make-video
Skill4 starsChanged 4 months agoMITView on GitHub
READ THIS FIRST for any request to make, edit, or publish a video in this repo — "make a video", "new video", "edit this clip", "create a reel/short", "faceless video", "turn this post/idea into a video", or any ambiguous video ask. This is the entry-point router — it asks what you're making (talking-head short 9:16 · long-form 16:9 · faceless motion reel · reel script only), scaffolds projects/<name>/, then runs the right pipeline end-to-end and ships to projects/<name>/outputs/. Route through this instead of picking video-edit / motion-design-edit / reel-script directly.
- Reads credentials
7 other projects use a version of this.
Adopted by
Most recently taken up by
- jayden-dang/skillstheir own version, first seen Nov 8, 2026
- murathany7/reel-studiotheir own version, first seen Nov 8, 2026
- jutsuhq/content-factorytheir own version, first seen Nov 1, 2026
- instavar/remotion-templatestheir own version, first seen Aug 4, 2026
- 87oo/ai-video-crewtheir own version, first seen Jul 17, 2026
- GoldLegendW80/llm-video-makertheir own version, first seen Jun 17, 2026
Other versions7 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 7
- av/skills★ 28same name
- GoldLegendW80/llm-video-maker★ 22same name
- jutsuhq/content-factory★ 6same name
- murathany7/reel-studio★ 5same name
- instavar/remotion-templates★ 3same name
- 87oo/ai-video-crew★ 1same name
- jayden-dang/skills★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository5 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- motion-design-editSkill
- reel-scriptSkill
- video-edit-longformSkill
- video-editSkill
The file
We split this skill into the rules it gives an agent
8 rules. 0 are common, 3 are distinctive, and 5 are very specific to this project.
0
Common
Most of the 6 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 16
Ask (via AskUserQuestion, one round, skipping anything the user already said):
- Very specific: line 27
Brand setup, once per install (routes A/B only): if repo-root BRAND.md doesn't exist yet, run the brand-setup flow in SETUP.md before your …
- Very specific: line 55
Default — the cloud transcription bridge, which needs ONE of ASSEMBLYAI_API_KEY or OPENAI_API_KEY in the repo .env (AssemblyAI is tried …
- Very specific: line 67
Mandatory: read the transcript back against what was actually said and hand-patch wrong words in words.json (keep start/end), then make …
- Distinctive: line 76
Follow EXECUTION.md §§0–5 exactly.
- Very specific: line 108
Run /reel-script on it → beat-tagged script (HOOK / CREDIBILITY / REFRAME / TIP 1-3 / CTA, each tagged pr_card/pr_punch/pr_proof).
- Very specific: line 142
Run /reel-script on the idea/post.
- Distinctive: line 169
After shipping or pausing, append the session to WORKLOG.md (state, decisions, what's next).
“Common” is measured: kept by at least 6 in 10 of the 6 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/make-video — one entry point for every video
You are the dispatcher for this video pipeline. Never guess a workflow: confirm the intent, then commit to exactly one route below.
1. Intake (always first)
Ask (via AskUserQuestion, one round, skipping anything the user already said):
- What are we making?
talking-head short— 9:16 reel from a camera cliplong-form— 16:9 YouTube video from a camera clipfaceless motion reel— paper-style motion graphics over a voiceover, no camerareel script only— just write the script
- Project name — then scaffold if missing:
mkdir -p projects/<name>/{raw,refs,working,outputs/previews} - Source — clip path, VO path, or the idea/post text. Copy source media
into
projects/<name>/raw/before doing anything else. - Brand setup, once per install (routes A/B only): if repo-root
BRAND.mddoesn't exist yet, run the brand-setup flow in SETUP.md before your first render — it asks primary/secondary/accent colors, background preference, and fonts, then edits the token files ($SKILL/remotion/src/templates/hfBrand.ts,$F5/remotion/src/motion/tokens.ts) and writesBRAND.mdso you never ask again. IfBRAND.mdalready exists, skip straight to rendering — the token files already reflect the user's brand. - If route C (faceless motion reel) with a VO in hand — also ask
visual style:
paper brand(cream/ink/orange-punch, the original reference-reel look — default) vsvisual os(the same brand-setup preset as routes A/B, or its own configured brand ifBRAND.mddiffers for this project). Write the answer toprojects/<name>/.pr_style(one word:paperorvisual-os) —series_pipeline.shreads it automatically on every render from then on, no need to ask again.
Shorthand used below (run from the repo root):
SKILL=.claude/skills/video-edit # 9:16 finisher (hf_* graphics)
F5=.claude/skills/video-edit-longform # 16:9 finisher (viz graphics)
MSKILL=.claude/skills/motion-design-edit # faceless pr_* graphics
PY="$(pwd)/.venv/bin/python" # repo-root shared venv (WhisperX local fallback)
OUT="$(pwd)/projects/<name>/outputs" # ALWAYS absolute2. Transcription (all routes)
Default — the cloud transcription bridge, which needs ONE of
ASSEMBLYAI_API_KEY or OPENAI_API_KEY in the repo .env (AssemblyAI is
tried first if both are set; pass --provider openai to force the other):
.venv/bin/python .claude/skills/make-video/scripts/transcribe_assemblyai_bridge.py <media>Prints the workdir (~/.cache/video-edit/<stem>_<sha1[:12]>/) and writes
words.json + .polished there. Fallback for sensitive clips: local WhisperX
($PY "$SKILL/scripts/transcribe.py" <media>).
Mandatory: read the transcript back against what was actually said and
hand-patch wrong words in words.json (keep start/end), then make sure
.polished exists. ASR mishears predictably (past examples: "instructions"→
"injections", "the floor"→"to Claude").
