cstack

These are the skills that I use almost every day. I am always adding more skills here. The skills here have saved me hours on on workflows. lmk your experiences with them and drop any feedback! These skills are also refined quite often.

Each skill lives in its own directory under skills/ and can be installed separately.

video-script writes the script and stops. video-edit makes edits of camera footage; animated-video makes films drawn in code. The deliverable decides which of the two owns the final timeline: an edit with animated shots is video-edit's, and animated-video builds the shots to its brief; an animated film is animated-video's, even with one camera insert. asset-library is shared by both.

Every one works alone. Without asset-library, video-edit and animated-video keep a manifest in the project instead of a shared library, and nothing else changes. Install all four to get reuse across projects.

Clone the repository:

git clone https://github.com/goodnight000/cstack.git
cd cstack

Choose a skill, then link it into your agent's skills directory. Run the appropriate commands from the repository root.

skill_name=reflect  # or rsi, video-edit, animated-video, video-script, asset-library

To install the video skills together, run either block below once per name in video-script video-edit animated-video asset-library, or wrap it in for skill_name in video-script video-edit animated-video asset-library; do … done.

For Codex:

mkdir -p "$HOME/.codex/skills"
test ! -e "$HOME/.codex/skills/$skill_name" && \
  ln -s "$PWD/skills/$skill_name" "$HOME/.codex/skills/$skill_name"

For Claude Code:

mkdir -p "$HOME/.claude/skills"
test ! -e "$HOME/.claude/skills/$skill_name" && \
  ln -s "$PWD/skills/$skill_name" "$HOME/.claude/skills/$skill_name"

If the selected skill is already installed, compare and back up that version before replacing it. If both agents share a skills directory, install once. You can copy the folder instead of linking it; a linked installation follows changes in this checkout. Reload your agent session after installation if needed.

Ask your agent:

Use the reflect skill to review this session, including the result and any avoidable effort.
Show me the findings and proposed changes before editing reusable instructions.

You can also name other threads and a time window when your agent has access to their history. Reflect does not provide its own history connector.

The review separates consequential findings from changes worth adding to reusable guidance. It can report an inefficiency without recommending another rule. It asks for approval before editing instructions and does not authorize publication, memory changes, or resuming the reviewed task.

Use the rsi skill to test one improvement to this agent workflow.
Reuse our Reflect findings, compare with the current version, and state a bounded
experiment plan. Keep installed instructions unchanged until I approve adoption.

RSI uses the project's existing runners and checks. It can reuse Reflect when installed, or reconstruct the necessary evidence directly. It adds no background service or test platform. Give it the task, experiment budget, and adoption scope; it preserves those bounds across the work. A successful local comparison is evidence for the tested scope, not proof of unlimited recursive improvement.

Use the video-edit skill to edit the footage in this project.
Start from the latest accepted version and the brief. Preserve complete takes,
use supporting visuals that match the spoken claims, and check the actual export.

On first use the skill runs scripts/preflight.py, which checks for FFmpeg, Python 3, and uv, which runs local transcription (Phonon on Apple silicon, Whisper elsewhere). It shows the install command for anything missing and runs it only if you agree. DaVinci Resolve (free) is recommended for a native, editable project; without it the skill renders with FFmpeg. Browser research and publishing need browser tools and your accounts.

With no brief, the skill asks where the video will be posted, how long it should be, and whether you want music, sound effects, or a hook headline, then follows its default short-form style. To keep your own defaults across updates, put them in ~/.video-edit/profile.md.

Of the bundled sound effects, 33 are CC0 Freesound recordings; the other 28 are not covered by the MIT license and their sources and unverified rights status are listed in their catalog.

Use the animated-video skill to turn this script into an animated video.
Confirm the format and show me one style frame before building the full piece.

For the best results, run it with Claude Opus 5.5 at high or max effort. Short unnarrated pieces take one agent and a storyboard; longer narrated films use the full production with parallel animators and a review panel.

The skill needs a shell, Node.js with Remotion, FFmpeg/ffprobe, and Python 3 with faster-whisper for word timings. Voiceover comes from the user or a separate tool. Music and effects come from the user's files, on-disk samples such as the video-edit sound library, or sound synthesized in code. The template holds style-neutral infrastructure only (timing, camera, still and render scripts, synthesis helpers). Optional parts from past films (a globe with a cable web, an iPhone running iMessage, synthesised sound effects) sit beside it to copy and restyle. Characters, art direction, and creative defaults are designed per project or supplied by you.

