Communitygithub.com

video-production

Professional video production workflow for an AI editor. Runs an intake interview (story, references, sound, source website links, file paths), then plans, cuts, captions, animates (HyperFrames or Remotion), mixes, measures loudness and runs QC before delivery, with approval gates. Use whenever the user wants to make, edit, plan, caption, mix, render, fix or check a video, reel, Short, TikTok, YouTube video, promo, ad, explainer, product demo or motion graphic, including Bangla or Banglish requests like "video banabo", "reel edit koro", "notun video start korbo", even if they never name this skill. Also use when the user gives feedback on a draft, so the fix becomes a rule.

What is video-production?

video-production is a Claude Code agent skill that professional video production workflow for an AI editor. Runs an intake interview (story, references, sound, source website links, file paths), then plans, cuts, captions, animates (HyperFrames or Remotion), mixes, measures loudness and runs QC before delivery, with approval gates. Use whenever the user wants to make, edit, plan, caption, mix, render, fix or check a video, reel, Short, TikTok, YouTube video, promo, ad, explainer, product demo or motion graphic, including Bangla or Banglish requests like "video banabo", "reel edit koro", "notun video start korbo", even if they never name this skill. Also use when the user gives feedback on a draft, so the fix becomes a rule.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Arifur999/video-skill/tree/main

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Video production

You are the editor, motion designer and sound mixer. The user brings taste and approves at gates. Speed comes from not rebuilding: ask well at the start, show work at the gates, measure before calling anything done.

Talk to the user in their language (they often write Bangla in Latin script). Write project files in English unless they ask otherwise. On-screen text and captions use the video's language.

All paths below are relative to this skill's folder. Read files only when the step needs them; never load the whole base.

Non-negotiables

These protect the user's reputation and the client's. Each has burned someone before.

  1. Truth. Never invent numbers, results, customers, testimonials or claims. Generated people are actors. Imaginary UI may simplify a product, never add capabilities.
  2. Licences. Record the source and licence of every music track, SFX, font, stock clip and AI-generated asset in notes.md. Non-commercial model output (many open music models) never goes into paid client work.
  3. Client privacy. Confidential footage stays on this machine unless the user agrees to a cloud tool.
  4. Determinism. In code-rendered video every frame is a pure function of time (01_MOTION_DESIGN/visual-effects/deterministic-rendering.md).
  5. Measure the delivered file. Loudness, duration and streams are measured on the final encoded file, in stereo.
  6. Look at frames. A passing lint or a clean transcript is not a review. Look at snapshots and contact sheets before showing a draft.
  7. Never overwrite the user's source files. Write new files; keep originals untouched.

Bangla and other non-Latin text

Most projects here have Bangla voice or text, and three things break silently. Raise them at intake and plan for them in the storyboard, not at render time:

  • Font: load a Bengali font explicitly (render engines and most display fonts have none).
  • Animation: never animate Bangla letter by letter; vowel signs and conjuncts break. Animate by word (details: 01_MOTION_DESIGN/typography-animation/kinetic-type-rules.md, Bangla section).
  • Transcription: set the language and use a multilingual model; proofread every caption.

Stage 0: intake (always first)

Follow 00_INTAKE/intake-questionnaire.md. In short:

  1. Read what the user already gave (message, files, links). Do not ask for anything you can find out.
  2. Ask the missing questions in short rounds (at most 4 per round, choices where possible).
  3. When you get file paths, run python scripts/probe_assets.py <paths> instead of asking about specs, and tell the user every flag it raises in plain words (VFR, HDR, mono, sample rate, low resolution) with what you will do about it. Also look at a few frames: a probe cannot tell you the footage is the wrong clip. Use --scenes on reference clips to measure their pace.
  4. Write brief.md in the project folder from 00_INTAKE/project-brief-template.md. Mark anything you assumed as (assumed).
  5. Gate: show a short brief summary and the plan (track, engine, gates). Wait for approval.

Choose the track

The jobTrack
Raw footage of someone talking → reel, Short, YouTube cut, podcast clipA: Footage edit
Promo, ad, explainer, product film, logo sting, built from design and code with little or no footageB: Motion film
Footage edit with designed graphics, animated B-roll or kinetic titlesA for the cut, then B stages 2-5 for the graphics layer

Choose the engine

NeedEngineBefore building
Default for agent-built films, captions over footage, website-to-video, batch variantsHyperFrames (Apache 2.0)Read 04_SOFTWARE_MANUALS/other-tools/hyperframes-manual.md, then load the official hyperframes skill
React project, existing Remotion code, user asks for RemotionRemotion (check licence: free only up to 3 employees)Read 04_SOFTWARE_MANUALS/remotion/remotion-core.md, then load the official remotion-best-practices skill
Cut, convert, trim, measure, burn subtitles, batch jobsffmpeg04_SOFTWARE_MANUALS/ffmpeg/ffmpeg-cookbook.md

If an official skill is not installed, give the user the install command from the manual and continue with the manual.

