Sixteen installable skills for Claude Code and other skill-aware agents: social-media judgement, short-form video planning, production handoffs, and toolchain checks. Eight need no dependencies. The video engine lives in the same repo and installs separately when you want execution, not just guidance — what each layer gives you.
terminal · via skills.sh
npx skills add scrollmark/social-skills # all 16
npx skills add scrollmark/social-skills -s hook-anatomy # just one
# Needs Node 22.20+. No Node? Same result without it:
# mkdir -p ~/social-skills
# curl -Ls https://github.com/scrollmark/social-skills/archive/refs/heads/master.tar.gz \
# | tar xz --strip-components=1 -C ~/social-skills
# ~/social-skills/install.sh
#
# Keep that directory: install.sh symlinks into it, so extracting to /tmp
# loses all 16 skills at the next purge — at once, and with no error.
terminal · symlinks into ~/.claude/skills
git clone https://github.com/scrollmark/social-skills.git
cd social-skills && ./install.sh
The npx route uses the
skills.sh CLI — the open agent-skills
ecosystem — and works with Claude Code and other supported agents.
Want just one skill? Every card below has its own one-liner. Some video
skills add an optional second step:
pip install from the repo.
The next section explains what that unlocks.
Install options
What each skill needs.
Eight skills need only the skill files. The rest can use the video
engine, and four also carry small scripts that run as soon as the
skill is installed.
planning, structuring, or critiquing a short-form video
brand-kit
engine
—
keeping a run of videos visually consistent
media-acquisition
engine · ffmpeg
1
deciding where a shot's pictures come from
audio-acquisition
engine · [audio] · [captions] · ffmpeg
2
a video needs a voice, a bed, or a fix for bad levels
subject-compositing
engine
—
planning a shot that composites a person into a scene
edit-handoff
engine
—
an editor is picking the cut up from here
video-production
engine · [qc] · ffmpeg · node
4
producing a video end to end, or a run has stalled
agent-interview
nothing
—
an agent has to interview someone before it can act
studio-setup
engine · [qc] · ffmpeg · node
1
checking or repairing a video toolchain
“Engine” is video-studio-engine, installed from this
repo rather than PyPI — the exact command is under
One repo. Four layers. Bracketed names are its
optional extras. Generated from the skills themselves, and checked on
every push.
Dependency guide
One repo, four layers.
The skill pack is useful even before you install a video toolchain.
Some skills are pure guidance, some ship small scripts, and some use
the Python and Remotion engine in the same public repo. The layers
below show what works immediately, what needs one
pip install, and what reaches for an external API key
or a licence you may not need to buy.
8 skills
Layer one · prose
Nothing to install.
The seven social-media skills are pure judgement — tone, hooks, trend timing, platform
dialect, post-mortems, voice. No external tools, no accounts, no keys. So is
agent-interview, which has nothing video-specific in it. These are the
same whichever way you install them.
Six programs use the Python standard library and take their input on the command line, so
they ship inside the skill folder and travel with
npx skills add:
doctor.py (what is reachable right now),
measure.py (durations and a track's energy structure),
normalize_audio.py (loudness after a render),
poster.py (thumbnail shortlist and contact sheet),
qc_render.py (the render checked against the plan it was built from) and
prekey.py (key a generated clip to alpha). Two of them are carried by
two skills each, so the skill stays self-contained.
They want ffmpeg and
ffprobe on your PATH. That is the one
thing the repo cannot ship for you.
The other thirty-six programs use third-party dependencies or read project
state: a props document, a composer directory, a styles tree. They live in
src/video_studio/ and run as
video-studio <command>. The base
install is small; heavier dependencies are opt-in extras, so you install
only the parts you use.
terminal · optional, video skills only
uv tool install --python 3.12 --refresh 'video-studio-engine[standard] @ https://github.com/scrollmark/social-skills/archive/refs/heads/master.tar.gz'
# Needs no git, no Xcode tools and no system Python — uv brings its own.
# Get uv first: curl -LsSf https://astral.sh/uv/install.sh | sh
# source "$HOME/.local/bin/env" # or open a new terminal
#
# --python 3.12 is pinned: on 3.13 the [audio] extra installs cleanly and
# silently has no voice. --refresh because GitHub's tarball is cached forever.
# Or one extra at a time — same URL, different bracket:
# [sourcing] + source_clips (yt-dlp)
# [audio] + tts_kokoro, gen_music, duck_music
# [captions] + gen_captions (local whisper word timings)
# [generate] + gen_veo, gen_boil
# [vision] + composite_subject, track_pointing (needs mediapipe)
# [export] + export_edit (OpenTimelineIO)
# [qc] + qc_analyze — the render checked against its plan, decoded
#
# The four model-backed QC checks stay opt-in, one extra each:
# [qc-ocr] captions and text layout [qc-yolo] planned subjects on screen
# [qc-face] lip sync [qc-clip] scene footage vs its prompt
It installs from the repo, not from PyPI — the name
video-studio-engine is
not published there, so a bare
pip install video-studio-engine
does not get you this. The URL is the install.
