A plain-English, read-only map of the whole video factory — what's happening, whether it's
healthy, and which file to touch to change anything. Nothing on this page
spends credits, writes, or deletes. It only looks.
🏭 Pipeline Map — the factory floor
Every stage your show passes through when someone clicks Generate, top to bottom in the order it happens.
How to read this
Top to bottom is the order things happen. Green "AI" tag = a model is thinking (costs money, can be creative or wrong).
Grey "deterministic" tag = plain code (free, predictable, can't hallucinate). If a show came out wrong, find the stage that
owns that thing — bad dialogue lives in the script stage; wrong faces live in the render stage. Click any stage to expand it.
🎬 Clip-Gen Deep-Dive — why the faces stay consistent
Our "secret sauce" explained as a film-crew story. You don't need this day-to-day — it's the answer to "why does AI video look stitched and ours doesn't."
The 11 steps
💬 Prompt Library — the words we tell the AI
Every personality, rule and style the show has comes from one of these prompts.
How to read this
Want the dialogue funnier or the posters darker? Find the prompt here, copy its file path, and tell Claude Code to open that file.
Anything marked Legacy is old code wired to nothing — leave it alone.
Prompt
Stage
Model
What it controls
File
🧠 Model Clusters — which AI does what, and what it costs
This is the bill, by vendor.
How to read thisCONFIRMED = the price is real and tracked. ESTIMATE = it's a guess, or it isn't being
counted at all. The big amber warnings mean the Cost page total is lower than reality — don't quote it as the final number.
Only the name of each secret key is shown here, never its value.
📡 Live Jobs & Logs — the heartbeat
Everything the engine is doing right now, newest first, across all customers. Refreshes every few seconds on its own.
How to read this
A green running row with a moving log = healthy. A red
failed = click it and read the last few lines.
Jobs live in memory — if the engine restarts, anything still running vanishes and the customer is refunded, so don't restart while a row is green.
System memory
Checking…
When this is high, starting a new heavy job risks running the machine out of memory. Read-only — it just reports.
Status
Job
Who / Show
Started
Duration
Log
Loading jobs…
🌿 Git / Work Tree — what changed recently
A diary of code changes for both halves of the product, side by side.
How to read this"Working tree clean" = everything is saved to history. "Uncommitted changes" = someone (maybe a Claude session) edited files
that aren't saved yet. The two repos must move together when you change the seam between them — this page lets you eyeball both at once.
Loading git status…
💾 Storage & Recovery — where the shows live
Where every show actually is, and the story of the ones that scared us.
How to read this
The "lost" shows were never lost — the app was looking in an empty folder. This page shows where every show actually is.
If it ever says a store is missing, the matrix below tells you where the original still lives.
Where each show exists
Each row is one show. A check (✓) means a copy lives in that store; a dash (—) means it doesn't.
iCloud = the master originals; Scratch = the working folder on the machine;
R2 = cloud backup; Supabase = the database record.
Show
iCloud (originals)
Scratch (working)
R2 (cloud)
Supabase (db)
🛠️ How to Change X — the decision desk
Start here whenever you want to change something but don't know where it lives.
How to read this
Find your goal, see the exact file, hit Copy instruction for Claude Code, and paste it into a session. The red
BOTH repos · lock-step cards are the dangerous ones — those need the engine deployed first.
📖 Glossary — jargon decoder
Whenever a word in this dashboard (or in Claude's reply) is unfamiliar, it's defined here in one sentence.