generate: a pay-per-use Claude skill that replaces a Higgsfield subscription A Claude Code skill that makes images and short vertical videos through kie.ai, the pay-as-you-go API that hosts the same models Higgsfield wraps (Nano Banana 2, GPT Image 2, Seedance).
generate: a pay-per-use Claude skill that replaces a Higgsfield subscription
以下摘录自该仓库的 README(片段) · 在 GitHub 读完整 README →
A Claude Code skill that makes images and short
vertical videos through kie.ai, the pay-as-you-go API that
hosts the same models Higgsfield wraps (Nano Banana 2, GPT Image 2, Seedance).
You pay per generation instead of per month: a still is about 4 cents and a
10-second 1080p clip is about 70 cents.
Inspired by the RoboNuggets video
"This 1 Claude Skill fully replaces your Higgsfield Subscription".
What it does
1. You type /generate a POV of waking up as a Roman senator, vertical.
2. Claude drafts a still on a cheap image model.
3. Claude looks at the still and runs a QA gate: size, orientation, no stray text, matches the brief.
4. Claude quotes the clip cost and your remaining budget, then waits for "go".
5. Claude animates the still, downloads the mp4, checks its real dimensions and duration, and logs the spend.
Spend caps (default $10 per day, $50 per month) are enforced by the script, and
every run is recorded in generations/ledger.json.
Install
Requirements: Python 3.9+ (standard library only), a kie.ai API key with a few
dollars of credit. ffmpeg is optional (frame extraction for QA).
On Windows PowerShell:
Then:
1. Copy .env.example to .env inside the skill folder.
2. Paste your key from https://kie.ai/api-key into KIE_API_KEY=.
3. In Claude Code, run /generate and describe what you want.
The same folder works in ~/.codex/skills/, ~/.openclaw/skills/ and
~/.hermes/skills/.
Layout
Dashboard
A single page shows your kie.ai credit balance, today's and this month's spend
against the caps, a per-day spend chart, and a review board of every still and
clip. The create panel is built for iteration, not single shots:
- **Prompt builder.** Subject, setting, camera, lighting, mood, details (or
camera move, subject motion, atmosphere, keep-still for video) compose the
final prompt, with the chosen preset's locked fragment and a negative prompt.
Presets live in public/presets.json.
- **Variations.** Generate 1 to 4 stills from one prompt at once. They show up
as a set; "Pick" stars one and rejects the rest.
- **Reference images.** Upload faces or products once, select up to three per
still, or use any board image as a reference for character consistency.
- **Review board.** Star, reject, note, search, filter to starred, hide or show
rejected, select two and compare side by side, click to view large.
- **One-click follow-ups.** Animate, Tweak (reload into the builder), Use as
ref, 16:9 and 1:1 reframes, a 2K or 4K version, and for clips Redo and Extend
(captures the last frame in the browser and uses it as the next source still).
- **Draft then finalize for Seedance 2.5.** Clips are made as 480p drafts
(28 credits per second, so about $1.40 per 10 s), one to four at a time.
Press Finalize on the keeper and kie.ai re-renders that exact draft at 1080p
(158 credits per second, about $7.90 per 10 s) with the same motion. Draft
ids stay valid for 7 days.
It runs two ways from the same public/index.html:
**Locally** (reads your generations/ folder, runs kie.py for you):