For photoreal or physically detailed worlds, the skill reads the references for scale, materials, light and motion, chooses between three.js in Remotion and an optional Blender render, and gets one moving test shot accepted before the full build. For the three.js route it starts from a tested realism kit in templates/parts/: sky and sun with environment lighting and haze, shadows that hold still under a moving camera, seeded grass, trees and stone, frame-driven wind, and cause-driven weathering. A runnable outdoor example shot (examples/realistic-3d) renders at under half a second per 720p frame and checks that re-rendered frames are identical. It is a detailed 3D starting point that still reads as CGI, not photographic quality.

Asked by video-edit (or by you) for a single animated or 3D shot inside a camera edit, it skips the story steps, works to the edit's frames and returns the clip with its sound events.

Use the asset-library skill to set up a library for my video projects.
Before sourcing anything for this edit, search it, and register what we build as soon as it works.

The library is one SQLite database plus your files, under one folder you choose, driven by scripts/library.py (Python 3, FFmpeg for probing and previews). There is no server. Media is versioned by content hash; characters, code and 3D sets are snapshotted as bundles with their files, dependencies, an example, a preview and a validation command, so a cut recorded with a version can always find what it used. At delivery the producing skill declares which library entries the cut selected, and the check fails until each has a use for that cut and version; whether an asset was worth keeping stays your call. Put the library's location and your reuse rules in ~/.asset-library/profile.md.

Use the video-script skill to brainstorm ideas for my next video.
Rank them by who has a stake, and give each a hook, a second sentence, and what we'll show.

It works best with a library of your past videos and their analytics: one Markdown file per video plus an index, in the library format. Point the skill at it, and at your voice and pace, in ~/.video-script/profile.md. To review a posted video, give it the insights screenshots; it reads the retention curve with scripts/retention_curve.py (Python 3 with Pillow and numpy) and checks the reading against the app's average watch time.

Reflect's evaluation cases cover feedback, constructive pushback, unnecessary repeat checks, necessary investigation, findings that warrant no instruction changes, and keeping one project's output from becoming a default when it seeds a new skill.

Animated Video's evaluation cases cover locking the format before building, treating a past film as evidence rather than a template, and making an explainer without a reference video.

RSI's evaluation cases cover missed skill discovery, exposed final tests, altered graders, authorized local adoption, creative judgment, and unproven improvement of the improvement process.

To assess behavior, give an independent agent SKILL.md, a case's request and session evidence, then its feedback at the appropriate point. Keep the case's checks hidden until reviewing the response. These are evaluation scenarios, not an automated test suite or a guarantee of future performance.

Run Video Edit's caption-helper check with Python 3:

python3 skills/video-edit/scripts/test_make_captions.py

This checks caption generation, time mapping, raw transcript preservation, and known warning/error behavior. It does not validate a Resolve installation, listening quality, visual taste, or an end-to-end video edit on your machine.

Run Asset Library's check, which builds a throwaway library in a temporary folder:

python3 skills/asset-library/scripts/test_library.py

It checks ingest and duplicate detection, bundle versions and their publishing, use recording, the project check against a cut declaration, and that an older database is backed up and migrated without losing rows. It says nothing about whether the assets in your library are any good.

Charade is a set of line icons that act out what they mean. Each one plays a short animation of its own verb when its button or link is hovered or focused: the lock unlocks, the bell swings, the trash eats a crumb. There are 203 icons in plain SVG and CSS, with no dependencies. Open icons/index.html from a local server to browse them.

Keep each skill self-contained and focused on a recurring need. Include examples or evaluation cases that expose meaningful mistakes. Keep transcripts, credentials, personal project history, and machine-specific paths out of published skills. For third-party contributions, preserve attribution and confirm license compatibility.

Skills stay general; personalization lives outside the repo, in each skill's ~/.<skill>/profile.md and the files it points to. Before changing a skill, ask whether the change would be right for a stranger making their own videos. If so, write it here generically. If it is right only for you (your audience, taste, numbers, machine or accounts), put it in your profile. When it is both, split it: the general rule here, your value and the evidence behind it in your profile. Bundle only assets whose license allows redistribution, and keep the creator's own footage, likeness and personal media out.

MIT.