skill-bank

Skill Bank

A collection of reusable AI agent skills by Bowofade Oyerinde / Bonifade Technologies.

Available Skills

course-creator

Plans and builds instructor-led courses one action at a time, for technology or other subjects. A course map contains sections, chapters, lessons, concepts, prerequisites, and a simple-to-complex project ladder. Teachers can reorder it without losing stable lesson IDs.

Installing course-creator includes its twelve focused subskills: curriculum mapping, content authoring, visual assets, Slidev decks, Remotion or HyperFrames lesson video, Kokoro voice narration, marketing copy, motion graphics, product launch, talking-head packaging, avatar video, and storytelling. The parent chooses only the subskills needed for the teacher's request. They can work alone or share the same course map and assets.

Independent actions include lesson writing; notes as PDF or an online page; optional exercises; publishable assignments, quizzes, and rubrics; lesson-by-lesson Slidev decks; optional local Kokoro voice narration; Remotion or HyperFrames videos using supplied text, audio, or teacher footage; captioning or overlay packaging of an existing talking-head clip; AI avatar, talking-photo or dubbed presenter video; and a timecoded CapCut Desktop handoff. Written material, decks, narration, and video can be produced separately or combined. Visual outputs can use the 580 bundled SVG images and icons. Videos use a shared course style and a per-video style.md.

For separate installation or invocation outside the parent, use curriculum-map, content-authoring, visual-assets, slide-decks, explainer-video, motion-graphics-video, voice-narration, marketing-copy, product-launch-video, talking-head-video, or avatar-video. Each accepts a standalone brief or existing files, and each is craft-first: it leads with the artifact it produces and treats course work as one optional branch rather than a requirement. The marketing-copy skill is general purpose — it writes for any product, service or brand — and draws from the LodgeStatus campaign notes for credible hooks and sales copy on skills and courses.

The product-launch-video skill is also included as a subskill in course-creator. It turns supplied screenshots and product information into a motion ad or screenshot-based demo, using genuine screens as evidence. It writes a per-video style.md and delivers editable Remotion or HyperFrames source plus an MP4 when rendering is available.

The talking-head-video skill packages an existing talking-head, interview or podcast clip with captions or designed graphic overlays — the footage itself is never altered. It transcribes locally and renders in HyperFrames. The avatar-video skill generates a presenter from a script — an AI avatar, a talking photo, or a dubbed clip — through the hosted HeyGen provider, and falls back to the script, narration and storyboard when the provider is not authenticated.

The skill teaches each concept through a relatable example, What, Why, Where, When, and How, a visual explanation, and a practical demonstration. Absolute beginner programming courses begin with computer basics. Each chapter has a small project; a larger section project follows at least three small projects.

The canonical course map is course-plan.json. The bundled validator checks prerequisite order, project placement, artifact paths, and table-of-contents synchronization:

python course-creator/scripts/validate_course.py path/to/course-plan.json --sync-toc

See the skill and its linked references for each action's output contract. The original written-course guidance is retained in the text production reference, technical depth reference, term bank, and roadmap notes.

Example prompts:

Map a beginner programming course from computer basics through a first web app. List every section, chapter, lesson, concept, and proposed project; do not write lessons yet.
Write the lesson on file extensions from my course map, then create its Slidev deck.
Make a Remotion lesson video from this deck and my recorded narration; include a CapCut Desktop editing pack.
Create a Remotion launch ad from these product screenshots and the product website; show only real UI states.

Install:

npx skills add oyenet1/skill-bank --skill course-creator

---

motion-graphics-video

Generate kinetic typography, animated charts, process flows, diagrams, logo reveals and visual concept explanations with Remotion or HyperFrames. Delivers an MP4 and editable source, with optional narration and captions. Works alone or as the motion graphics subskill inside course-creator for lesson diagrams and chapter titles, reusing the course style and lesson IDs.

npx skills add oyenet1/skill-bank --skill motion-graphics-video -g

Example: “Create a 20-second motion graphics video explaining the water cycle with animated arrows and labels.” For a course: “Animate this lesson's process diagram using the course style and register the video under the lesson.”

