noodle-skills

Ready-to-install agent skills, each wrapping a nanoodle visual AI workflow. Drop a folder into your agent's skills directory and Claude Code (or any agent that reads markdown playbooks and runs shell commands) can generate posters, restyle photos, sing jingles, narrate poems, and render video teasers — plus a coding-agent toolbelt: write alt text from real pixels, judge screenshots against a spec, mock up UI screens, draft self-audited PR descriptions, and mint OG cards — headless, on your own NanoGPT key.

Each skill is just two files: a SKILL.md playbook and a workflows/*.noodle-graph.json saved from the nanoodle editor. The nanoodle CLI (npx nanoodle) executes the graph; the agent supplies the inputs and hands back the outputs.

Which repo do I want? These are prebuilt skills — each one runs a single ready-made workflow. Want one skill that teaches your agent to design and run arbitrary graphs instead? That's nanoodle-skill. Agent speaks MCP rather than skills? Use nanoodle-mcp.

These skills spend real money. Every run bills your NanoGPT balance — cents per run, but real cents. Each SKILL.md carries an honest per-run ballpark, and agents are told to mention cost before running repeatedly.

The skills

Skill	What it does	Inputs	Output	Cost / run (approx.)
poster-generator	One-line idea → LLM prompt-writer → poster image	Idea	Poster.<ext>	~$0.04
photo-restyler	Photo + plain-English instruction → restyled image	Photo (file), Style	Restyled.<ext>	~$0.04
jingle-maker	Brief → lyrics LLM + style LLM → sung ~30s track	Brief	Jingle.<ext>	~$0.05–0.10
narrated-poem	Picture → vision → four-line poem → spoken audio	Picture (file)	Narration.<ext>	~$0.01
video-teaser	One-liner → shot-description LLM → 5s video clip	Tagline	Teaser.<ext>	~$0.06

For coding agents

Skill	What it does	Inputs	Output	Cost / run (approx.)
alt-text-writer	Vision model reads the actual pixels → alt text + caption JSON	Image (file)	Alt text (text)	<$0.01
visual-judge	Screenshot vs written spec → neutral describe → pass/fail JSON verdict	Screenshot (file), Spec	Verdict (text)	~$0.01
ui-mockup	Feature sentence + style guide → screenshot-style mockup render	Feature	Mockup.<ext>	~$0.04
pr-describer	Diff → draft → in-graph audit → final PR description	Diff (file)	Final PR body (text)	~$0.01
og-card-maker	Post/release line → house-style 1200×630 card + honest alt text	Post	Hero.<ext>, AltText (text)	~$0.05

Costs are ballparks from the NanoGPT catalog at the time of writing; the CLI prints the actual metered cost (and remaining balance) after every run.

Install

A skill is a plain directory — installing is copying (or symlinking) it:

# Claude Code, user-level
cp -r skills/poster-generator ~/.claude/skills/poster-generator

# Claude Code, project-level
cp -r skills/poster-generator <your-project>/.claude/skills/poster-generator

Other agents only need the folder somewhere they can read SKILL.md and run commands (~/.agents/skills/, ~/.grok/skills/, …).

Requirements at run time: Node.js >= 20 and a NanoGPT API key in NANOGPT_API_KEY (or a .env file passed via --env-file). npx nanoodle fetches the executor on first use — nothing else to install. The skills are tested against the nanoodle 0.4.x CLI (the version npx nanoodle currently fetches); this repo's devDependency pins the same range.

Try one from this repo

export NANOGPT_API_KEY=...        # spends from this balance
cd skills/poster-generator
npx nanoodle run workflows/poster.noodle-graph.json \
  --input "Idea=a synthwave night market on the moon" \
  --out ./poster-out

Inspect any workflow's interface offline (no key, no network):

npx nanoodle inspect skills/jingle-maker/workflows/jingle.noodle-graph.json

Design your own

These ten are templates, not a ceiling. The recipe is five steps:

1. Build and test a workflow visually at nanoodle.com (runs in your browser on your own key), then press 💾 Save to download the noodle-graph.json.
2. Name your nodes in the editor — custom node names become the --input keys and output filenames.
3. Drop the file into your-skill/workflows/.
4. npx nanoodle inspect it and check the derived keys.
5. Write SKILL.md documenting the exact run command.

The full guide lives in the nanoodle-js repo: docs/agent-skills.md.

Tests

npm install
npm test

Tests are fully offline and spend nothing: they verify each graph's derived input/output keys against its SKILL.md, that every model-bearing node pins a model, and they execute every playbook's documented run command verbatim against a local stub API. The suite runs against the pinned devDependency, nanoodle 0.4.x — the same version npx nanoodle fetches.

Limitations

• Runs bill your NanoGPT balance — there is no dry-run mode beyond inspect.
• Model ids are pinned to catalog entries that existed when this repo was last updated; NanoGPT occasionally renames or retires models. If a run fails with a model error, pick a replacement with --set "<nodeId>.model=<id>" or re-save the graph in the editor.
• Output quality is the model's, not the workflow's — rerun or tweak the prompts/system inputs for better takes.
• Media generation is not instant: music and video runs can take minutes. Don't kill the process early.

License

MIT.