Thanks for thinking about contributing. OD is small on purpose — most of the value lives in **files** (skills, design systems, prompt fragments) rather than framework code. That means the highest-leverage contributions are usually one folder, one Markdown file, or one PR-sized adapter.
This guide tells you exactly where to look for each type of contribution and what bar a PR has to clear before we merge it.
| If you want to… | You're really adding | Where it lives | Ship size |
|---|---|---|---|
| Make OD render a new kind of artifact (an invoice, an iOS Settings screen, a one-pager…) | a **Skill** | [`skills/<your-skill>/`](skills/) | one folder, ~2 files |
| Make OD speak a new brand's visual language | a **Design System** | [`design-systems/<brand>/DESIGN.md`](design-systems/) | one Markdown file |
If you're not sure which bucket your idea is in, [open a discussion / issue first](https://github.com/nexu-io/open-design/issues/new) and we'll point you at the right surface.
---
## Local setup
The full one-page setup lives in [`QUICKSTART.md`](QUICKSTART.md). The TL;DR for contributors:
Node `~24` and pnpm `10.33.x` are required. `nvm` / `fnm` are optional; use `nvm install 24 && nvm use 24` or `fnm install 24 && fnm use 24` if you prefer managing Node that way. macOS, Linux, and WSL2 are the primary paths. Windows native should work but isn't a primary target — file an issue if it doesn't.
You don't need any agent CLI on your `PATH` to develop OD itself — the daemon will tell you "no agents found" and fall back to the **Anthropic API · BYOK** path, which is the fastest dev loop anyway.
---
## Adding a new Skill
A skill is a folder under [`skills/`](skills/) with a `SKILL.md` at the root, following Claude Code's [`SKILL.md` convention][skill] plus our optional `od:` extension. **No registration step.** Drop the folder in, restart the daemon, the picker shows it.
### Skill folder layout
```text
skills/your-skill/
├── SKILL.md # required
├── assets/template.html # optional but recommended — the seed file
├── references/ # optional — knowledge files the agent reads
│ ├── layouts.md
│ ├── components.md
│ └── checklist.md
└── example.html # strongly recommended — a real, hand-built sample
```
### `SKILL.md` frontmatter
The first three keys are the Claude Code base spec — `name`, `description`, `triggers`. Everything under `od:` is OD-specific and optional, but **`od.mode`** decides which group the skill shows up in (Prototype / Deck / Template / Design system).
```yaml
---
name:your-skill
description:|
One-paragraph elevator pitch. The agent reads this verbatim to decide
if the user's brief matches. Be concrete: surface, audience, what's in
featured:1# any positive integer surfaces it under "Showcase examples"
preview:
type:html # html | jsx | pptx | markdown
entry:index.html
design_system:
requires:true# does the skill read the active DESIGN.md?
sections:[color, typography, layout, components]
example_prompt:"A copy-pastable prompt that nicely shows what this skill does."
---
# Your Skill
Body is free-form Markdown describing the workflow the agent should follow…
```
The full grammar — typed inputs, slider parameters, capability gating — lives in [`docs/skills-protocol.md`](docs/skills-protocol.md).
### Bar for merging a new skill
We're picky about skills because they're the user-facing surface. A new skill must:
1.**Ship a real `example.html`.** Hand-built, opens straight from disk, looks like something a designer would actually deliver. No lorem ipsum, no `<svg><rect/></svg>` placeholder hero. If you can't build the example yourself, the skill probably isn't ready.
2.**Pass the anti-AI-slop checklist** in the body. No purple gradients, no generic emoji icons, no rounded card with a left-border accent, no Inter as a *display* face, no invented stats. Read the **Anti-AI-slop machinery** section of the README for the full list.
3.**Honest placeholders.** When the agent doesn't have a real number, write `—` or a labelled grey block, not "10× faster".
4.**Have a `references/checklist.md`** with at least P0 gates (the stuff the agent has to pass before emitting `<artifact>`). Lift the format from [`skills/guizang-ppt/references/checklist.md`](skills/guizang-ppt/) or [`skills/dating-web/references/checklist.md`](skills/dating-web/).
5.**Add a screenshot** at `docs/screenshots/skills/<skill>.png` if the skill is featured. PNG, ~1024×640 retina, captured from the real `example.html` at zoomed-out browser scale.
6.**Be a single self-contained folder.** No CDN imports beyond what other skills already use; no fonts you didn't license; no images larger than ~250 KB.
If you fork an existing skill (e.g. start from `dating-web` and remix into a `recruiting-web`), keep the original LICENSE and authorship in `references/` and call it out in your PR description.