---

🔬 spec-driven-development

Specification-Driven Development (SDD) — transforms vague ideas into structured, implementation-ready specifications.

Takes any rough requirement and produces:

• requirements.md — User stories + testable acceptance criteria (WHEN/SHALL/IF)
• design.md — Architecture, components, data model, API contracts
• tasks.md — Ordered, dependency-aware implementation checklist

Install:

npx skills add oyenet1/skill-bank --skill spec-driven-development

Triggers on: "spec this", "SDD", "plan this feature", "break this down", "generate requirements", "write a spec", and more.

---

Installing Skills

These skills work in a coding agent that can read files and run commands; the desktop app is not required. Install Node.js with npm first so npx is available. The installation commands below work in Linux/macOS terminals and Windows PowerShell or Command Prompt.

Choose a video skill

Skill	Use it for
product-launch-video	Product launch ads and demos with real desktop/mobile screenshots
motion-graphics-video	Kinetic typography, animated diagrams, charts, processes and lesson titles
explainer-video	Explainers, lessons, motion graphics and product walkthroughs
talking-head-video	Captions and graphic overlays on supplied presenter footage
avatar-video	Avatar or talking-photo presenters; provider or hardware requirements apply
voice-narration	Local narration audio
slide-decks	Slide presentations
course-creator	The full course workflow with its bundled subskills

Install one skill globally (available across projects):

npx skills add oyenet1/skill-bank --skill product-launch-video -g
npx skills add oyenet1/skill-bank --skill motion-graphics-video -g
npx skills add oyenet1/skill-bank --skill explainer-video -g
npx skills add oyenet1/skill-bank --skill talking-head-video -g
npx skills add oyenet1/skill-bank --skill avatar-video -g
npx skills add oyenet1/skill-bank --skill voice-narration -g
npx skills add oyenet1/skill-bank --skill slide-decks -g

Run only the command for the skill you need, or install several together:

npx skills add oyenet1/skill-bank --skill product-launch-video --skill explainer-video --skill voice-narration -g

Install the complete course bundle:

npx skills add oyenet1/skill-bank --skill course-creator -g

Omit -g to install into the current project. The installer prompts for the agent to use; add -a codex to target Codex explicitly. To inspect available skills or check your global installation:

npx skills add oyenet1/skill-bank --list
npx skills list -g

Complete installable skill catalog

All fifteen top-level skills, including the video-shortcuts router, can be installed independently. Use npx skills add oyenet1/skill-bank --skill NAME -g -a codex with any name below. Omit -a codex to choose another supported agent.

Skill name	Direct associated skills installed on first use
spec-driven-development	None
curriculum-map	None
content-authoring	None
marketing-copy	None
visual-assets	None
voice-narration	explainer-video, product-launch-video
slide-decks	visual-assets, explainer-video
explainer-video	voice-narration, product-launch-video, slide-decks, visual-assets
motion-graphics-video	voice-narration, visual-assets
product-launch-video	explainer-video, voice-narration
talking-head-video	avatar-video, explainer-video, product-launch-video
avatar-video	talking-head-video, voice-narration
course-creator	All twelve course subskills already bundled; no separate sibling copies
storytelling	slide-decks, visual-assets, voice-narration, explainer-video
video-shortcuts	Router only; install the matching video/narration/storytelling skills alongside it

The associated graph is resolved recursively, including dependencies of dependencies. npx skills add itself copies only the selected skill; its first-use launcher installs the associated graph and prepares runtimes.

Install every top-level skill for Codex in one command:

npx skills add oyenet1/skill-bank --skill '*' -g -a codex -y

The CLI supports many agents, but some do not support global installation. Selecting an agent explicitly avoids unrelated agent-specific installation errors. Older generated YAML descriptions containing unquoted colons caused the CLI to skip motion-graphics-video and avatar-video and report only eleven skills. The generator now serializes valid YAML and checks every target.

Measured installation and LodgeStatus video-generation results are in the installation audit, including failures, recovery, cached checks and workflow coverage.