3. Routes
A. Talking-head short (9:16) → video-edit (hf_* graphics)
Follow EXECUTION.md §§0–5 exactly. Summary:
- Phone portrait clip stored landscape+rotation flag? Bake it first
(EXECUTION §0 ffmpeg command) → use the
.portrait.mp4from there on. video-editis a finisher, not a cutter — it expects pre-cut footage. Raw/uncut multi-take recordings need filler-word and silence removal done first (by hand, or with your own cutting tool) before this route.- Transcribe (§2 above) → author
<workdir>/broll_plan.jsonwithhf_*kinds (EXECUTION §2 table). Vary the templates — rotatehf_stat,hf_title,hf_chart,hf_listicle,hf_lower_third,hf_before_after…; reserve the mascothf_calloutfor a beat or two, never the through-line. $PY "$SKILL/scripts/lint_plan.py" <workdir>/broll_plan.json— must pass.- Preview:
OUT_DIR="$OUT" CAPTION_EMPHASIS="…" bash "$SKILL/scripts/render.sh" <clip>→ iterate on the plan until the user likes it. - Final (only after explicit sign-off): same command plus
QUALITY=final SCORE=0.
B. Long-form (16:9) → video-edit-longform (viz graphics)
Same loop as route A but against $F5: transcribe → author the beat plan with
its viz/native kinds ($F5 knowledge + template_library.md) → lint →
preview bash "$F5/scripts/render.sh" with absolute OUT_DIR → sign-off →
final. Notes:
- Colors/fonts come from
$F5/remotion/src/motion/tokens.ts— run the brand setup in SETUP.md before your first render so this matches your own brand, not the shipped example preset. - The bespoke one-off compositions from the original bundle (channel-specific
intros/outros) were removed when this skill was cleaned up for reuse — only
the generic
EditedVideoengine and reusabletemplates//motion/library remain. Don't try to reintroduce channel-specific one-offs here.
C. Faceless motion reel → reel-script + motion-design-edit (pr_* graphics)
No VO yet (user brought an idea, notes, or a Facebook/LinkedIn post):
- Run
/reel-scripton it → beat-tagged script (HOOK / CREDIBILITY / REFRAME / TIP 1-3 / CTA, each taggedpr_card/pr_punch/pr_proof). - Save it to
projects/<name>/working/script.md, hand it to the user, and stop. They record the VO and drop it inprojects/<name>/raw/.
With a VO in raw/ — follow EXECUTION.md "Faceless motion-graphics
workflow" exactly. Summary:
- Build the black 1080×1920 base with the VO muxed in, at the VO's exact
ffprobe duration (never
-shortest) →projects/<name>/working/base.mp4. - Transcribe base.mp4 (§2 above) + verify/patch the transcript.
Check for
raw/captions.txt— if present, parse it (scripts/parse_captions.py) and use its lines verbatim as on-screen text; it's authoritative for wording, not a paraphrase source. Without it, compress the transcript into on-screen text yourself as before. - Read
MOTION-STYLE.md(repo root) first, then author<workdir>/broll_plan.json: group ideas into evolving multi-layer scenes (split on nearly every sentence),speech_anchoron every beat, 5-part arc per$MSKILL/knowledge/paper_style.md.projects/motion-reel/working/broll_plan.jsonis the reference plan. python3 "$MSKILL/scripts/lint_plan.py" <workdir>/broll_plan.json projects/<name>/working/base.mp4- Preview + final, both scored (music + SFX) by default:
bash "$MSKILL/scripts/series_pipeline.sh" <name> preview/bash "$MSKILL/scripts/series_pipeline.sh" <name> final(only after explicit sign-off). This reads.pr_styleautomatically and writes toprojects/<name>/outputs/previews/andoutputs/respectively — no per-project script needed. See EXECUTION.md's "Recurring series" subsection for the full pattern. - If this project will run more than one episode, scaffold a
project-local
STYLE.mdcreative contract (palette table, component rules, beat arc, timing rules) — mirrorprojects/claude-code-edit/STYLE.md's structure — so every future episode reuses the same locked look instead of re-deriving it.
D. Reel script only → reel-script
Run /reel-script on the idea/post; save the beat-tagged script to
projects/<name>/working/script.md. Offer route C as the follow-up once
they've recorded the VO.
4. Hard rules (every route)
- Brand: routes A/B always render in whatever brand
BRAND.mdrecords — run the brand-setup flow (SETUP.md) once per install before the first render ifBRAND.mddoesn't exist yet (see §1 step 4). Ships with an example "Visual OS" preset (navy ink, electric-blue accent, Claude-orange punch, light bg; Georgia titles · Inter body · JetBrains Mono peripheral chrome only) until you customize it. Whatever palette you land on, keep it disciplined: one clear accent color, one distinct punch/execution color, and don't let peripheral-chrome fonts bleed into headline/body roles. Route C (pr_*) supports two selectable styles — Paper Brand (default, its own cream/ink/orange palette, independent of the brand-setup flow) or the configured Visual-OS-style brand, chosen per project via.pr_style(see §1 step 5 above). Token files are the source of truth:$SKILL/remotion/src/templates/hfBrand.ts,$F5/remotion/src/motion/tokens.ts,$MSKILL/remotion/src/templates/paperBrand.ts. - Never auto-run the final render. Preview → user says "ship it" → final.
- Absolute
OUT_DIRalways — render.sh cd's into its remotion dir; a relative path strands the output inside the skill. pr_proofwithoutimage_pathrenders a blank card — source a real screenshot/photo or explicitly flag the placeholder before calling it done.- Finals belong in
projects/<name>/outputs/, previews inoutputs/previews/, intermediates in the~/.cache/video-edit/workdir — never inraw/. - After shipping or pausing, append the session to WORKLOG.md (state, decisions, what's next).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

