Film-trained AI production control for video, agents, and non-engineer product work.
I translate director judgment into source selection, AI video shot packages, agent handoff, retake gates, artifact review, and lesson writeback for non-engineer operators.
我把导演判断、影像生产、AI coding agent 和产物验收接成一条可审查的控制链。
Beijing Film Academy photography/directing training gave me the image, narrative, rhythm, continuity, and field judgment. Four months of intense vibe-coding turned that judgment into AI-native execution surfaces: what source an agent should read, what a tool output is allowed to prove, when a visual artifact needs a retake, and what can be written back as reusable method.
Public site · Start in 5 minutes · Methods · Templates · Cases · Examples · Roadmap · Changelog · Contributing · License · Open a public-safe issue
| If this is the failure | Start here | Copy this |
|---|---|---|
| AI video shots look fine alone but break as a sequence | Filled synthetic shot package | AI video continuity locks |
| A director note like "make it oppressive" cannot become agent work | DirectorShotIR teardown | DirectorShotIR field card |
| A non-engineer product idea is too messy to code or hire against | Project Clinic teardown | Project Clinic startup packet |
| Public pillar | What I make concrete |
|---|---|
| AI video production control | Script intent, visual reference roles, shot/package logic, continuity locks, QA boundaries, retake notes, and low-context operator handoff. |
| Director-style agent workflows | Turning performance, blocking, camera, transition, asset roles, review rules, and retake targets into agent-readable work. |
| Project Clinic for non-engineers | Turning a messy product/workflow idea into source selection, capability route, first task packet, review checklist, and outcome-delta note. |
| Video admission and verification | Separating raw artifacts, static evidence, human playback, relation checks, owner seal, and final acceptance boundaries. |
I started from Beijing Film Academy photography/directing. The point is not to pretend I am a conventional full-stack engineer with a film hobby. The useful edge is translation: image, narrative, rhythm, continuity, taste, and on-set operator reality becoming AI-native work systems.
Many people can use an AI coding agent casually. The harder layer is knowing:
- what source the agent should read first;
- what a tool output is allowed to prove;
- when a generated artifact is only candidate evidence;
- how to turn a creative failure into a field-level retake;
- what should be written back as reusable method rather than fake project truth.
That is the public lane here.
| If you are facing | Do this first | Then copy |
|---|---|---|
| AI video clips look fine one by one but fail as a sequence | Read the filled synthetic shot package | AI video continuity locks |
| A director note like "make it oppressive" cannot become agent work | Read the DirectorShotIR synthetic teardown | DirectorShotIR field card |
| A non-engineer product idea is not ready for coding or hiring | Read the Project Clinic synthetic teardown | Project Clinic startup packet |
| A workflow has demos, screenshots, reports, and no current truth | Start with the workflow rescue map | Outcome-delta writeback note |
| A video artifact exists but nobody knows if it is reviewable | Use the video admission ladder | AI video continuity locks |
The fastest route is: read one filled synthetic example, copy one blank template, fill only public-safe or owner-approved material, then open an issue only if the question can be discussed without private assets.
The pixel identity is the public face. The diagram below is the operating model: intent becomes source selection, then agent handoff, then artifact review, then bounded retake or writeback.
These case tracks expose method and boundaries, not private assets or final project truth.
| Case | Public value |
|---|---|
| Script to dispatchable AI video packages | How script intent becomes visual reference roles, image evidence, package review, retake logic, and operator handoff. |
| Filled synthetic script-to-package sample | A complete public-safe example using an invented scene, not private script or production material. |
| DirectorShotIR and crew-style agent work | How one vague "AI brain" becomes role-based creative responsibility: director, art, camera, critic, QA. |
| Project Clinic for non-engineers | How one messy request becomes a source bundle, first task packet, review checklist, and reusable lesson. |
Use this repository as a public reference for:
- AI video continuity review before more generation spend;
- director-style agent task packets instead of vague prompts;
- non-engineer project startup packets before hiring or coding;
- video admission gates that stop weak evidence from becoming final claims;
- public-safe templates you can copy into private owner workspaces.
If a template or synthetic example saves you a false start, a star is a useful bookmark. Stars are not proof of product quality, client outcome, or final artifact acceptance.
- Case before tool: a tool is worth naming only when it changes a real production, product, or verification decision.
- Evidence before claim: model output, sidecar advice, dry-runs, screenshots, and fixtures stay candidate material until the right owner accepts them.
- Director judgment into operator language: taste, continuity, rhythm, and camera intent become inspectable instructions, not vague prompts.
- Boundary first: private project material stays in its owner workspace. The public layer shows method, synthetic examples, templates, and review contracts.
Useful conversations usually start with one of these:
- an AI video or short-drama workflow that keeps failing at continuity, rhythm, asset control, or review;
- a non-engineer product idea that needs a source/tool/agent execution route before hiring engineers or opening a large build;
- an AI workflow with many artifacts but no reliable acceptance boundary.
Use the issue templates in this repository for public, method-level discussion. Do not upload private scripts, raw media, provider screenshots, account state, customer material, or final project acceptance truth to public issues.
Read CONTRIBUTING.md before opening a pull request. The accepted contribution types are intentionally narrow: synthetic examples, template clarity, broken-link fixes, and public-safe method language.
These repositories are supporting slices. They are not the headline. Manual profile pinning guidance: GitHub profile pinning.
Reference repos and what they prove
| Repository | Use this if | Public status | Not claimed |
|---|---|---|---|
mmi-gateway |
You need an intake pattern for turning messy material into review-required candidate evidence. | Reference slice | Not the current product identity. |
codex-sidecar-subagents |
You need the read-only advisor pattern where Codex keeps file access, verification, and integration judgment. | Reference slice | Not autonomous truth or final acceptance. |
epistemic-os |
You need claim-scope guardrails so weak evidence does not become confident AI release language. | Guardrail slice | Not a complete runtime system. |
netfix |
You need a practical Mac network rescue utility for operator downtime. | Utility satellite | Not part of the AI video method library. |