For a fresh isolated installation audit with per-stage timing, run:

python3 scripts/smoke_install_video_skill.py
# Discovery/copy only, without runtime downloads:
python3 scripts/smoke_install_video_skill.py --files-only

The script writes logs and report.json under a new temporary sandbox. It measures skill copying, first-use setup, cached setup and read-only verification. Installed skills, npm/UV/browser caches and downloaded models stay isolated; compatible host Node/npm, Python, UV and FFmpeg may be reused. Use --source PATH to test a local checkout or --sandbox-dir EMPTY_DIR to retain a chosen location.

Prepare the runtime on first use

npx skills add copies files and has no setup hook. Each runtime skill now instructs the agent to run automatic first-use setup before production. This installs its associated skills, including indirect dependencies, from the bundled verified snapshot. The complete course bundle has its own launcher. For immediate setup from a repository checkout, run python3 scripts/install_video_skill.py explainer-video --target /path/to/skills. For manual setup, open the installed skill directory and use the command for your OS:

OS	Command
Linux/macOS	sh tools/setup.sh --yes
Windows PowerShell	powershell -ExecutionPolicy Bypass -File tools\setup.ps1 --yes
Windows Command Prompt	tools\setup.cmd --yes

Setup downloads the supported private runtimes and associated skills. Use --check instead of --yes for a read-only verification of cached runtimes (it fails if anything required is missing), or add --no-skills to prepare runtimes without installing associated skills. First setup needs internet access and space for tools/models. No desktop application is needed. Windows/macOS end-to-end rendering remains unverified; local avatar generation also needs supported hardware, and hosted providers need credentials.

Example requests after installation:

Use product-launch-video to make a 30-second mobile demo of my app at http://localhost:3000. Capture real mobile screens; ask for private login access if needed.
Use explainer-video to make a 60-second narrated walkthrough using both desktop and mobile screenshots of my website.
Use talking-head-video to caption my recorded presentation.

Which video skill should I install?

Skill	What it does
explainer-video	Explainers, tutorials and walkthroughs
product-launch-video	Product adverts, promos and demos
motion-graphics-video	Animated text, charts, diagrams and logo reveals
talking-head-video	Captions and overlays on existing footage
avatar-video	Generated presenters and talking photos
voice-narration	Speech audio from text

For presentation slides, use the supporting slide-decks skill.

For general explainers, start with explainer-video. For product advertising, start with product-launch-video. You do not need every skill.

For Codex, install a chosen skill with:

npx skills add oyenet1/skill-bank --skill explainer-video -g -a codex

Replace explainer-video with an exact name from the table. Omit -g for installation in the current project. Explicitly invoke it in a prompt, for example: "Use explainer-video to make a video about solar energy. Ask me for any missing important details before starting."

Repository edits do not update previously installed skill copies. To try local changes before they are published, run this from the repository root:

npx skills add . --skill explainer-video -g -a codex

Start a new agent session after replacing an installed skill so its updated instructions are loaded.

Install the six video and narration skills together

For Codex, run one command:

npx skills add oyenet1/skill-bank --skill explainer-video --skill product-launch-video --skill motion-graphics-video --skill talking-head-video --skill avatar-video --skill voice-narration -g -a codex

To use this checkout's latest changes before they are pushed, run the same command from the repository root with . instead of oyenet1/skill-bank. To install every skill in the repository, including non-video skills:

npx skills add oyenet1/skill-bank --skill '*' -g -a codex

-g makes the skills available across projects; omit it for the current project. Start a new Codex session after installation. Skill files install first; required renderer and narration runtimes are prepared on first use. Avatar backends have their own hardware or provider requirements.

Use each video skill

Name the skill in your request and supply any files or URLs it needs:

Skill	Example request
explainer-video	Use explainer-video to make a 60-second beginner explanation of solar panels for YouTube, with English narration.
product-launch-video	Use product-launch-video to make a 30-second Instagram advert for my app at [product URL], using real product screens.
motion-graphics-video	Use motion-graphics-video to animate these three steps in a 20-second portrait video: sign up, choose a plan, book a room.
talking-head-video	Use talking-head-video to add readable captions and key-point overlays to [recording path].
avatar-video	Use avatar-video to make a presenter video from [script path], using [presenter photo path].
voice-narration	Use voice-narration to read [script path] in British English with a male voice and export MP3.

