---
name: gpu-onboarding
description: >
  First-run onboarding: when a user is new to the video-factory GPU setup,
  walk them through the H3 production economics (Turbo 4-8 step workflow,
  provider lanes), ask four questions about their workload, and make the first
  deployment. Use when .pi/gpu/providers.json is missing or empty and the user
  hasn't chosen a lane, or when they ask which provider/GPU to use.
---

# GPU Onboarding (Claude Code)

Goal: a new user goes from "I installed this" to "my first video is queued"
without reading a manual — and picks a provider from economics, not brand
recognition.

## Step 1 — The one-paragraph context

Say, in plain words:

> H3 generates up to 15 seconds at 2K with audio in one pass, so a "minute" of
> video is four clips stitched together. The big cost lever is the workflow:
> Turbo LoRA 4–8 steps + SageAttention + INT8 runs 4–10x faster than the
> default 20-step setup, which is what most public "slow H3" numbers come from.

## Step 2 — The decision card

Full detail lives in `docs/h3-economics.md` in this repo. Present:

```
H3 PRODUCTION — planning numbers (2026-08-12 snapshot, verify live)
  1 finished minute = 4 × 15s clips (H3 maxes at 15s @ 2K with audio)
  anchors: RTX 6000 480p/15s = 56s @ 4-step Turbo · 5090 15s/0.4MP ≈ 3 min
  rule: Turbo LoRA 4-8 step + SageAttention2 + INT8 — never vanilla 20-step
  per finished minute: Salad 5090 ~$0.06 · Vast 5090 ~$0.14 · RunPod PRO 6000 ~$0.11
  $15 buys: Higgsfield 1 min · Vast 5090 ~100 min · Salad 5090 ~250 min
```

## Step 3 — Four questions (then stop talking)

1. **Batch size** — finished minutes per batch? (30+ min → keep the instance
   alive between jobs; small batches → destroy aggressively.)
2. **Quality floor** — 0.4 MP quick pass or full 2K? (4-step Turbo = fast lane;
   8-step = quality; 2K needs the big cards.)
3. **Budget ceiling** — write it into `.pi/gpu/policy.json`, don't just quote
   it. Defaults: confirm above $1, per-job $5, daily $40, monthly $400.
4. **Provider lane** — present the three, don't make them pick blind:
   - **Vast 5090 (~$0.69/hr)** — best balance; the default recommendation.
   - **Salad 5090 (~$0.29/hr batch)** — ~2x cheaper, distributed fleet,
     availability caveats; the aggressive lane for retryable batches.
   - **RunPod PRO 6000 ($1.69/hr)** — the measured 56s/4-step anchor card;
     most predictable operations; choose for polish over price.

## Step 4 — First provision (all via bash — see provider-adapters skill)

1. `gpu-setup` flow if CLIs are missing.
2. Decide model: `minimax-h3` for the flagship lane; for large cheap batches
   offer `wan-2.2-5b` on a 24GB card as the draft lane and H3 for finals — the
   split is often the cheapest production setup.
3. Search live offers → provision → install ComfyUI + weights → **benchmark
   the exact workflow** (compare to docs/h3-economics.md anchors) →
   health-check with a test generation → write the deployment to
   `.pi/gpu/deployments.json` (status ready) → queue the job.
4. Record measured per-clip time + cost: append a job to `.pi/gpu/jobs.jsonl`
   and note the benchmark in the model manifest (`.pi/gpu/models.json`).

## Step 5 — The one-line close

> Everything is recorded in .pi/gpu — tomorrow the agent will know what's
> deployed, what it costs per minute, and what's queued.

## Variants

- No CLIs: run gpu-setup first, then return here.
- User has a provider already: skip the lane explanation, ask budget + quality,
  deploy on what they have.
- User wants zero infra: point at the hosted options in docs/h3-economics.md
  (Higgsfield / official H3 API / cheap hosted ~$0.32 per 15s clip).