### Skills that already ship — pick one to imitate
- Multi-frame mobile flow: [`skills/mobile-onboarding/`](skills/mobile-onboarding/), [`skills/gamified-app/`](skills/gamified-app/)
- Document / template (no design system required): [`skills/pm-spec/`](skills/pm-spec/), [`skills/weekly-update/`](skills/weekly-update/)
- Deck mode: [`skills/guizang-ppt/`](skills/guizang-ppt/) (bundled verbatim from [op7418/guizang-ppt-skill][guizang]) and [`skills/simple-deck/`](skills/simple-deck/)
---
## Adding a new Design System
A design system is a single [`DESIGN.md`](design-systems/README.md) file under `design-systems/<slug>/`. **One file, no code.** Drop it in, restart the daemon, the picker shows it grouped by category.
### Design system folder layout
```text
design-systems/your-brand/
└── DESIGN.md
```
### `DESIGN.md` shape
```markdown
# Design System Inspired by YourBrand
> Category: Developer Tools
> One-line summary that shows in the picker preview.
## 1. Visual Theme & Atmosphere
…
## 2. Color
- Primary: `#hex` / `oklch(...)`
- …
## 3. Typography
…
## 4. Spacing & Grid
## 5. Layout & Composition
## 6. Components
## 7. Motion & Interaction
## 8. Voice & Brand
## 9. Anti-patterns
```
The 9-section schema is fixed — that's what skill bodies grep for. The first H1 becomes the picker label (the `Design System Inspired by` prefix is stripped automatically), and the `> Category: …` line decides which group it lands in. Existing categories are listed in [`design-systems/README.md`](design-systems/README.md); if your brand truly doesn't fit, you can introduce a new one, but **try existing categories first**.
### Bar for merging a new design system
1.**All 9 sections present.** Empty section bodies are fine for hard-to-find data (e.g. motion tokens), but the headings have to be there or the prompt grep breaks.
2.**Hex codes are real.** Sample directly from the brand's site or product, not from memory or AI guesses. The README's "brand-spec extraction" 5-step protocol applies to maintainers too.
3.**OKLch values for accent colors** are nice-to-have. They make palettes lerp predictably across light/dark.
4.**No marketing fluff.** The brand's tagline is not a design token. Cut it.
5.**Slug uses ASCII** — `linear.app` becomes `linear-app`, `x.ai` becomes `x-ai`. The 69 imported systems already follow this convention; mirror it.
The 69 product systems we ship are imported from [`VoltAgent/awesome-design-md`][acd2] via [`scripts/sync-design-systems.ts`](scripts/sync-design-systems.ts). If your brand belongs upstream, **send the PR there first** — we'll pick it up automatically on the next sync. The `design-systems/` folder is for systems that don't fit upstream, plus our two hand-authored starters.
That's it — daemon will detect it on `PATH`, the picker shows it, the chat path works. If the CLI emits **typed events** (like Claude Code's `--output-format stream-json`), wire a parser in [`apps/daemon/src/claude-stream.ts`](apps/daemon/src/claude-stream.ts) and set `streamFormat: 'claude-stream-json'`.
1.**A real session works end-to-end** with the new agent — paste the daemon log into the PR description showing it streamed an artifact through.
2.**`docs/agent-adapters.md`** is updated with the CLI's quirks (does it require a key file? does it support image input? what's its non-interactive flag?).
3.**The README's "Supported coding agents" table** gets one row.
API-mode chat sends `max_tokens` to the upstream provider on every request. The web client picks that number from a three-tier lookup in [`apps/web/src/state/maxTokens.ts`](apps/web/src/state/maxTokens.ts):
1. The user's explicit override in Settings, if set.
2. Otherwise, the per-model default in [`apps/web/src/state/litellm-models.json`](apps/web/src/state/litellm-models.json) — a vendored slice of [BerriAI/litellm][litellm]'s `model_prices_and_context_window.json` (MIT). It covers ~2k chat models across Anthropic, OpenAI, DeepSeek, Groq, Together, Mistral, Gemini, Bedrock, Vertex, OpenRouter, and friends.
3. Otherwise, `FALLBACK_MAX_TOKENS = 8192`.
To pick up a newly-launched model, regenerate the vendored JSON:
The script fetches LiteLLM's catalog, filters to `mode: 'chat'` entries, projects each to its `max_output_tokens` (or `max_tokens` fallback), and writes a sorted snapshot. Commit the regenerated `litellm-models.json` alongside whatever PR triggered the refresh.
The OVERRIDES table in `maxTokens.ts` is for the rare case where LiteLLM is missing or wrong for a model id we actually use — for example, `mimo-v2.5-pro` (LiteLLM only ships MiMo via the `openrouter/xiaomi/...` and `novita/xiaomimimo/...` aliases, neither of which matches the canonical id Xiaomi's direct API uses). Keep it small; everything that LiteLLM gets right belongs upstream.