Replace bracketed placeholders with real URLs or file paths. Missing important choices should trigger clarification before production. Add "ask me for missing details before starting" to make that preference explicit, or "choose for me" to delegate choices. You can combine skills in one request, for example: "Use product-launch-video and voice-narration to make a narrated advert for [product URL]. Ask me for missing details before starting."

Shortcuts followed by your prompt

The portable video-shortcuts skill routes short prompts across coding agents:

Shortcut	Production skill	Example
/tutorial	explainer-video	/tutorial explain solar energy to beginners
/product-launch	product-launch-video	/product-launch promote my app at [URL]
/motion-graphics	motion-graphics-video	/motion-graphics animate these three steps
/talking-head	talking-head-video	/talking-head caption [video path]
/avatar	avatar-video	/avatar use [photo] to present [script]
/narration	voice-narration	/narration read [script] in British English
/story	storytelling	/story explain an API using a catchy story; give me a post and a video

Each shortcut has a clarification breakpoint: use details in your prompt or previous answers, ask a batch of questions for missing important inputs, and wait before scripts, storyboards, authoring or rendering. Video questions cover subject/product, purpose, audience, platform/layout, duration and audio; source files, language, voice and CTA are asked where relevant. Complete prompts proceed directly. Say "choose for me" to delegate creative choices.

Install all skills, including the router, for your agent:

npx skills add oyenet1/skill-bank --skill '*' -g

The installer lets you choose the agent. Use . instead of oyenet1/skill-bank from this checkout to try unpublished changes. Restart the agent after installing.

Portable invocation: if your host treats custom / commands as unknown, send Use /tutorial explain solar energy to beginners as ordinary prompt text, or explicitly invoke video-shortcuts followed by /tutorial and your prompt. The router and production skills use Markdown instructions, not a Codex-only API. A skill install does not automatically register commands in every host menu.

Native slash adapters: this repository ships command templates for Claude Code and Codex. Install matching production skills first, then from the repository root:

# Claude Code: /tutorial followed by your prompt
python3 scripts/install_video_shortcuts.py --host claude

# Codex: /prompts:tutorial followed by your prompt
python3 scripts/install_video_shortcuts.py --host codex

All seven shortcuts install together. Restart the agent to load them. Existing conflicting command files are preserved unless you explicitly pass --overwrite. --target PATH installs into a different command directory. Other coding agents can use portable prompt aliases or adapt commands/claude/ to their documented command format; native registration is specific to the host.

Codex's custom prompt mechanism is deprecated in favor of skills, so the portable router is the shared workflow. Claude Code command templates follow its command format.

Social-media-ready output

All audience-facing text and video workflows now share a social media output standard:

• Ask for missing publishing platform, audience, format and length before production.
• Open with a catchy, accurate hook and deliver a complete explanation and payoff.
• Format posts for the selected feed, with clear text and relevant CTAs.
• Keep video objects, Mermaid flows and captions readable on a phone and clear of platform controls.
• Use smooth purposeful motion and verify the actual export, including sound-off comprehension.
• Check current official platform requirements before final export.

The goal is retention and shareability. Virality is not guaranteed, and the skills do not invent trends, metrics, testimonials or performance. Full tutorials still explain the subject; internal source files and course plans retain their purpose. No social post is published automatically by this policy.

Included object images

visual-assets now ships 56 reviewed device/object images from devmock in assets/device-objects/. The course bundle includes the same bank in course-creator/subskills/course-creator-assets/assets/device-objects/. Phones, laptops, tablets, monitors, workstations and a person using a tablet are available as real image files, alongside the existing object/diagram guidance.

The catalog lists each object and its dimensions. The manifest records source filenames, hashes and the screening decision. Stock-watermarked and visibly provider-credited previews are excluded; included files are copied unchanged, with no watermark removal. Some contain screen placeholders or baked-in backgrounds, so inspect the chosen object before composing it.

