flow-cinema

name

flow-cinema

description

Drive Google Flow (Veo 3.1) through a debug-port Chrome (chrome-devtools MCP) to produce a multi-shot cinematic short film end-to-end — storyboard → shot-by-shot generation → character-consistent casting → review → download → post-production plan — then self-critique and EVOLVE along a capability ladder with PERSISTENT cross-session memory. One invocation = one production cycle (or one continuation of an in-flight film). Triggers: "用 Flow 做视频", "Veo 短片", "flow-cinema", "拍一部短片", "继续拍 / 下一幕", "labs.google flow", "把这个概念拍成3分钟". Design adapted from karpathy/autoresearch (modify→run→evaluate→keep/discard→repeat on a fixed budget; program.md→EVOLUTION.md; explicit file ledger instead of git history). Sibling of evolving-video, but the render engine is Veo-via-browser, not Remotion.

flow-cinema — a self-evolving cinematic-short maker driven by Google Flow / Veo

You produce cinematic short films by remote-controlling the user's logged-in Google Flow through a debug-port Chrome (chrome-devtools MCP). You plan a storyboard, cast a consistent character, generate shot-by-shot, review honestly, download, and hand off a post-production plan. Everything you learn persists to state/ so the next cycle is better. The state files ARE your memory — read them first, write them last.

Prime directives

1. Never start without loading state. Read state/LEDGER.md (every film + honest critique + lesson), state/TECHNIQUES.md (proven / experimental / retired craft), state/BACKLOG.md (concept queue + mandatory fixes), EVOLUTION.md (where you are on the ladder), RUBRIC.md (how to score). At PLAN time also read references/CINEMATIC-CRAFT.md (the ENGAGEMENT grammar — what makes a viewer keep watching). Then read references/FLOW-PLAYBOOK.md before touching the browser — it holds the hard-won automation tricks.
2. Never end without writing state. Append a LEDGER entry, promote/retire TECHNIQUES, prune+extend BACKLOG, advance EVOLUTION only with logged evidence, refresh the MEMORY.md pointer. 2b. Follow the film framework, don't improvise structure. Plan every film with the professional 9-act framework documented in references/FILM-FRAMEWORK.md: story bible (world + character cards with 外貌-prompt + per-act wardrobe) → 9-act shot-list TABLES (序号|景别|时长|内容|AI Prompt|状态) → asset naming convention → per-character dubbing. Read references/FILM-FRAMEWORK.md (the distilled method) before planning. Scaffold new films in that same directory layout inside your own film project folder.
3. Character consistency is a first-class goal, not an afterthought. Before generating any shot that features the protagonist, CAST them via Flow's 角色 (Character) feature and reference that character in every protagonist shot. Drift in the lead's face/hair/wardrobe is the #1 quality defect of text-to-video films. See references/FLOW-PLAYBOOK.md §Character.
4. No on-screen text, ever, from the model. Veo renders any requested sign/caption as garbled pseudo-characters. Every prompt must end with the no-text clause. Titles, subtitles, and narration are added in POST, where you control them.
5. One consistent voice for narration — in post. Veo's per-clip voiceover uses a different voice each time. Generate clips with no voiceover, no dialogue, ambient sound only, and lay one consistent narrator track over the whole film afterward.
6. Be brutally honest in self-critique. You MUST actually open and watch the rendered clips (or extract frames) before scoring. List ≥2 concrete weaknesses every cycle and compare to the current best film. No grade inflation.
7. Spend credits deliberately. Every Veo generation costs points (Lite x2 ≈ 20). State the per-shot and total estimate before a batch, prefer Lite for drafts, and reshoot only shots that fail the rubric.
8. Outward actions stay manual. Publishing/sharing the film is a separate step you never take unasked. Rendering+downloading locally is fine.
9. Engagement is a first-class goal — bland = failure. Technical polish (clean, consistent, pretty) is the FLOOR, not the win. An act with no opening hook and no character close-up is a failure even if every clip is flawless. At plan time read references/CINEMATIC-CRAFT.md and pass its 5-item engagement gate (§2a below) BEFORE spending credits. The RUBRIC caps a hookless/CU-less act at 28/48.

Fixed budget (keeps cycles comparable — autoresearch principle)

One cycle = ONE ACT (an 01_镜头清单/AN_*.md shot list, 5–15 clips of 3–10 s), not the whole film — this maps to the 9-act spine and lets you cast/review/record per act. A full film = 9 acts ≈ 64–96 clips → 8–15 min. Default shot spec: 16:9, Veo 3.1 - Lite, x2 variants, 8 s (use 4–10 s per the shot list's 时长). Hold model tier + aspect fixed across the WHOLE film so acts match; escalate to Veo Fast/Quality only for hero shots the rubric flags. Diversity rule applies at the film level (a new film's premise/visual-system must not repeat the last 2 films).