The engine is independent of the skills and installed separately:
npx skills add never touches it. The
cards below spell out which skills still work without it and which
become planning guidance only.
licences, not code
Layer four · yours to licence
A Remotion licence, or not.
The Remotion composer ships in this repo now —
composer/, React driven entirely by the
props document the engine writes. What is still not ours to ship is Remotion
itself: the project declares it as a dependency and vendors none of it, because
Remotion requires a paid Company Licence above three people and for hosted or automated
rendering. The code is here; that license stays between you and them. It is Node, not
Python, so npm install in
composer/ once, and then
npx remotion render turns a storyboard
into a file.
That licence is now optional. Scrollmark's own editor renders too —
MIT, no seat threshold, no paid tier — and an agent drives it over MCP. It does not read
the props document: it has its own project format, so the agent builds the timeline from
the same storyboard through the editor's own commands. The trade is honest rather than
hidden. Remotion renders headless, so it suits a run nobody is watching; the editor
needs to be running, which suits one where somebody is.
ElevenLabs voice has a script now —
video-studio tts_eleven — but it bills
per character and returns no word timings, so captions for an ElevenLabs read come from
gen_captions. The local voice is free
and runs offline; reach past it deliberately. The automated quality gate at the last step of a
production run used to be an unpublished internal CLI; it is now
qc_render.py, bundled, needing no install. It proves a render matches
its plan — it cannot tell you the plan was worth rendering, so the looking is still
yours.
Remotion · Nodeor the editor · no licenceElevenLabs · paid APIffmpeg on PATH
Without the engine
What you still get.
Layer three is optional. Here is what each video skill still gives you without it.
loses nothing
Unaffected
The whole social half.
All seven social-media skills and agent-interview are untouched by any
of this.
studio-setup loses the least of the video skills:
doctor.py ships with it, so the status
report and the triage work immediately. Only the install plan —
video-studio setup and its
auto/system/manual classification — needs the package.
audio-acquisition is the best-off of the production skills. It carries
measure and
normalize_audio, so the
measure-don't-trust discipline and the music-costs-6-dB fix are both executable out of
the box. Voice, generated music and licence-filtered effects are
[audio].
holes in it
Partly executable
Useful, but incomplete.
video-formats still works for eight of its ten formats. Those are
structure you can follow in any editor. Boil is roughly 40% inoperative
without [generate]: its mark
vocabulary, shape files, and palette workflow are
gen_boil, so the interview questions
have no answers to pick from. PointerPopups does not work at all as written
without [vision] — its entire grammar
is the output of a hand-tracking pass.
brand-kit lets you agree a brand with someone and hand-write the files.
The preset mechanism that lists, shows, saves and applies one is
video-studio styles, so applying a
brand becomes manual editing.
media-acquisition keeps prekey.
The provider ladder and the landmine list will still tell a human which library to open,
but the searching, generating, licence filtering and the five-way check on what came back
are [sourcing,generate].
briefings
Package-side
Briefings, not workflows.
subject-compositing is entirely package-side: segmentation, the feathered
occlusion seam, the arm punch-through and the gesture timing all need
[vision], because MediaPipe and OpenCV
cannot be bundled into a skill folder. What survives is the footage brief — what a
segmenter needs from a take — and it is worth having before anyone films.
edit-handoff cannot produce a file without
[export]; the exporters read the
engine's per-project timeline document. What generalises is the judgement: which format
each application actually imports, and that no interchange format carries effects.
video-production ships doctor, normalize, poster and qc_render, but it sequences the
whole engine — props, preview, preflight — and a renderer on top, whether that is
Remotion or the editor. With none of those present there is nothing to sequence. What survives is the shape of a production
run: which decision precedes which, and where money enters.
Each of those skills states its own dependency in a
## Requires section near the top of its
SKILL.md — that is where the authoritative
answer lives. The longer version of this page is
in the README.
The full deck
Sixteen skills, three jobs.
Understanding social media, producing short-form video, and running the working method underneath both. Each skill is a compact playbook: patterns and judgement calls, not API wrappers. They pair well with the SocialGPT MCP server, which brings your real numbers into the conversation.
Group one · 7 skills
Understanding social media.
Tone, hooks, trends, platform conventions, creator voice. All seven run standalone — no external tools, no accounts, no keys.
Interpretation
read-the-room
Helps agents interpret captions, comments, replies, and DMs by reading tone, sarcasm, emoji, parasocial language, and comment-section context.
Reach for it interpreting comments, captions, replies, or DMs
Craft skills for making a short-form video, not only planning one. The judgement is prose you can follow anywhere; execution comes from scripts bundled with the skill or from the engine in the same repo.
Structure
video-formats
Provides ten reusable short-form video structures, from talking-head clips to hand-drawn motion graphics, with guidance for planning and critique.
Reach for it planning, structuring, or critiquing a short-form video
8 of 10 formats need nothingBoil + PointerPopups: pip install
A skill here is prose: a SKILL.md,
the reference files it needs, and a description saying when to reach for it. If you
know something about the feed that Claude keeps getting wrong, that is a pull request.