Storytelling and other visual skills use this bank through visual-assets, which is part of their associated skill graph. Update or reinstall visual-assets to receive the images in an existing installation. Selected images are copied into the output project for portability. They are illustrative mockups, not proof of real product behavior; their reuse rights are recorded separately from this repository's code license.

Storytelling: posts, videos or both

Use /story followed by your topic, or select storytelling within another workflow. If the output is unclear, the agent asks post / video / both and waits for the missing audience, purpose and delivery details. Example prompts:

Use /story explain API requests to beginners as a catchy LinkedIn post.
Use /story explain how water moves through a plant in a 45-second narrated portrait video. Ask for any missing details.
Use /story explain authentication to beginner programmers; give me both a post and a Slidev video. Ask for missing details.

The default storytelling method comes from Paul Smith's Sell with a Story. Select a story for what the audience should understand, feel or do, then follow hook → context → challenge → conflict → resolution → lesson → action. The working framework applies that method to original posts, Slidev scenes and programming explanations. Narrative functions can share a beat; this does not impose seven slides on every story. Stories open with a relatable problem or curiosity hook, explain what happens and why, and finish with a clear payoff. Illustrative stories stay labelled; product claims and real customer results need evidence.

• Post: ready-to-copy text adapted to the selected platform, without video setup.
• Video: editable Slidev scenes, a script, storyboard, portable assets and MP4 when rendering succeeds. Slidev is used whenever storytelling video is selected, unless you explicitly choose another framework.
• Both: one grounded story adapted into complete post text and spoken narration.

Programming explanations use Mermaid for control flow, request/data movement, interactions and state changes, linked to concrete inputs, outputs and code. Other stories show the relevant objects, parts or substances and their processes, with labelled flows rather than text-only scenes. Assets show what narration actually describes. Motion follows entrance → readable hold/action → exit, with consistent easing, smooth transitions and highlights timed to the explanation.

Slidev's Mermaid integration and animation primitives support this authoring. The existing static PNG export helper does not preserve live animation; the storytelling workflow requires live Slidev capture and checks the exported MP4. If capture is unavailable, it reports the blocked render and preserves the editable sources instead of claiming a static montage has smooth animation.

To add the storytelling skill and shortcut router to an existing installation:

npx skills add oyenet1/skill-bank --skill storytelling --skill video-shortcuts -g

Choose your coding agent when prompted. Installing every skill with --skill '*' also includes storytelling. After updating, reinstall the native command adapters if you use them, then restart the agent to load /story (Claude Code) or /prompts:story (Codex). Other agents can use Use /story ... as ordinary text with the portable router.

Automatic product screenshots

product-launch-video and explainer-video can capture real public or logged-in web product screens using their bundled tools/capture_product_screens.py. Choose desktop, mobile, or both; mobile capture uses responsive browser device emulation. Credentials come from private environment variables or an existing session outside exported assets. The capture tool prepares its own private Playwright/Chromium runtime and outputs PNGs with provenance. See the capture plan and login guide. Native mobile apps require supplied captures or a separate emulator tool.

When a prompt leaves important details unclear, the skill asks one batch of questions and waits for your answers before production. It uses recommended defaults only when you explicitly ask it to choose for you.

Editing Skills

For the current video production steps, installer gaps, subtitle contract, and the Tauri app integration plan, see video desktop integration audit.

Kokoro narration bootstraps its own uv, Python 3.12 environment, packages and verified model files on first use via tools/kokoro/start.sh. Slide decks and video skills ship tools/ensure_video_runtime.py; their first-use preflight checks and installs Node, Slidev, Remotion, HyperFrames, Chromium, FFmpeg and FFprobe as needed. For example, from an installed product-launch-video skill directory, run python3 tools/ensure_video_runtime.py product-launch-video to see the verified executable paths. --check reports status without installing. npx skills add copies files and does not run a post-install hook, so setup runs when the skill is first used.