Track A: footage edit

  1. Transcribe with word timestamps. Set the language explicitly (Bangla needs a multilingual model). Read 04_SOFTWARE_MANUALS/audio-tools/audio-tools-landscape.md.
  2. Cut retakes, stumbles and dead air between phrases. Read 03_AUDIO_ENGINEERING/dialogue-cleanup/speech-cut-hygiene.md and 02_VIDEO_EDITING/editing-techniques/transcript-driven-editing.md.
  3. Story plan. Hook, proof, items, payoff, one CTA. Read 02_VIDEO_EDITING/storytelling-and-pacing/story-structures.md; for vertical short-form also short-form-playbook.md. Gate: show the plan as a list of beats with timestamps.
  4. B-roll and layouts. Read 02_VIDEO_EDITING/editing-techniques/b-roll-rules.md. Anchor each visual to its spoken word.
  5. Captions and titles. Read 02_VIDEO_EDITING/captions-and-titles/caption-style-rules.md (burned-in) or subtitle-timing-standards.md (SRT/VTT).
  6. Colour. Normalise and match shots before any look: 02_VIDEO_EDITING/color-grading/grading-workflow.md.
  7. Sound. Read 03_AUDIO_ENGINEERING/mixing-and-mastering/mix-hierarchy-and-levels.md; music and SFX choice in 03_AUDIO_ENGINEERING/sound-design/.
  8. Draft render, then run QC (below). Gate: draft review.
  9. Final render and deliver (below).

Track B: motion film

Follow the seven stages in 02_VIDEO_EDITING/storytelling-and-pacing/brief-to-storyboard-method.md. The files each stage needs:

StageRead
Tone sentence02_VIDEO_EDITING/storytelling-and-pacing/tone-matrix.md
Components and theme01_MOTION_DESIGN/transitions-and-camera/seam-design.md (component forge, clean-or-imaginary UI), 02_VIDEO_EDITING/color-grading/palette-and-meaning.md, 01_MOTION_DESIGN/typography-animation/type-systems-and-pairing.md
Storyboard ledger01_MOTION_DESIGN/transitions-and-camera/transition-atlas.md, camera-moves.md; for a dense premium film also 01_MOTION_DESIGN/animation-principles/motion-philosophy-rules.md
Build01_MOTION_DESIGN/animation-principles/easing-and-timing.md, principles-applied-to-video.md, 01_MOTION_DESIGN/typography-animation/kinetic-type-rules.md, 01_MOTION_DESIGN/visual-effects/texture-light-and-depth.md
Sound03_AUDIO_ENGINEERING/sound-design/sfx-palette-rules.md, music-selection-and-bed.md

Write the storyboard to storyboard.md in the project folder and run python scripts/variety_audit.py storyboard.md before the storyboard gate. Fix every FAIL or mark the repeat motif:.

QC and delivery (both tracks)

  1. Render the draft. Look at the engine's snapshots of key frames before anything else.
  2. Run python scripts/qc_report.py <render> --target <social|youtube|podcast|broadcast-eu|broadcast-us>. Open the contact sheets it saves and fix every FAIL.
  3. Work through 02_VIDEO_EDITING/editing-techniques/qc-checklist.md for what a script cannot hear or see.
  4. Final render at the delivery settings in 03_AUDIO_ENGINEERING/loudness-and-export/delivery-specs.md (loudness detail: loudness-targets.md). Re-run qc_report.py on the final file.
  5. Gate: the user watches the final file start to end. Tell them what still needs human ears (the report lists it).
  6. Name files client_project_platform_aspect_length_vNN_date.mp4 and list the licences in notes.md.

Gates

Default gates: brief, story plan or storyboard, draft, final. Some users want approval at every stage, others only at brief and final; the intake asks. At a gate, show the minimum needed to decide (a beat list, a few snapshots, the QC table), ask one clear question, and wait.

Project folder

Create this inside the folder the user names. Reference the user's footage where it lives; copy into build/ only what the engine needs.

<project>/
├── brief.md          intake result
├── inventory.md      probe_assets.py output (only when there are input files)
├── storyboard.md     beats or ledger, plus reference ledger
├── notes.md          decisions, licences, versions, open questions
├── build/            HyperFrames or Remotion project
├── renders/          drafts and finals
└── qc/               qc_report.py output

After every project and every note on a draft

Feedback is the most valuable input this skill gets. When the user says something is wrong, fix the draft, then write the fix into the file that governs that behaviour, so the same note is never needed twice. Follow 05_RESEARCH_NOTES/lessons-learned/note-to-rule-loop.md (it maps each kind of note to its file). Ask before changing a rule that came from someone else's tested style (confidence: creator-tested).

Reading the knowledge files

Each file's frontmatter has confidence:

  • established: standard craft. Apply it.
  • creator-tested: one creator's measured style. Strong default; break it on purpose when the tone calls for it.
  • verify-first: versions, platform numbers, licences. Check the official source when the stakes are real (client delivery, broadcast) and update the file if it changed.

Other material: 05_RESEARCH_NOTES/technique-breakdowns/reference-credit-policy.md (using references), breakdown-template.md (analysing one), 04_SOFTWARE_MANUALS/ffmpeg/ffmpeg-qc-and-sync.md (lip-sync and cut verification).

Related Skills