onuryildriim/claude-kurgu-atolyesi

video-kurgula — yeni videoyu kanalın kendi standardında kurgula ve teslim et

The user drops a finished talking-head export (already cut, e.g. from CapCut/Final Cut) at public/main.mov, puts supporting assets (screenshots, screen recordings) and the music track into public/assets/, says "bunu kurgula", and expects a finished delivery — graphics, full-screen scenes, sound effects, music, QC, docs.

Before anything: the brand

Read marka/MARKA.md first. It holds the channel's identity: who the channel is for, tone of voice, colours, fonts, closing card, things the user never wants. If it still contains the <!-- DOLDURULMADI --> marker, stop and run the marka-kimligi skill first (a short interview) — editing without a brand produces a generic video. src/theme/theme.ts must match MARKA.md; the house rules the user taught you live in references/editorial-style.md → "Ev kuralları".

Chat in the user's language (default Turkish). On-screen text is in the video's language (MARKA.md).

Operating mode

• Default: no approval gates. Make the editorial decisions yourself using references/editorial-style.md, deliver, then take revision notes. Ask only on a true blocker (no video found, source the pipeline cannot handle, missing tool, brand not set up). If MARKA.md says the user wants a content check or plan approval first, do that.
• Music is supplied by the user (an audio file in public/assets/). Never generate music and never spend paid API credits (ElevenLabs etc.) unless the user explicitly asks — state the cost first. Details: references/audio.md.
• Long jobs: start detached (nohup … &), tell the user the ETA per step and a ! tail -f <log> command, and keep working on whatever does not depend on the job. Verify output FILES (frame counts, ffprobe), never just exit codes.
• Keep work/progress.md current after every step — it is the resume point if the session dies. On "devam", read it first and continue from the first unchecked box.
• Never modify or delete: public/main.mov, originals in public/assets/, anything in projects/, and any file already delivered in out/ (new versions get new names).
• You cannot hear audio or see HDR. Verify by measurement and stills, and say plainly which checks need the user's ears/eyes.

Workflow

Commands, durations and known traps for every step: references/pipeline.md (read it before step 1).

0. Open the project. Find the new video (path in the message, a file dropped into public/, or public/main.mov already replaced). Pick a short kebab-case slug, then node scripts/05-new-project.mjs --slug <slug> [--video <file>] [--assets <dir|files…>] (--dry-run first). It archives the previous finished project to projects/<old>/ (moves, never deletes) and scaffolds work/.
1. Probe + fitness check: npm run probe. Default master path: 3840×2160, HEVC 10-bit, Rec.2100 HLG (iPhone HDR). SDR / 1080p / 30 fps sources go through the SDR branch — read "Different source formats" in references/pipeline.md first.
2. Prepare: npm run luts (only if work/luts/*.cube are missing) → npm run proxy → npm run assets.
3. Transcribe: optional work/transcribe.vocab.txt (names, brands, jargon) → npm run transcribe. Review the raw text; write work/transcript/corrections.tr.json with only high-confidence fixes, list unclear passages under uncertain — never invent speech. Deliverable: out/final.tr.srt.
4. Asset inventory: look at frames of every asset (contact sheets) and write work/asset-manifest.json (what it shows, where it fits, limitations, privacy notes).
5. Edit plan: read references/visual-language.md and references/animation-scenes.md, then split the transcript into chapters (→ chapter tags + the YouTube chapter list), choose the layout (L0–L5) of every passage, one persistent device per chapter, the 2–4 thesis lines, and the 1–3 passages that get an animation scene. Write work/edit-plan.json following references/editorial-style.md (schema: src/plan/schema.ts, format example: library/examples/edit-plan.example.json). npm run plan:check. 5b. Build what the plan needs (new event types / bespoke src/scenes/*): npx tsc --noEmit && npx eslint src && npm run plan:check.
6. Look before rendering: node scripts/65-design-stills.mjs (+ --at <secs>) and node scripts/66-cut-check.mjs --remotion. Read the sheets yourself: panel on the face after a cut → other side or end the event before the cut; text must never shift while on screen; nothing clipped.
7. Music: find the user's track in public/assets/, measure duration + loudness, note it in work/progress.md.
8. Render graphics (detached): npm run render:gfx → check every segment's frame count against the manifest.
9. Master (detached): npm run composite → npm run qc (all-pass; promotes to out/final-hlg.temiz-ses.mov) → read the QC contact sheets of every event (entrance / hold / exit).
10. Sound: npm run audio → npm run qc:mix (all-pass; promotes to out/final-hlg.mov, writes out/stems/). Make two or three --preview a-b clips of the busiest passages for the user.
11. Docs: work/qc-report.md, out/README.tr.md, final work/progress.md with the line TAMAMLANDI (references/delivery.md).
12. Delivery message (references/delivery.md): files, what was added, measurements, what needs their ears/eyes, which music file was used, how to ask for changes.

Typical wall time for a 7–8 min 4K video: ≈ 3–4 h, mostly rendering (first video: +2–3 h while missing components are built). Say so up front.

Vertical videos (Reels / Shorts)

• From a finished long video: scripts/90-shorts.mjs (see references/pipeline.md).
• A pre-cut vertical video (Reel) that only needs zooms, band graphics and SFX: build a vertical composition + script on request (1080×1920, graphics in the band above the head, one block at a time).
• Rules: references/visual-language.md §9 and the vertical part of references/editorial-style.md.

Revisions on a delivered edit

• Text / timing / side / removing an event: edit work/edit-plan.json → plan:check → stills → only the affected segments re-render → composite → qc → audio → qc:mix.
• Sound only (louder music, another effect, "sfx": false on an event): work/audio/selection.json or the plan → npm run audio && npm run qc:mix (~5 min, the video stream is copied, not re-encoded).
• Keep the previous delivery: rename it before promoting a new one; save a copy of the plan (work/edit-plan.revN.json).

Learning the user's taste (this is how the workshop becomes theirs)

• Anything the user corrects twice is a house rule: add it to references/editorial-style.md → "Ev kuralları", one line, with the date and the user's own words in quotes. Tell the user you added it.
• If the user says "bunu kural yap", "bunu hep böyle yap", "bir daha yapma" — add it immediately, once is enough.
• Brand-level changes (colour, font, closing card, tone) go to marka/MARKA.md and src/theme/theme.ts.
• A screenshot the user likes ("böyle olsun") goes to marka/style-refs/ with a line in its README.
• Never copy another channel's look verbatim: learn the grammar from references, keep the user's brand.

New event types

The kit has panels, lists, stats, steps/funnels, bars, search, versus, cards, phone, insert, screen, stage boards, kinetic theses, chapter tags, social and subscribe cards (src/events/*). For a new one: add a zod object in src/plan/schema.ts (timed parts carry at), a component (full-screen ones use the Scene shell in src/events/SceneEvents.tsx), a case in src/compositions/GraphicsLayer.tsx, the new timed keys in timesOf (scripts/50-validate-plan.mjs) and in the cue loops of scripts/75-audio.mjs. Rules: reserve space for everything that appears later (nothing on screen may move), all motion is a pure function of the frame, opaque surfaces in HLG, schematic diagrams carry a "Şematik gösterim" label, no invented numbers.