First-use bootstrap on any OS

Every runtime-backed skill ships a small cross-platform bootstrap that detects the operating system and CPU architecture, prepares the private runtime, and offers to install the associated sibling skills that are not present. Run the launcher for the host:

OS	Command
Linux, macOS	sh tools/setup.sh
Windows (PowerShell)	powershell -ExecutionPolicy Bypass -File tools\setup.ps1
Windows (Command Prompt)	tools\setup.cmd

The launcher runs tools/bootstrap.py, which reads the shipped tools/dependencies.json and:

1. installs missing Node, FFmpeg/FFprobe, the selected renderers and Chromium into the private user data directory;
2. prepares Kokoro narration and local speech recognition where the skill needs them;
3. installs missing associated skills from its bundled, hash-verified source snapshot into the exact requested directory. For example, explainer-video installs voice-narration, product-launch-video, slide-decks and visual-assets. No npx or Git is needed for these sibling installations; existing skills and user edits are preserved.

Dependency order, parallel setup, and reuse

Setup respects prerequisites while overlapping independent work. The desktop app's Prepare tools action resolves Node once, then runs renderer profiles and FFmpeg/FFprobe setup concurrently, with at most four workers. Each renderer installs its packages before preparing its browser. Duplicate profile names are handled once. The app joins all active workers before reporting success or failure; only a completely verified setup reports ready.

Standalone setup and scripts/install_video_skill.py use the same bootstrap. It runs up to four independent branches concurrently: video tools, speech recognition, Kokoro narration, and associated skill copies. Renderer routes within the video-tools branch remain sequential because they share Node and media directories. Sibling copies remain sequential because their cyclic dependency graphs overlap. Associated skill files are copied recursively; their own runtimes are prepared when those skills are first used.

The dependency layers, from prerequisites to consumers, are:

1. Host and download utilities: Windows/macOS/Linux on x64 or ARM64, writable user data, internet for first downloads, and enough disk space. Launchers reuse Python 3.10+ or prepare private Python. Without Python, Unix setup uses curl/wget, archive extraction and SHA-256 utilities; Windows uses PowerShell. Linux browser system libraries must already be available; setup does not invoke a distribution package manager.
2. Runtime managers: Node.js 22.12.0 or newer with npm for rendering; private UV and Python 3.12 environments for narration and transcription. Compatible existing runtimes are reused.
3. Media tools: FFmpeg and FFprobe. Standalone setup uses working system binaries or ffmpeg-static and @derhuerst/ffprobe-static; the desktop app uses its existing native media installer and platform archives.
4. Renderers and their browsers: Slidev uses @slidev/cli, @slidev/theme-default, playwright-chromium and Chromium; Remotion uses @remotion/cli, remotion, react, react-dom and Chrome Headless Shell; HyperFrames uses hyperframes, gsap and its managed browser. Each engine also installs its npm dependencies recursively.
5. Narration: Kokoro uses kokoro-onnx, soundfile, espeakng-loader, misaki and imageio-ffmpeg. Its locked indirect dependencies include addict, attrs, cffi, cloudpickle, dlinfo, flatbuffers, joblib, numpy, onnxruntime, packaging, phonemizer, protobuf, pycparser, regex and typing-extensions. The two model files, kokoro-v1.0.onnx and voices-v1.0.bin, total approximately 354 MB.
6. Transcription: faster-whisper, ctranslate2, PyAV (av), their recursively resolved Python dependencies, and the small.en speech model. small is the multilingual option; older macOS has an alternative PyAV pin.
7. Skills: visual-assets has no associated skills. explainer-video, product-launch-video, voice-narration and slide-decks form a connected group that also installs visual-assets. motion-graphics-video brings in that group through its narration/asset associations. talking-head-video and avatar-video install each other and that group. Cycles are visited once and existing skills are preserved. course-creator already contains all twelve course subskills, so it does not install separate sibling copies.
8. Optional features: Local avatars require compatible accelerator hardware, backend-specific Python/PyTorch packages and model downloads; ordinary setup does not install these models. Hosted avatars require HeyGen credentials. Generated scene images require an OpenAI credential and quota. Docker and Pocket TTS are optional alternatives, not default bootstrap dependencies.

