Working in This Pipeline¶
Ground rules that apply across all your work — regardless of what task you're doing. Read once, internalize, return when you're about to do something you're unsure about.
Security & secrets¶
The pipeline uses two pieces of secret material:
G-Labs API key- Authenticates your generation requests. Static across sessions until rotated.
Cloudflared tunnel URL- Public HTTPS endpoint that proxies to your local G-Labs. Changes every session.
What never goes into git¶
- API keys (G-Labs, OpenAI, Anthropic, anything else)
- Personal Cloudflare account credentials
- Service account JSON files
.envfiles
The Manager already knows what .gitignore rules to apply. If you ever find yourself manually adding a file to a commit, double-check it's not a secret.
If you suspect a key was committed accidentally, rotate it immediately, then scrub history (or ask Claude to help with this).
What's safe to share¶
- The
.nbflowfiles themselves (no keys inside — they reference R2 URLs, not raw assets) - Generated images / videos (R2-hosted)
- The tracker spreadsheet (already shared with the team)
- The wiki (it's public — designed to be safe)
Cost awareness¶
Every generation costs real money. Most workflows burn small budgets, but a careless re-run cycle can chew through a meaningful amount.
Where budget goes¶
| Operation | Rough cost shape |
|---|---|
| NanoBanana 2 image gen (4 candidates) | Low per gen — usually the cheapest part |
| Veo 3.1 video gen (4 candidates, 8s each) | Significantly more per gen — the expensive part |
| Reruns | Each rerun is a full gen — bundling 3 reruns of one node is 3× the cost of the original |
| Iteration via prompt-tuning | Each round is its own gen — set a sensible iteration cap |
Cheap ways to validate before spending¶
mode=imagesvalidation pass- Tell the Generation Runner to do an image-only pass before running video. You see all the still frames cheaply; if they look right, then run video. Catches "the avatar's pose is wrong" before paying for the video gen. Use this whenever a workflow is new or substantially changed.
Single-node test before full fan-out- When fanning out to a new account, run one test gen in that tab before committing to the full workflow run.
Skip aggressive regen attempts when not needed- The Generation Runner auto-reruns flagged images up to 3 attempts. If you're just doing a quick validation, you can cap this lower.
Full detail in Chapter 12 — Workflow Optimization.
Pipeline limits¶
Things to know that you'll bump into eventually:
| Limit | What to know |
|---|---|
| Generation concurrency | Default 3 parallel, cap 5. Going higher doesn't help — upstream APIs throttle |
| Veo clip length | ~8 seconds per clip. If you need longer, split into multiple clips and stitch in post |
| R2 storage | Generous quota; we're not close to it. But don't dump huge raw video files there unnecessarily |
| PatchWork file size | Very large .nbflow files (1000s of nodes) start to slow down the web app — split workflows if you're approaching that |
| API rate limits | Soft caps on requests-per-minute. Concurrency=3 stays under these comfortably |
Working efficiently¶
Focus over juggling- Working on 5 workflows in parallel sounds productive but isn't. Each context switch costs you state recovery time. Finish one through to a clean handoff point before starting another.
Use the tracker- When you genuinely need to track multiple workflows, the master Google Sheets tracker is the north star. It tells you what's testing, what's approved, what version is current. Sync it (
tracker-syncskill) after meaningful changes so it doesn't drift. Bump versions, never overwrite- Every meaningful change to a workflow bumps the V0-N counter and archives the previous version. This protects you when something breaks. Detailed rules in Chapter 3 — Lvl 1-2 Variants and Chapter 5 — Lvl 3-4 Variants.
Run the sanity check before every Generation Runner pass- The pre-generation sanity check catches schema bugs that would otherwise burn a Generation Runner pass. Always run it. It takes seconds.
What to do when something goes wrong¶
The fast answer: jump to Chapter 13 — Troubleshooting. It has symptom → cause → fix lookups for the most common failures.
The slower answer: read the agent's response carefully. Most failures emit a useful message. Don't retry blindly — understand what went wrong before re-running.
When you're ready¶
→ Next: Glossary — every term defined. Don't read it cover-to-cover; bookmark it for when you encounter a word you don't recognize.