Skip to main content
skill CX Codex

flux-3-generate — Claude Code skill: what it does, how to install

flux-3-generate takes a FLUX 3 video API job from submit to a checked MP4, including polling and drafts. Use it when you render video through the API.

MIT 185 installs ★ 138 updated 2026-09-09

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add black-forest-labs/skills --skill flux-3-generate -a claude-code
Open on GitHub

What flux-3-generate does

The skill walks one job through the asynchronous FLUX 3 video API: build the request, submit it once, poll, download, check the file. Each request sets a mode. t2v works from text, i2v pins your keyframes as frames, v2v extends a start_video, and draft_enhance re-renders a cached draft at full quality from draft_cache.

The schema is strict, so an unknown field or a missing required one returns 422. Polling runs every 6 seconds. Pending, Reasoning and Generating mean wait, Ready is success, and Error, Request Moderated, Content Moderated and Task not found are final failures. A 429 during polling means the task still exists, so never resubmit: you would pay for a second render. Signed result URLs last about an hour. Download every entry in result.samples and result.draft_caches right away, not only the singular result.draft_cache.

Set BFL_API_KEY. It goes out in the x-key header, and the skill checks that the variable exists without printing it. The last step checks the MP4: it decodes, duration, resolution and aspect ratio match the request, and audio is there only when you asked for it.

What it helps with

  • submit a FLUX 3 video job and poll it to Ready
  • turn still images into pinned frames with i2v
  • extend an existing clip from its last frames with v2v
  • try a cheap draft first, then re-render only the one you picked
  • save every draft_caches file before the signed links expire

Example prompts for the agent

  • Submit a t2v job to the FLUX 3 video API and poll it until it's Ready
  • Make a draft with draft: true, download all draft_caches, then check the MP4
  • Extend my clip with v2v using start_video and save the task id and polling_url

How to install flux-3-generate

1Skills CLI (Vercel)

npx skills add black-forest-labs/skills --skill flux-3-generate -a claude-code

2Manually: put the files in place

~/.claude/skills/flux-3-generate/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/flux-3-generate/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about flux-3-generate

How do I install flux-3-generate in Claude Code?

Run npx skills add black-forest-labs/skills --skill flux-3-generate -a claude-code. The skill lands in ~/.claude/skills/flux-3-generate/SKILL.md.

What does flux-3-generate do?

It submits a job to the FLUX 3 video API, polls its status, and downloads the MP4 and any drafts. Then it checks the file technically: decoding, duration, resolution and audio.

What should I do when the API returns 429?

A 429 on submit means no task was created, so wait for a free slot and send the same body again. A 429 on polling means the task exists, so keep polling the same polling_url and do not resubmit.

Source text

SKILL.md · MIT
---
name: flux-3-generate
description: Use when submitting or polling a FLUX 3 video API job. Also covers drafts, downloads, and technical validation.
metadata:
  author: Black Forest Labs
  version: "1.0.0"
  tags: flux, flux-3, bfl, video-api, generation, polling, drafts, qc
---

# FLUX 3 Generate

Execute a ready plan through the asynchronous API, preserve expiring artifacts, and
verify the MP4. Read the [FLUX 3 API reference](https://docs.bfl.ai) before building a
request; it is authoritative wherever this skill disagrees. Authenticate with the
`x-key` header from `BFL_API_KEY`; check the variable exists without printing it, and
never put keys in prompts, logs, or saved request bodies.

## Request shape

Every request names its `mode` (required, exact strings) and carries the matching media
field:

| `mode` | Media field | What the model does |
| --- | --- | --- |
| `t2v` | none | Generates the clip from your text alone |
| `i2v` | `keyframes` (required) | Puts your images on screen, pixel for pixel, as pinned frames |
| `v2v` | `start_video` (required) | Continues from the final frames of your clip |
| `draft_enhance` | `draft_cache` (required) | Replays a cached draft at full quality; accepts no other fields, not even `prompt` |

- The schema is strict: any field it does not know returns a `422` naming it
  (`Extra inputs are not permitted`). A missing required field 422s the same way.
- Settings on the generating modes: `aspect_ratio` (`auto`, `21:9`, `2:1`, `16:9`,
  `4:3`, `1:1`, `3:4`, `9:16`), `duration` (integer 5-20 or `"auto"`), `resolution`
  (`hd` or `fhd`), `generate_audio` (bool), `draft` (bool), `version`.
- Constraints couple (a video input at higher resolution may cap duration): read the
  reference rather than assuming fields are independent.
- Three or more bare keyframes, and all `[seconds, image]` timestamped keyframes, need
  an integer `duration`. Media travels as a public URL or inline base64; keyframe
  images are at least 256x256.

