Under the Hood¶
This section is for people who want to improve the system. Operators don't need to read it — Claude and the specialized agents apply all of these rules automatically when you ask for generations.
Read it if you're:
- An engineer maintaining the agents (
manager,nanobanana-prompter,veo-prompter, etc.) - Adding a new model or a new generation backend and need to know what rules exist
- Diagnosing why a generation came out wrong and want to see what the system should have caught
- Onboarding into the technical side and want to understand what's actually happening in the prompts
If you're a day-to-day operator producing videos, skip this section. Everything below this header is plumbing.
What's in here¶
| Page | Covers |
|---|---|
| AI Tells | Taxonomy of generation artifacts the auto-QA loop scans for (extra fingers, melted faces, garbled limbs, etc.) — and the visual judgment criteria it uses to flag a regen |
| Image Prompt Rules | The full rule set the Image Prompter agent applies when constructing NanoBanana prompts — framing, in-frame rule, POV rule, background motion, character refs, text in images |
| Video Prompt Rules | The full rule set the Veo Prompter agent applies — macro vs. micro motion, location-change semantics, POV constraints, speaking-scene patterns, Veo3 slot requirements |
| Architecture Patterns | The structural patterns every .nbflow follows — dynamic-feeds-template, one-gen-node-per-(template+media), minimal templates, variableName wiring contract |
| Tool & Model Reference | Per-model card for every backend — NanoBanana 2, Veo 3.1, Seedance — with model identifiers, prompting best practices, and known failure modes |
| Pre-generation Sanity Check | The Python validator the Generation Runner executes before any generation pass — catches schema bugs, model mismatches, dangling links, stale per-node refs, and the dynamic variableName mismatch class |
.nbflow Schema |
JSON schema reference for the .nbflow workflow files PatchWork produces and consumes — node types, link tuples, multi-account specifics |
Relationship to CLAUDE.md¶
The canonical agent rules live in manager/CLAUDE.md. The pages here are extended elaborations — they explain why a rule exists, how it's enforced, and what the failure mode looks like when it's violated.
| If you want to | Read |
|---|---|
| Change agent behavior | manager/CLAUDE.md (source of truth) |
| Understand why a rule was added or what it's trying to prevent | These pages (background context) |
| Add a new model or backend | These pages + CLAUDE.md together |
What's NOT in this section¶
- Operator workflows — that's the main chapters (1-14)
- Agent definitions — those live in each agent's own folder under
claude-projects/ - Workflow-specific prompt details — those live in
Assets/{workflow}/per project - Skill scripts — see the Skills section for user-invocable skills