THE AGENT SKILLS DIRECTORY
Reusable skills for AI agents. Install any of them into a Cohub world's agent with one command.
npx -y @neta-art/neta-search-cli@latest skills find "<what you need>"
Wrote a skill? Publish it here →
Basic Cohub abilities. Most agents already have these mounted; listed here for reference.
Paste this into your agent's chat — it fetches the skill and follows it.
Use the skill: red-blue-pen-world. Run `npx -y @neta-art/neta-search-cli@latest skills use @sumino/red-blue-pen-world --via site` and follow what it prints. Read its complete output, redirecting it to a temporary file first if necessary, and resolve relative paths from the skill directory it prints.npx -y @neta-art/neta-search-cli@latest skills use @sumino/red-blue-pen-worldTurn one character into a tiny figure inside a huge, hand-drawn world, in just red and blue pen on white paper.
# Red & Blue Pen World
One plate, one world, one very small person. The whole scene is drawn in exactly two inks — positive red and dark blue — on pure white paper. There is no fill, no grey, no gradient: volume comes from how tightly the lines are packed, from grid rows, hatching and dot clusters. Red carries plants and the principal contours; blue carries buildings, shadows and secondary structure. A single full-body figure stands somewhere in the middle ground, small enough to be a detail of the landscape rather than its subject.
**Demo:** `assets/demo-plate.webp` — a c. 1536×1024 plate in the finished style.
**Language:** reply in the user's language. Text the user sees in the deliverable (titles, captions, board group names) follows the user's language; the prompt itself stays English.
## What the user gets
- One panoramic plate (landscape, default 3:2), the world filling the frame top-down at an angle.
- Dense layered detail: plant-heavy foreground, built middle ground, receding distance, simple cloud lines.
- One character as a tiny whole-body figure (~1/12 of the frame height) doing one still, readable action.
- Color audit passed: only the two inks plus paper.
## When to use
- "red and blue pen drawing", "two-ink panorama", "pen-and-ink world map illustration", "duotone line art";
- 红蓝线稿、红蓝双色笔绘、全景线稿、红蓝版画、小人世界观插画.
- a character who should feel small inside a large world.
Do not use for:
- a single character portrait — `generate` with a character sheet;
- a coloured illustration of any kind — this skill forbids a third colour;
- a web page or poster — this is a plate.
Out of scope: do not change the world's settings (config, style, slug) for this plate, and do not delete or overwrite existing world files — write a new Work each time.
## Hard rules
0. **Sandbox CLI path.** Every `neta` command is `node /mods/neta/neta …`.
1. **Two inks, nothing else.** Red and dark blue on pure white. No grey, no halftone, no gradient, no colour mixing, no "just a little" accent.
2. **Density, not tone.** Say hatching, row lines, grid lines, dot clusters — and say that shading is line density, never fill.
3. **Red plants, blue architecture.** Keep that division in the brief, or the plate reads as one flat colour.
4. **The figure is tiny and complete.** Whole body head to feet, roughly one twelfth of the frame height, one still action, no close-up, no second figure.
5. **Layered depth.** Foreground plants / middle buildings and paths / distant skyline and clouds — all three named in the brief.
6. **No text.** No title, signature, watermark or halftone dots.
7. **Audit before delivering.** Run `scripts/audit-colours.mjs`; a third hue above 3% or a missing ink is a rework.
8. **Place it when it is a deliverable.** Board card via board-ops `place.mjs`.
## Inputs and when to ask
| Input | Counts as given when |
|---|---|
| who the figure is | the user named a character, or a world character atom is attached |
| the world / place | the user named a place, or the world's own setting answers "where does this character live" |
| ratio | the user named one; otherwise deliver 1536×1024 |
All three given → build straight through to delivery, no questions. Missing the character *and* the place → ask once, in one message, then stop. Missing only the ratio → pick 1536×1024 and say so.
## Pipeline
### 0. Read the character and the place
```bash
node /mods/neta/neta atom get "world/characters/<Name>.md" # identity, outfit, pose habits
node /mods/neta/neta find images "<Name>" # the reference to pass to the model
```
Collect: 4+ visible features for the figure, one action that is true to the character, and the place's signature structures (what makes this world's architecture recognizable).
### 1. Write the brief
Copy `templates/brief.example.json` and replace every placeholder: `[theme]`, `[place]`, the three layers, and `character`. Nothing else is needed.
```bash
node scripts/build-prompt.mjs /tmp/plate.json > /tmp/plate-prompt.txt
```
The script refuses thin layers (3 foreground / 4 middle / 2 background), a figure with no action, and a look shorter than 40 characters. It also warns when the character or world is a well-known licensed franchise — those are rejected by providers (see below).
