agentleFS
imtamiliniyanimtamiliniyan/claude-video-skills/.claude/skills/make-video/SKILL.md

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

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
  1. Distinctive:

    Ask (via AskUserQuestion, one round, skipping anything the user already said):

    line 16
  2. Very specific:

    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 …

    line 27
  3. Very specific:

    Default — the cloud transcription bridge, which needs ONE of ASSEMBLYAI_API_KEY or OPENAI_API_KEY in the repo .env (AssemblyAI is tried …

    line 55
  4. Very specific:

    Mandatory: read the transcript back against what was actually said and hand-patch wrong words in words.json (keep start/end), then make …

    line 67
  5. Distinctive:

    Follow EXECUTION.md §§0–5 exactly.

    line 76
  6. Very specific:

    Run /reel-script on it → beat-tagged script (HOOK / CREDIBILITY / REFRAME / TIP 1-3 / CTA, each tagged pr_card/pr_punch/pr_proof).

    line 108
  7. Very specific:

    Run /reel-script on the idea/post.

    line 142
  8. Distinctive:

    After shipping or pausing, append the session to WORKLOG.md (state, decisions, what's next).

    line 169

“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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/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):

  1. What are we making?
    • talking-head short — 9:16 reel from a camera clip
    • long-form — 16:9 YouTube video from a camera clip
    • faceless motion reel — paper-style motion graphics over a voiceover, no camera
    • reel script only — just write the script
  2. Project name — then scaffold if missing: mkdir -p projects/<name>/{raw,refs,working,outputs/previews}
  3. Source — clip path, VO path, or the idea/post text. Copy source media into projects/<name>/raw/ before doing anything else.
  4. 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 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 writes BRAND.md so you never ask again. If BRAND.md already exists, skip straight to rendering — the token files already reflect the user's brand.
  5. 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) vs visual os (the same brand-setup preset as routes A/B, or its own configured brand if BRAND.md differs for this project). Write the answer to projects/<name>/.pr_style (one word: paper or visual-os) — series_pipeline.sh reads 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 absolute

2. 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:

  1. Phone portrait clip stored landscape+rotation flag? Bake it first (EXECUTION §0 ffmpeg command) → use the .portrait.mp4 from there on.
  2. video-edit is 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.
  3. Transcribe (§2 above) → author <workdir>/broll_plan.json with hf_* kinds (EXECUTION §2 table). Vary the templates — rotate hf_stat, hf_title, hf_chart, hf_listicle, hf_lower_third, hf_before_after…; reserve the mascot hf_callout for a beat or two, never the through-line.
  4. $PY "$SKILL/scripts/lint_plan.py" <workdir>/broll_plan.json — must pass.
  5. Preview: OUT_DIR="$OUT" CAPTION_EMPHASIS="…" bash "$SKILL/scripts/render.sh" <clip> → iterate on the plan until the user likes it.
  6. 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 EditedVideo engine and reusable templates//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):

  1. Run /reel-script on it → beat-tagged script (HOOK / CREDIBILITY / REFRAME / TIP 1-3 / CTA, each tagged pr_card/pr_punch/pr_proof).
  2. Save it to projects/<name>/working/script.md, hand it to the user, and stop. They record the VO and drop it in projects/<name>/raw/.

With a VO in raw/ — follow EXECUTION.md "Faceless motion-graphics workflow" exactly. Summary:

  1. 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.
  2. 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.
  3. 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_anchor on every beat, 5-part arc per $MSKILL/knowledge/paper_style.md. projects/motion-reel/working/broll_plan.json is the reference plan.
  4. python3 "$MSKILL/scripts/lint_plan.py" <workdir>/broll_plan.json projects/<name>/working/base.mp4
  5. 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_style automatically and writes to projects/<name>/outputs/previews/ and outputs/ respectively — no per-project script needed. See EXECUTION.md's "Recurring series" subsection for the full pattern.
  6. If this project will run more than one episode, scaffold a project-local STYLE.md creative contract (palette table, component rules, beat arc, timing rules) — mirror projects/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.md records — run the brand-setup flow (SETUP.md) once per install before the first render if BRAND.md doesn'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_DIR always — render.sh cd's into its remotion dir; a relative path strands the output inside the skill.
  • pr_proof without image_path renders a blank card — source a real screenshot/photo or explicitly flag the placeholder before calling it done.
  • Finals belong in projects/<name>/outputs/, previews in outputs/previews/, intermediates in the ~/.cache/video-edit/ workdir — never in raw/.
  • After shipping or pausing, append the session to WORKLOG.md (state, decisions, what's next).

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.