Exact direct versions live in runtime_requirements.json. The complete Kokoro package versions/hashes live in requirements.lock, and avatar package/model requirements live in avatar_models.json. Renderer and transcription transitive versions are resolved at installation; there is no single checked-in lock covering all platforms and features.

Optional local avatar installations, after a compatible hardware probe:

1. MuseTalk 1.5 (CUDA): Python 3.10, PyTorch/torchvision/torchaudio, NumPy, OpenCV, soundfile, einops, OmegaConf, ffmpeg-python, imageio, imageio-ffmpeg, Pillow, tqdm, diffusers, accelerate, transformers, huggingface-hub, librosa, MoviePy, mmengine, mmcv, mmdet, mmpose and xtcocotools. Source builds can additionally need antlr4-python3-runtime, chumpy and its build dependencies. Models: MuseTalk UNet, SD VAE, Whisper, DWPose, face parsing, ResNet18 and S3FD; about 4.5 GB of model downloads.
2. MuseTalk (Apple MPS): Python 3.10 and the common MuseTalk packages above, with mediapipe in place of the CUDA OpenMMLab stack; about 4.0 GB of models.
3. LatentSync 1.5 (CUDA): Python and the pinned PyTorch stack, accelerate, av, decord, diffusers, einops, face-alignment, ffmpeg-python, huggingface-hub, imageio, imageio-ffmpeg, kornia, librosa, lpips, matplotlib, mediapipe, NumPy, OmegaConf, OpenCV, Pillow, python-speech-features, safetensors, scenedetect, soundfile, tqdm and transformers. Models include LatentSync, SyncNet, Whisper, face detection/alignment, I3D, VGG16, ViT, KonIQ and SD VAE; about 10.2 GB of model downloads. See the avatar manifest for exact pins, source builds, accelerator memory and per-file hashes.

To see dependencies of dependencies for an installed renderer, run npm --prefix /path/to/video-runtime/remotion ls --all --json (replace remotion with slidev or hyperframes). Each profile also retains its resolved package-lock.json. For a private Python environment, run uv pip freeze --python /path/to/environment/bin/python; on Windows use Scripts/python.exe. These show the versions actually installed on your platform, including indirect packages. A fresh full setup needs disk space for caches, extracted browsers, Python environments and models; use a sandbox directory on disk rather than a quota-limited temporary filesystem.

Repeated setup verifies and reuses ready packages, browsers, model files and environments rather than downloading them again. Missing or failed components are retried. The standalone bootstrap retries a failed component once while preserving completed components; a second failure remains an explicit setup failure. Python package downloads use a 120-second read timeout and at most four concurrent UV downloads per branch, honoring any user-supplied UV_HTTP_TIMEOUT or UV_CONCURRENT_DOWNLOADS values. These settings avoid UV's short default read timeout and limit contention during parallel setup (UV settings). Browser readiness checks inspect the actual browser, not a stale marker file. The installer also reuses an existing skill directory containing SKILL.md; it preserves user edits and refuses incomplete directories or symlink destinations. --resume remains available after a failed install. --check stays read-only. Cancellation stops active bootstrap child process trees; native renderer workers share the app's cancel flag. The native FFmpeg download keeps its existing cancellation granularity: setup waits for that download to return before finishing cancellation.

Use --check to report the plan and install nothing, --yes to run without an interactive prompt, --no-skills to prepare runtimes only, and --target DIR to name the skills directory. Unsupported OS/architecture combinations report the missing runtime rather than selecting a binary for another platform. No global npm/pip install is ever performed and PATH is never modified.

The explainer and product launch skills ship tools/assemble_video.py for joining rendered scene files with narration or footage audio and timed subtitles. See the scene assembly contract. The explainer skill also includes tools/render_slidev_video.py for exporting Slidev click states into that assembly flow. Those skills include tools/render_remotion_video.py for packaging an editable Remotion composition with its audio and captions. They also include tools/render_hyperframes_video.py for editable HTML compositions. All video skills ship tools/transcribe_captions.py to create word-timed caption files from final spoken audio, with recognition marked for review. The four video skills also ship tools/run_video_job.py for a single request with renderer setup, progress events, verified output, subtitles, cancellation, and retry. See the desktop job contract.