### 2. Generate
```bash
NETA_DEFAULT_IMAGE_MODEL=gpt-image-2 node /mods/neta/neta generate image \
--param size=1536x1024 --param quality=high \
--image "world/media/studio/characters/<Name>.png" \
--work-title "<Name> · Red & Blue Pen World" \
--prompt "$(cat /tmp/plate-prompt.txt)"
```
Model and size facts that cost time if you guess:
- `gpt-image-2` handles this style best and accepts reference images. Declared sizes that actually work: `1024x1024`, `1536x1024`, `1024x1536`.
- **Ultra-wide or unusual sizes fail** — `2048x1152`, `1440x2560`, `864x1536` come back as HTTP 404 `Invalid URL` even though the declaration lists a 512–3840 range. Use 1536×1024 for a panorama; crop or extend later if a different ratio is demanded.
- Models exposing `aspect_ratio` (e.g. Gemini-style) take `16:9` / `3:2` instead of a pixel size — check with `neta generate image params` before sending flags.
### 3. Look, then audit
Open the generated file and check the list in `references/qa.md`. Then:
```bash
node scripts/audit-colours.mjs "<plate>.png"
```
Expect roughly 25–50% red, 5–20% blue, the rest paper. Failures mean a rework with the brief tightened, not a colour fix.
### 4. Failures and their fallbacks
| Symptom | Cause | Do this |
|---|---|---|
| HTTP 500 `unexpected EOF` | upstream hiccup | retry once into the same card: `--target-work <workId>` |
| HTTP 404 `Invalid URL` | unsupported size | drop to 1536×1024 or 1024×1536, retry |
| HTTP 400 `prompt_blocked` / "copyright restrictions" | a licensed character or franchise | **do not rephrase or retry** — provider policy. Offer the composite route below |
| Third hue in the plate | model added colour | rework: repeat "two inks only, no colour mixing, no grey" at the top and bottom of the prompt |
| Figure too big / cropped | scale drift | add "the figure is about one twelfth of the picture height, whole body visible, small in the frame" |
| Grey shading | "realistic lighting" leaked in | forbid halftone and grey explicitly; ask for "line density, dot clusters, grid rows" |
### 5. Licensed characters (the honest route)
Providers refuse famous characters and franchises at generation time. Redirecting the prompt to dodge that filter is not acceptable — instead composite the user's own image:
```bash
# 1. size the supplied art to the figure's height on the plate
magick supplied.png -resize x90 /tmp/figure.png
# 2. turn it into red/blue duotone line work: greyscale → threshold → colour map
magick /tmp/figure.png -colorspace Gray -contrast-stretch 2%x2% -threshold 55% /tmp/figure-bw.png
magick /tmp/figure-bw.png -fill "#b3201e" -opaque black /tmp/figure-red.png
# 3. composite onto the plate at the chosen spot, then run the colour audit again
magick plate.png /tmp/figure-red.png -geometry +740+520 -compose over -composite final.png
```
State plainly that the character came from the user's own file, not from generation.
### 6. Place on the Board and hand it over
```bash
node /mods/neta/.agents/skills/board-ops/scripts/place.mjs \
--title "<Name> · Red & Blue Pen World" --work <work.path>
```
## Deliver
The deliverable is one image Work plus its Board card. To hand it over:
1. Place the card (above), so the plate is findable on the canvas.
2. Give the Work a handover line of its own, using an interaction-components **link** card whose target is `{ "type": "work", "workId": <the id from the generate output> }` — never a bare path, never only an inline image.
3. One line on what the plate holds (which inks, which layers, where the figure stands), then the next-steps options.
## Files
```
scripts/build-prompt.mjs # brief JSON → prompt, with layer checks and a licensed-name warning
scripts/audit-colours.mjs # two-ink audit: red / blue / paper / foreign hues
templates/brief.example.json # generic lighthouse-town example, all placeholders present
references/qa.md # visual checklist, audit interpretation, composite recipe
assets/demo-plate.webp # finished plate, for style reference
```
## Pitfalls
- **"White background" is not enough** — say "pure white paper, no grey, no halftone", or shadows arrive as soft grey.
- **One ink dominating** — a blue building on blue hatching turns the plate into a monochrome; keep red on plants and contours so the eye has two temperatures.
- **The figure is the hard part** — models enlarge it and add detail; repeat the one-twelfth rule and "whole body, no close-up".
- **Detail without order** — "crowded but orderly" plus named layers keeps the composition legible; without it you get texture soup.
- **Post-generation crops break the ratio story** — if the user needs 9:16 or 1:1, extend the canvas by stretching the edge pixel rows, or crop from a 1536×1024 plate with generous margins.