German uses formal `Sie` because OD speaks to a mixed audience of solo creators, agencies, and engineering teams; until project feedback shows that an informal `du` voice fits better, formal German is the least surprising default. Locale PRs should translate UI chrome, core docs, and display-only gallery metadata in `apps/web/src/i18n/content.ts`, but should not translate `skills/`, `design-systems/`, or prompt bodies that agents execute. Those source prompts are maintained as workflow inputs, and keeping one source language avoids multiplying prompt QA across locales. When adding or renaming a skill, design system, or prompt template, update the German display metadata and run `pnpm --filter @open-design/web test`; `content.test.ts` fails if German display coverage drifts. Daemon errors, export filenames, and agent-generated artifact text are known limitations unless a PR explicitly scopes them.
For step-by-step instructions on adding a new locale (UI dictionary, README, language switcher, regional terminology), see [`TRANSLATIONS.md`](TRANSLATIONS.md).
We're not pedantic about formatting (Prettier on save is fine), but two rules are non-negotiable because they show up in the prompt stack and the user-facing API:
1.**Single quotes in JS/TS.** Strings are single-quoted unless escaping makes them ugly. The codebase is already consistent — please match.
2.**Comments in English.** Even if the PR is translating something into Deutsch or 中文, code comments stay in English so we can keep one set of greppable references.
- **Don't narrate.** No `// import the module`, no `// loop through items`. If the code reads obviously, the comment is noise. Save comments for non-obvious intent or constraints the code can't express.
- **No new top-level dependencies** without a paragraph in the PR description on what we get vs. what bytes we ship. The dep list in [`package.json`](package.json) is small on purpose.
- **Run `pnpm typecheck`** before pushing. CI runs it; failing it earns a "please fix" comment.
---
## Commits & pull requests
- **One concern per PR.** Adding a skill + refactoring the parser + bumping a dep is three PRs.
- **Title is imperative + scope.** `add dating-web skill`, `fix daemon SSE backpressure when CLI hangs`, `docs: clarify .od layout`.
- **Body explains the why.** "What does this do" is usually obvious from the diff; "why does this need to exist" rarely is.
- **Reference an issue** if there is one. If there isn't and the PR is non-trivial, open one first so we can agree the change is wanted before you spend the time.
- **No squash-during-review.** Push fixups; we'll squash on merge.
- **No force-push to a shared branch** unless the reviewer asked.
We don't enforce a CLA. Apache-2.0 covers us; your contribution is licensed under the same.
- Which agent CLI was selected (or whether you were on the BYOK path).
- The skill + design system pair that triggered it.
- The relevant **daemon stderr tail** — most "the artifact never rendered" reports get diagnosed in 30 seconds when we can see `spawn ENOENT` or the CLI's actual error.
- A screenshot if it's UI.
For prompt-stack bugs ("the agent emitted a purple gradient hero, the slop blacklist was supposed to forbid that"), include the **full assistant message** so we can see whether the violation was the model or the prompt.
---
## Asking questions
- Architecture question, design question, "is this a bug or a misuse" → [GitHub Discussions](https://github.com/nexu-io/open-design/discussions) (preferred — searchable for the next person).
- "How do I write a skill that does X" → Open a discussion. We'll answer it and turn the answer into [`docs/skills-protocol.md`](docs/skills-protocol.md) if it's a missing pattern.
---
## What we don't accept
To keep the project focused, please don't open PRs that:
- **Vendor a model runtime.** OD's whole bet is "your existing CLI is enough". We don't ship `pi-ai`, OpenAI keys, or model loaders.
- **Rewrite the frontend away from the current stack without prior discussion.** Next.js 16 App Router + React 18 + TS is the line. No Astro, Solid, Svelte, or other framework rewrites unless maintainers explicitly want that migration.
- **Replace the daemon with a serverless function.** The daemon's whole point is owning a real `cwd` and spawning a real CLI. Vercel deployment of the SPA is fine; the daemon stays a daemon.
- **Add telemetry / analytics / phone-home.** OD is local-first. The only outbound calls are to providers the user explicitly configured.
- **Bundle a binary** without a license file and authorship attribution next to it.
If you're not sure whether your idea fits, open a discussion before writing the code.
---
## License
By contributing, you agree your contribution is licensed under the [Apache-2.0 License](LICENSE) of this repository, with the exception of files inside [`skills/guizang-ppt/`](skills/guizang-ppt/), which retain their original MIT license and authorship attribution to [op7418](https://github.com/op7418).
