delwerhossain/carvendors-video-studio

CarVendors Video Studio v4

Talk to the user in short, simple Bangla. Everything in the videos is UK English. $CS = this skill folder (~/.claude/skills/carvendors-studio, Windows %USERPROFILE%\.claude\skills\carvendors-studio). $PY = the skill's Python: Mac ~/.cache/carvendors-studio/venv/bin/python, Windows %USERPROFILE%\.cache\carvendors-studio\venv\Scripts\python.exe (made by the install; has numpy + scipy).

0. Rules that never bend

1. No invented facts. Price, mileage, engine, year, "new arrival", "sold", finance, economy, plates only with evidence. Old TikTok text and comments are not evidence.
2. Real stays real. Real footage and real showroom photos only; no AI-made cars, people, showrooms, signs.
3. Brand: CarVendors.co.uk is the main brand; "Featured dealer: B&F Cars" is secondary. CarVendors helps people find/connect with cars; it does not sell the car. Both are our companies.
4. Full video, balanced length: keep the WHOLE source in order (no scene cuts). Add one hook and one outro. Total ≈ source + 3-7 s. Trim only black/frozen/duplicate frames.
5. Cheap by design: use the fast lane (batch.py + build.py). Do not hand-write composition HTML, do not re-analyse hooks (the shared catalog has them), look at ONE contact sheet per video. Ask only when blocked.
6. Never print or copy API keys (~/.config/video-studio/.env). Never overwrite an export (_v2, _v3…).
7. Unsure where something is or what the user means → check $PY $CS/tools/paths.py show and the CarVendors Drive folder (paths.py get carvendors — it holds all context: sources, logo, showroom photos, banners, customer reviews, text-style previews), then ask one short question.

1. Where things are (automatic, Mac + Windows)

$PY $CS/tools/paths.py show finds this PC's Google Drive "My Drive" and reads the team config My Drive/Projects/CarVendors/_studio/config.json (paths relative to My Drive + team defaults). No setup when Drive is running. Overrides for one PC: paths.py set <key> <path> / paths.py drive <My Drive path>. Keys: car_sources (folders 1-200 with video.mp4, comments.txt, sometimes Voice/), hooks, bgm, sfx, fonts, logos, showroom_pics, text_refs, outros, exports (finished videos for upload), studio (shared catalog + ledger), projects_root (local job folders, ~/Projects/videos).

2. The bulk loop (default)

$PY $CS/tools/batch.py plan 1-20      # job dirs ~/Projects/videos/bnf-vNNN + draft job.json + contact sheet + text boxes
  → for each video: look at work/qa/contact.jpg, edit job.json (§3), set "ready": true
$PY $CS/tools/batch.py render 1-20    # builds + renders every ready job, exports to Drive, writes the hook ledger
$PY $CS/tools/batch.py status 1-200

One video by hand: $PY $CS/tools/build.py <job.json> (check only) → optional npx hyperframes snapshot <job>/work/hf --at <t> → build.py <job.json> --render. Work in batches of 10-20; after each batch look at the frame sheets (output/*_frames.jpg) before the next.

3. What to decide per video (the only AI work)

From the contact sheet + source_info + text_hint in job.json:

• slug: short model name, e.g. mitsubishi-asx.
• Text look: follow references/text-style.md exactly (the preview images in text_refs): white Montserrat Black text, ONE accent = brand red #E8202A pills/tags, white cards with charcoal text, green only for tiny dots/ticks, logo clean top-centre. Never charcoal/black boxes, pills or badges on footage.
• Old text (each text_hint box, src = source px, canvas = 1080x1920 px), cheapest good option: keep (harmless, readable, on-brand, e.g. a plain model name) → cover with covers[] (text-cover, style card = white card + red kicker tag, on canvas coords, ALWAYS with a line, e.g. kicker "FEATURED" + "Spotted at B&F Cars") when it shows prices, "TikTok", other businesses, rude or unverified claims → remove (remove[] = src box, inpaint) only if a cover would hide the car. Never an empty cover.
• Overlays (0-3, key beats only, never over the car's face/badge/wheels): label-card on the first shot · punch-text ("Iconic Abarth / [Style]", "Perfect for / [City Driving]", "Sporty / [Interior]") on detail or interior shots · hero-model (big "ABARTH / 595") when the sky is empty · detail-line "In the / details." on a close-up. Pill = brand red always. On white cloud / bright windscreens: move the text or lineBg 0.35-0.5 (soft glow, no box). Times are seconds from the start of the main video.
• hook: leave "auto" (least-used, not used in the last N edits). "car_hook": true for the Car Hooks folder. A named hook: {"file": "Car Hooks/Car Flip.mp4", "in": 0, "out": 3.2} — whole action, cut only dead time after the payoff. transition: cut (default) · whip · flash · zoom — exactly one.
• voice: only if ready_voiceover exists and the user wants it; transcribe and check facts first. Everything else (logo badge, music, outro, grade) comes from team defaults; override only for a reason.

Full field list: docstring of tools/build.py. Example: examples/job-v002.json.

4. Shared hooks, outro, music (same on every PC via Drive)

• $PY $CS/tools/hooks.py list|stats|pick <id>|sheet <name>|scan — catalog of all hooks (length, action peak, suggested in/out, preview strip) and usage counts from _studio/ledger/*.json. Run scan after adding hooks.
• Outro: put the team outro video in the outros folder and set defaults.outro_video in _studio/config.json; every new build uses it. Until then: real showroom photo outro (defaults.outro_photo).
• Music: defaults.music list in the config (library tracks, licence checked, credit line in the post). Logo: defaults.logo (logo-badge, clean style, top-centre — the reference look).

5. Before saying a video is done

build.py reports hyperframes check — it must say "Check passed" (fix contrast by position or lineBg). Look at the frame sheet: old text gone or covered, no text on the car, logo readable, hook → main cut clean. Report in Bangla: output path(s), hook used, what was covered/removed, any blocker. Give UK-English post copy when asked (include the music credit line for NCS tracks).

6. Read more only when needed

• references/text-style.md — THE text look (palette, type, the seven elements, placement). Read before overlays.
• references/components.md — every component's variables (label-card, punch-text, text-cover, logo-badge, hero-model, detail-line, shape-text, verified-feature, caption-line, localized-contrast, outro-real-showroom).
• references/design.md — full design system, colour, type presets (PUNCH / CINEMATIC), old-text ladder, human-editor finish, colour grading, brand/claims.
• references/audio.md — voice (ElevenLabs tools/tts.py, model chosen per video, pre-approved for this work), music/SFX, loudness. references/quality.md — quality gates and history.
• New ads from scratch (story + AI scenes + voice): plan with the user first; same rules, build.py for assembly.