, inspect the actual artifacts and separately perform transcription and on-screen text checks not completed by the tools.

2. Adapt the verified structure to your topic and use rediumvex to produce prompts. Retain timing and shot intent while replacing assets, dialogue and brand elements; generate in Higgsfield manually and select usable clips.

3. For voiceover and captions, prepare readable assets and scripts for the selected MoneyPrinterTurbo workflow. Check how local assets are supplied before composing a short video; convert analysis JSON from other projects yourself.

The proposed combination is the Instagram research skill, Melvynx skills and vyralcontent/content-skills. Inputs include your brand information and recorded footage; outputs are a filmable script and an edited video.

1. Organize references using either the Instagram research skill's thresholds or Melvynx /outliers account-median filter. Record the two thresholds separately with query times to avoid mixing figures from different datasets.

2. Extract reusable narrative structures, then use viral-hooks and viral-instagram-reels to check openings, dialogue and the test plan. Sends Per Reach and Trial Reels are strategies discussed in the skills; available metrics and test entry points depend on the account interface.

3. Film the adapted script yourself. If using /steal, check the style files and test video first, then process your raw footage with /short-editing; verify proper names, caption cuts and the CTA word by word.

Use JohnChow90/video-shotcraft shot cards and its Remotion workflow for product motion, and worldwonderer/drama-skills for drama. Evaluate HKUDS/ViMax when an executable generation framework is needed. Produce and review the two deliverable types separately.

Product and SaaS demos

1. Prepare actual product screens, the Logo, feature descriptions and video duration, then choose a presentation approach from the shot cards. Choose either the original list's JohnChow90 version or the Vincentwei1021 source linked by its README; avoid mixing revisions.

2. Have the Agent build a Remotion project or adjust an example from the selected cards, then inspect aspect ratio, text and product operation. The original source also mentions remotion-dev/skills; it is not included among this report's 10 verified projects.

3. Render an MP4 after the preview passes. Check UI text, buttons and cropping frame by frame, and confirm commercial use complies with the separate Remotion and asset licenses.

Narrative drama and character continuity

ViMax requires Python 3.12+ and uv; the README lists Linux/Windows environments. The Web UI also needs Node 18+. Create private configs/agent.local.yaml from configs/agent.example.yaml, then enter models, endpoints and credentials for the selected LLM, image and video services.

1. In drama-skills, create the script, character/scene/prop assets and storyboards, specifying appearances and reference images to preserve. Check the relationship between keyframes and motion prompts before production.

2. Create a separate project when using ViMax and organize the approved story, reference images and requirements into accepted inputs. Configure LLM, image and video providers and generate a small number of shots first; this report has not verified an automatic import interface between the two projects.

3. Compare faces, clothing, props, spatial positions and dialogue in every shot. If appearance drifts, replace reference images, regenerate or cut the shot before assembly.

The listed MIT projects, video-shotcraft's Apache-2.0 license and dependencies each have conditions. No license declaration was found in the reviewed raydenai/viral-video-creation snapshot. Public readability does not grant reuse rights; ask the maintainer before commercial adoption.