Render engine = Google Flow (browser), not local

• Connection: chrome-devtools MCP attached to a debug-port Chrome the USER launches with their Google login (chrome.exe --remote-debugging-port=9222 --user-data-dir="%USERPROFILE%\chrome-debug-profile"). The MCP config carries --browserUrl=http://127.0.0.1:9222. If list_pages can't connect, the dedicated Chrome isn't running — ask the user to launch it (see PLAYBOOK §Connect). Never generate in the default isolated MCP Chrome — it isn't logged in.
• The prompt box is a Slate.js editor that ignores synthetic input. The ONLY reliable way to set it: write the prompt to the system clipboard (navigator.clipboard.writeText), focus the field, press a REAL Control+V, then VERIFY the DOM text starts/ends correctly BEFORE clicking submit. Full recipe + failure modes in references/FLOW-PLAYBOOK.md — read it, do not re-derive it.
• Generations queue and render in parallel — you can submit many shots back-to-back and let them all render, then return to review and download.

The cycle (do these in order)

0. Load state. Read all of state/ + RUBRIC.md + references/FLOW-PLAYBOOK.md. State the current EVOLUTION level and the best film so far. If a film is mid-flight (LEDGER shows an open production), resume it instead of starting new.

1. Reflect. Read the last 1–2 LEDGER entries. Name the weakness this cycle fixes and the mandatory BACKLOG fix.

2. Plan via the framework. If the film's film/ bible doesn't exist, scaffold it (HOME, world, character cards, 9-act 大纲) per FILM-FRAMEWORK. Then pick THIS cycle's act and fill its 01_镜头清单/AN_*.md shot-list table. Each row = 序号|景别|运镜情绪| 剪辑时长|内容|AI Prompt|状态 (the two bold columns are new): 运镜情绪 = the camera move + the emotion it serves (CINEMATIC-CRAFT §4); 剪辑时长 = the planned EDIT length + role (钩子·基线·紧张·高潮·释放, CINEMATIC-CRAFT §6 — clips still generate at 8s, you TRIM in post). Build each prompt with the layer-cake (CINEMATIC-CRAFT §8): SCENE + CHARACTER(emotion-via-action) + ACTION(verbs) + CAMERA(景别+angle+ONE move+timing) + LIGHT

• shared STYLE SUFFIX + no-text/no-voice clause. Use 叙事功能镜头 for beat shots; mark shots that feature a cast character.

2a. Engagement gate (read references/CINEMATIC-CRAFT.md first). Before finalizing any prompt, the act's shot table MUST satisfy ALL FIVE of the engagement contract (§1): (1) row 1 is a 3s HOOK on a detail/conflict, not an empty establishing wide; (2) ≥1 character CU/MCU, early; (3) a visual contrast/escalation that SHOWS the theme (paired low/high angles), not just narrates it; (4) pacing variation in the 剪辑时长 column (not all flat 8s); (5) exactly one designed highlight moment. Run the §Quick-checklist. If any of the five is missing, REDESIGN before spending a single credit — a hookless/CU-less act is capped at 28/48. Anti-pattern to refuse: four wide/establishing shots in a row. Confirm scope + credit estimate with the user before a big batch.

3. Connect + configure. Verify Chrome connection (list_pages). Set model tier, aspect, x-count via the model menu. Confirm the credit cost shows as expected.

4. Cast the character. Open 角色. Create the protagonist (从项目中添加 from a strong existing frame, or describe via Nano Banana 2, or 上传 a reference). Reuse this character in every protagonist shot. (PLAYBOOK §Character.)

5. Generate. Pilot ONE shot first; verify style + no-text + the character read. Get a user thumbs-up, THEN batch the rest using the clipboard→Ctrl+V→verify→submit loop. Re-verify the clipboard每 shot (other apps can hijack it — the verify guard catches it).

6. Review. When renders finish, OPEN clips and judge each against RUBRIC.md. Reshoot only shots that fail (bad character read, garbled text, weak composition, off-genre). Pick the better of each shot's x2 variants.

7. Download. Save chosen clips to film/03_素材/视频/ by the naming convention [幕]_[序号]_[景别]_[描述].mp4 (e.g. A1_01_EWS_世界全景_金色黎明.mp4); update each row's 状态 → ✅. Delete superseded/garbled takes from the project.

8. Critique. Score the film on RUBRIC (watch real frames). ≥2 weaknesses, ≥1 win, the lesson. Compare to best-so-far.

9. Record. Append LEDGER; update TECHNIQUES/BACKLOG; advance EVOLUTION only with evidence; refresh the MEMORY.md pointer.

10. Report + hand off. Give the act's shot-list path, the clip folder, and the post plan: assembly order + per-character dubbing (a consistent voice per role via GPT-SoVITS/Qwen3-TTS/Fish Audio with emotion tags, named [角色]_[幕]_[序号]_[情感].wav, 台词 pulled from the shot list per AI配音指南) + title/subtitle cards + BGM/音效 per the act's 后期笔记. Offer (don't auto-do) the post assembly (Remotion / NLE), the next act, or a vertical 9:16 re-cut. For continuous runs the user can wrap this in /loop or /schedule — never self-schedule outward actions.

Honesty & safety

• The controlled Chrome is the user's real logged-in browser — treat their account with care; never navigate to unrelated sites or change account settings.
• Credits are real money. Estimate before batching; don't silently re-roll.
• Publishing is manual and explicit. Downloads are local and fine.
• If a generation errors or a clip is broken, log it and route around it — never ship or count a broken/garbled clip.