Home Extensions · 01 — Skill

Skills

A skill is a procedure the model reads before it acts: when to use it, how to shape the request, and where the API will refuse.

Two skills SKILL.md + CLI Python standard library Agnes 2.5 Flash v1.0.0

What a skill is

A folder it can read, a script it can run.

An MCP server hands the model a tool. A skill hands it a procedure: the trigger words, the order of the request, the two API rules that will bite, and the command that does the work. One is a socket; the other is a recipe with the failures written in the margin.

Both skills here share one shape — frontmatter that decides when it loads, a script that needs nothing but the Python standard library, and a reference file for the sharp edges. Nothing runs in the background, and nothing holds a key except a .env file you own.

Loads by description

The frontmatter says which requests should wake it — including the words a Chinese user actually types. No menu, no command list.

The script does the talking

The curl-shaped mistakes are encoded once, in Python, instead of being re-derived every turn by a model that is guessing.

The key stays local

A .env beside the skill, resolved flag → environment → file. AGNES_API_KEY never enters the prompt, the log or the answer.

Skill A written procedure plus a script. Loaded when its description matches the request. SKILL.md
Plugin Code inside the agent runtime. Changes how a run behaves rather than what it can call. dsh ACP
Connector A door into the agent from somewhere else — a chat app, a microphone. bridge
MCP Server A separate process exposing tools over stdio JSON-RPC. stdio
The two

One draws a still. One shoots five seconds.

Both hand the job to an outside model, and both come back with a file on disk.

SKILL-01agnes-image· 2.5 flash

Draw it, then edit what you drew.

Text-to-image, image-to-image and multi-image composition. The skill's real work is the prompt shape — subject, scene, style, light, composition, quality — and, for an edit, naming what must survive it.

  • Three modes, one script. text · image · multi-image; local files become data URIs on the way in.
  • Tiers, not pixels. Ask for 1K–4K plus a ratio — exact sizes like 1920×1080 get normalised anyway.
  • Saved, then shown. The script prints Saved: <path>, and the image itself is presented — not described.
Download · zip agnes-image-skill.zip
agnes-image · text-to-image simulated
prompt
A classical bust in profile, torn paper collage, warm ivory light, editorial composition, high detail.
ratio
ready
Click a prompt or ratio, then Generate. Four samples come back; click one to keep it. The demo is a front-end simulation — no request leaves the page.
SKILL-02agnes-video· 2.5 flash

Five seconds, first frame to last.

Text-to-video, keyframe control and reference mode. The hard part is the waiting: the free tier allows one request a minute, so the skill creates a task, polls every 61 seconds, then downloads the mp4 — and says so, so the agent does not panic and relaunch it.

  • Create, poll, download. One command runs the whole task; status lines are the only progress you get.
  • Keyframes. First and/or last frame, as a public URL or a local path.
  • References. Up to 5 images and 3 audios, cited in the prompt as <Picture 1> and <Audio 1>.
  • Flash limits, checked first. 720P only, 4–12 seconds, no video input. The script refuses before the API has to.
Download · zip agnes-video-skill.zip
agnes-video · keyframe · 5s · 720P simulated
mode
mode=keyframe
0.0s
1.4s
3.1s
5.0s
ready · task not created
stdout stays JSON-only, one line per event
Press Generate to watch the real shape of the task: create, poll, download. On the free tier those two polls are 61 seconds apart; here they are compressed into a second.
Install

Two folders and a key.

No build step and no dependencies — the scripts import the standard library and nothing else.

  1. 01Unzip into the skills folder. Each zip becomes one folder: agnes-image and agnes-video, side by side in Quanta's skill directory.
  2. 02Copy .env.example to .env. Paste AGNES_API_KEY. No key yet? The skill asks instead of failing quietly.
  3. 03Restart Quanta. The page's skill list picks both up from their frontmatter — no registration step.
$ tree skills -L 2
skills/
├── agnes-image/
│   ├── SKILL.md
│   ├── .env.example
│   ├── scripts/generate_image.py
│   └── references/api_docs.md
└── agnes-video/
    ├── SKILL.md
    ├── .env.example
    ├── scripts/generate_video.py
    └── references/api_docs.md

$ python3 skills/agnes-image/scripts/generate_image.py \
    --prompt "bust, torn paper, ivory light" --size 2K --ratio 3:4
Saved: ./output.png

The reference file is the part people skip. It holds the parameter table, the size matrix and the errors — which is why the agent does not have to rediscover them on every run.

Next

The parts that watch the run.