Skills catalog / Stillwake
Stillwake v1.0.0
Turn the still artwork you already have into illustrations that breathe.
Nothing gets redrawn. Stillwake takes the pictures your site already serves and gives them small, restrained movement: a lamp that breathes, paper that settles, a wheel that turns. The page keeps its layout and the art keeps its character.
It lists the picture files actually on your pages, animates one as a test before you spend on the rest, then works through the batch. An automatic check rejects any clip that drifts from the original shape or jumps where the loop restarts, so a picture cannot quietly turn into a different picture.
Visitors on a phone, and anyone whose system is set to reduce motion, keep the original still image.
Validated: complete and repeatable
Distribution details are being finalized. Read how we admit skills →
Who it’s for
Who it’s for and what your agent needs
- Developers who look after their own site's illustrations and icons
- Designers adding movement to an interface that is already live
- Teams with a settled illustration style they do not want changed
- Builders who use an AI assistant to make changes to their own site
Minimum (stills-only)
- An AI coding agent with file and shell access
- An image-to-video tool with start-image conditioning
- ffmpeg and ffprobe
- Python 3.10+ for background removal
- A media backend route per the bundled routing guide (reference: Codex CLI stills, Grok CLI image-to-video; backups: Higgsfield MCP, Gemini or Antigravity hosts)
Full motion
- A background-removal model runtime such as rembg with BiRefNet
- img2webp or an equivalent animated-WebP encoder
What’s inside
What’s inside the bundle
- Stillwake Agent Skill
- A guide to which image and video services to use, with one recommended route and backups
- A method for listing the pictures your site actually serves, with duplicates removed
- One tracking file that follows every asset through generation, processing, and integration
- A strict prompt rule that keeps the movement small and stops the model inventing new objects
- A list of suitable movements per subject, so a lamp glows and a page lifts instead of everything wobbling
- A calibration pass that proves one picture end to end before any paid batch runs
- Batch generation that survives interruptions and quota stops, and picks up where it left off
- Recipes for every step from raw clip to a finished, web-ready animated file
- Automatic pass or fail checks, with thresholds set against real production artwork
- Review sheets that show every result side by side on light and dark backgrounds
- Hover behavior for mouse users, and the fallback for reduced-motion settings
- A restart plan, a list of common failures with what to do about each, and a finishing checklist
- Installation and troubleshooting guides
Proof
What the evidence shows
Stillwake and a capable assistant working without it were given the same three requests and the same three illustrations. Stillwake shipped two of the three and refused the third, because no attempt at it held the original shape. The assistant working without Stillwake shipped all three, and two of those quietly redrew the picture: their outlines drift 18 and 43 pixels from the original, where Stillwake rejects anything past 8. This shows the checks do their job. It is not evidence that Stillwake is faster, and we will not present it that way.
- Eight attempts were made in total. Six were rejected: three by the automated checks for changing the outline, and three by the review step that asks a person to look at the result before it ships.
- The two that shipped stay within 8 and 3 pixels of the original outline, loop without a visible jump, keep their transparency, and come in at 502 KB and 437 KB against a 900 KB ceiling.
- The run also found three faults in Stillwake itself, all now fixed in version 1.0.0: a prompt that let the generator turn flat paper into a photograph, a resolution default that quietly wrecked composition, and a missing check for a subject slowly redrawing itself while every other measurement looked healthy.
- On a page, the original picture is what loads and what stays for anyone who has reduced motion turned on. This was checked in a browser: no motion file is requested at all in that case, and an illustration with no animation simply never swaps.
The third illustration was refused, not delivered late. Every attempt at it was caught by the automated silhouette check, which compares each frame's outline against the original and rejects anything past 8 pixels; the two attempts measured 19 and 27. That check exists because of this run: the earlier version had sixteen checks, all of which those attempts passed.
Not included
What it doesn’t include
- Subscriptions to the image or video services you choose
- Paid API credit for those services
- Generated media: the clips are produced during the run, through your own accounts
- The background-removal model itself, or a license to run it
The bundle contains standard Markdown and instructions that do not depend on one particular AI provider. It avoids tool names and file paths that only work inside one particular assistant. It is built for assistants that support the Agent Skills format, or that can be pointed at SKILL.md and told to follow it, with the capabilities listed above; exact skill installation steps vary by client.
Install, updates, support
Install, updates, and support
- Download the zip and extract it into a directory named stillwake, with SKILL.md at the directory root.
- Copy the complete stillwake directory into the skills directory your AI coding assistant recognizes, at the user level or the project level.
- Restart or refresh your assistant, confirm stillwake appears in the available skills, then ask the agent to use it to bring your existing illustrations to life.
Updates. Shipped version: 1.0.0.
Support covers
Installation of the skill bundle, help understanding what a step means, and reproducible defects in the bundled documentation and defaults.
Outside the boundary
Provider outages, provider or model billing, custom animation or creative production services, background-removal model licensing, hosting, legal clearance, and guarantees that a third-party model will accept a prompt.