# Agent Skills \[Accurate rxfy context for AI coding assistants] rxfy ships **two agent skills** — structured reference libraries that AI coding assistants (Claude Code, GitHub Copilot Agent, Codex, and others) load on demand to get accurate, rxfy-specific guidance without hallucinating APIs. Without a skill an agent has only its training data: it may invent hook signatures, confuse normalized ids with entity data, or silently omit the `name`/`key` fields that SSR requires. With a skill it has the real API surface and the common pitfalls already in context. ## Two skills — one for setup, one for usage | Skill | Use it when | | ------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `rxfy-setup` | **Getting rxfy into a project** — starting a new app from a template, or adding rxfy to an existing app at a chosen depth. Routes the two paths, gives the install commands, records the setup variant, then hands off. | | `rxfy` | **Working in a project that already has rxfy** — declaring models/states, reading data, SSR, sync, debugging. A task-indexed reference library that orients to your integration level. | rxfy is one framework you adopt at progressive levels — **Store → +SSR → +Sync**. `rxfy-setup` gets you to a level; `rxfy` is how you work once you're there. During setup, `rxfy-setup` records which variant you chose (in `CLAUDE.md` or memory) so `rxfy` never has to re-detect the project type. ## Install ```bash # setup / installation front-door npx skills add vanya2h/rxfy --skill rxfy-setup # working in an rxfy project npx skills add vanya2h/rxfy --skill rxfy ``` Installing both is fine — they cover different phases and don't overlap. ## How the `rxfy` skill works The skill's `SKILL.md` is a small router, not a wall of text. It carries the two invariant rules (`data$` emits ids, not entities; `patch` edits in place while `stale` triggers a refetch) and a **library table** the agent reads from by task. On first contact it detects — from files on disk — whether it's in a scaffolded template or an existing app, and which level the app is at, then opens only the matching mini-guides. A client-only project never pulls sync or grants context. ## What the skill covers **Store + React (always):** * Models (`createModel`) and states (`defineState`), including [plain value fields](/rxfy/define-state#plain-value-fields) * The normalized data flow: `useStateData` returns ids, `useModelStore` gives entities * `useAtom`, `usePending`, `` — async rendering patterns * Mutations, `set` vs `setRaw`, pagination with `useStatePagedData` * Lens composition with `createLens` / `keyLens` * Common mistakes: inline observable creation, `store.get()` on unloaded ids, atom stability, missing-data-after-SSR **SSR:** * `name`/`key` requirements, buffered (`onAllReady`), streaming (``), and two-pass (`collectStateData`) modes **Sync (real-time layer):** * Binding Drizzle tables with `defineResource` and writing through `sync.create` / `sync.update` / `sync.delete` * `patch` vs `stale`: which writes publish what, and why lists refresh instead of mutating in place * The wire protocol and the `rxfy-ws` transports (`createWsServer`, `createWsClient`) * `createSyncClient`, the `StoreProvider` `syncClient` prop, and `updatesAvailable$` / `applyUpdates` * Grants and sync hydration: `sync.serve` (signs a channel grant as `$grant`), `sync.hydration`, `createSyncClient` with grant renewal, `readSsrGrants` **Templates:** * What each `create-rxfy-app` template (`vite-spa`, `next`, `vite`) already wires, where the model/state lives, and where to add the next entity — so the agent extends the scaffold instead of rebuilding it