## Run a job

1. **Submit once.** Persist the returned task `id` and `polling_url` with the
   secret-free request.
2. **Poll every 6 seconds** to a terminal state with the `x-key` header, following
   redirects. Planning time on a complex brief is normal, not a stall. The API status
   strings are exact. KEEP POLLING on `Pending`, `Reasoning`, `Generating`. STOP on
   `Ready` (success) and on `Error`, `Request Moderated`, `Content Moderated`, or
   `Task not found` (failures). None of these terminal statuses ever later becomes
   `Ready`.
   - `429` on submit means your org's concurrency ceiling is full and **no task was
     created**. There is nothing to poll. Wait for a slot, then resubmit the same body.
   - `429` on poll means the task exists and is unaffected. Back off and keep polling
     the same `polling_url`. Never resubmit, or you pay for a second render.
   - `503`: bounded backoff.
3. **Download immediately.** Observed signed sample and draft cache URLs expire roughly
   one hour after the result is ready. The `se=` query parameter on the URL itself is
   the authoritative expiry. Read it rather than assuming a fixed window. Video
   outputs arrive in `result.samples`, which is a list. Draft bundles arrive in
   `result.draft_caches`, which is also a list. The singular `result.draft_cache` is
   the first element of that list, not the complete set. Iterate every element of both
   lists and download all of them before doing anything else, recording the local path
   and task ID per element. A client reading only the singular `result.draft_cache` key
   silently drops paid artifacts that cannot be recovered once the signed URLs expire.
   Save the downloaded files, plus task ID and status.
4. **Validate**: container decodes, duration/resolution/aspect match intent, audio
   present exactly when requested. Technical validity is not creative approval; hand
   review to the specialist that owns it.

## Drafts

Add `draft: true` while a concept is unproven: a fast low-step preview plus a
`draft_cache` costs far less than a full render. Judge a draft on event legibility,
composition, and continuity; softness is the low-step trade and resolves at full
quality. Enhance only the chosen cache (it replays the same generation, so preserve
draft-to-cache identity when several are in play).

Between reruns change one consequential dimension. After two structurally similar
misses, return to the creative skill instead of accumulating adjectives.

## References

- [references/api-quick-reference.md](references/api-quick-reference.md)

© Black Forest Labs · MIT · black-forest-labs/skills

Similar skills

All skills for AI and ML →

From the same set8

flux-3-prompt-doctor flux-3-prompt-doctor evaluates video briefs for FLUX 3 before generation. It resolves conflicting constraints and routes requests to the correct pipeline mode. CX Codex +1 203installs flux-3-video Routes general FLUX 3 video requests to the exact sub-skills handling prompt repair, visual staging, audio, and generation. CX Codex +1 185installs flux-3-cinematic-inserts flux-3-cinematic-inserts structures FLUX 3 text-to-video prompts for standalone insert shots and B-roll. It locks camera movement, physics, and scene continuity around a single visual event. CX Codex +1 185installs flux-3-keyframes-continuation flux-3-keyframes-continuation is a Claude Code skill for building FLUX 3 videos from images or clips with keyframes, storyboards, and continuation planning. CX Codex +1 183installs flux-3-audio-dialogue flux-3-audio-dialogue helps you direct FLUX 3 sound layer by layer: lines, voiceover, ambience, effects and music. Useful when a clip's audio comes out muddled. CX Codex +1 181installs flux-3-product-ads flux-3-product-ads — a Claude Code skill for turning FLUX 3 video and voice into a finished product ad. It guides shot planning, audio-based timing, assembly, and quality checks. CX Codex +1 136installs flux-3-archival-formats flux-3-archival-formats — a skill for Claude Code that configures FLUX 3 video generation for period and archival looks. It defines physical recording formats and specific degradation artifacts instead of moody adjectives. CX Codex +1 136installs

Project sponsors

Companies that keep this analytics open