To prepare dependencies at install time from a clone of this repo, use the bundled installer and choose the skills directory for your agent:

python3 scripts/install_video_skill.py explainer-video --target ~/.agents/skills
python3 scripts/install_video_skill.py product-launch-video --target ~/.agents/skills

On Windows Command Prompt, run the same command with py -3 and a Windows skills directory, for example py -3 scripts/install_video_skill.py explainer-video --target %USERPROFILE%\.agents\skills. The installer selects managed Node archives for Windows, macOS, or Linux (x64 and arm64) and stores Kokoro's Python, packages, and models in the user data directory. Linux setup uses the same path for Debian, Arch, Fedora, RPM packages, and AppImage; it does not invoke apt, pacman, or dnf. Unsupported CPU/OS combinations report the missing runtime instead of silently choosing a binary for another platform.

This copies the generated skill, installs its video runtime, and prepares local Kokoro narration and speech recognition where required. It also accepts motion-graphics-video, slide-decks, talking-head-video, avatar-video, voice-narration, and course-creator. Existing valid skill directories are reused without overwriting their files. If setup fails after the files are copied, rerun the same command with --resume to retry the missing prerequisites. For avatar-video, it checks local capabilities without downloading models and lets the provider dispatcher prefer a verified local backend or the editable script, voice and storyboard fallback when no hosted provider is available.

install_video_skill.py invokes the same parallel bootstrap as first-use setup, including indirect associated skill copies. Installations made through npx skills add or a manual copy complete that setup on first use.

Skills are generated. The source of truth is skills-src/; the course-creator/subskills/ folders and the top-level skills are build output. Do not edit generated folders by hand — the next build overwrites them.

Path	What it holds
skills-src/manifest.yaml	which capabilities exist, their names, and which shared modules each ships
skills-src/<capability>/skill.md	one capability — frontmatter for both targets, shared craft, {{mode:…}} blocks for target-specific text
skills-src/<capability>/bootstrap	the capability's bootstrap: block also emits tools/bootstrap.py, the setup.sh/.cmd/.ps1 launchers, and tools/dependencies.json
skills-src/_shared/	the intake modules: intake, brand, video, voice, preflight, objects
skills-src/_shared/prompts/<category>/	script and storyboard patterns, one file per pattern
skills-src/_shared/library/<name>/	full-text source material shipped verbatim to the capabilities that declare library:
skills-src/<capability>/files/	extra files a target ships (e.g. the product launch workflow)

One capability emits two install targets: the course-creator bundle subskill (course-creator-<id>) and the standalone skill (<id>). Tokens in a skill.md resolve per target — {{ref:brand}} links the shared module, {{doc:narration.md}} links a course-creator reference, {{sibling:voice-narration}} links another capability, and {{asset:tools/kokoro/README.md}} resolves a shipped path at the right depth.

# regenerate every target (requires pyyaml)
python scripts/gen_skills.py

# verify committed output matches the source — run before committing
python scripts/gen_skills.py --check

# behaviour tests
cd scripts && python -m unittest

The tool trees follow the same rule: course-creator/tools/kokoro/ is canonical and voice-narration/tools/kokoro/ is generated from it, so the standalone install works alone.

Intake

Every producing skill runs the same intake protocol before it makes anything. See skills-src/_shared/intake.md: present inputs are used, values determined by supplied context are recorded with a reason, and missing inputs that would change the deliverable are asked in one batched block with options and a recommended pick. Production waits for your answers; there is no automatic question timeout. Say "choose for me" to delegate choices. Brand values and evidence are never invented.

Brand identity lives in a brand.md that skills read and write — colour, type, tone, vertical, claims and accessibility all inherit from it, and each per-video style.md records only its deltas. A local project can be read with python scripts/extract_brand.py <path>; secrets and dotfiles are never read.

License

MIT — see LICENSE.