Write a CSF3 story for ONE React component, covering only its materially-different states (no Cartesian), with a factory when 3+ stories share a shape. Use for 'write a story for X', 'document this component', 'add a Storybook story'.
Installs into .claude/skills of the current project.
Are you the author of Sb Stories?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/strongeron-sb-stories)
---
name: sb-stories
description: "Write a CSF3 story for ONE React component, covering only its materially-different states (no Cartesian), with a factory when 3+ stories share a shape. Use for 'write a story for X', 'document this component', 'add a Storybook story'."
compatibility: "Requires bash, python3, and Node.js (Storybook; the --strict gate runs tsc + eslint via npx); git optional (--diff mode), jq optional (design-system hint in check-story-ready)."
allowed-tools: Bash Read Glob Grep Write Edit
license: MIT
metadata:
author: strongeron
version: '2.4.0'
bundle: storybook-workbench
vendor:
# Skill-local files live in this skill's scripts/ + references/; shared ones (e.g. discover-runtime.py,
# validate-stories.sh, anti-patterns.md, composition-patterns.md) resolve from shared/ ($CORE) in dev
# and are copied into dist/ by build.sh on export.
scripts: [validate-stories.sh, check-story-ready.sh, scaffold-factory.sh, extract-states.sh, extract-prop-shapes.sh, page-patterns.py, scaffold-page-story.py, discover-runtime.py]
wrappers: false
references: [with-mcp.md, without-mcp.md, storybook-surface.md, anti-patterns.md, validate-workflow.md, factory-patterns.md, extraction-workflow.md, test-wiring.md, directory-structure.md, composition-patterns.md]
templates: [controlled-component-story.tsx]
---
# sb-stories — one component, its real states
The default Build mode. The component exists in `src/components/`; you write its visible-states story.
## Before authoring (ask yourself)
- **Is it actually used?** Check `components.real[]` in `.storybook/project-inventory.json`. If it's
in `dead[]`, remove it — don't write a story. If it's `vendor` (shadcn `ui/`), deprioritize.
- **Which states change behavior/appearance materially?** Refuse Cartesian combinations. Per-primitive
minimum tables in `references/anti-patterns.md` (Button 8, Input 8, Modal 5, Form 6).
Read `.storybook/component-states.json` instead of guessing — **if it's missing, generate it first:**
`scripts/extract-states.sh`. (Bigger extraction context: `references/extraction-workflow.md`.)
To render all those states on **one canvas**, use the `StateGrid` wrapper (variants × states →
`StateMatrix`) from `sb-wrappers` — don't hand-roll the grid. These are the component-time wrappers;
the data wrappers (ProjectInventory / DesignSystemHealth / AppFlowGraph …) come after their own steps.
- **Which states does THIS app actually ship?** Read `.storybook/component-usage.json` (`sb-inventory`'s
`extract-component-usage.sh` — generate it if missing). Use `props.<prop>.values` to **prioritize** the
variants real call sites pass; for anything in `declaredButUnused` (e.g. `variant=danger` never used),
still author the state for completeness but **tag it `['usage:unused']`** and note "not used in this app"
— don't pad the catalog with states prod never renders.
- **You don't author per-component docs by hand — they're composed once.** Every component's autodocs
page already gains a **"Real usage in this app"** band: the `UsageSection` block (wired once into
`preview.ts` `docs.page` by `sb-setup`) renders a **"Where it's used"** map per component — the pages it
lands on, what nests it, what it renders, the tokens it pulls — read from the usage graph
(`component-pages.json`). So when you add a story, **don't hand-add a usage block**; just run
`refresh-usage.sh` so the graph is current and the band populates. To explore the whole graph
interactively (any token / component / page → everywhere it's used, clickable), scaffold the
**`UsageExplorer`** wrapper (`sb-wrappers`). Details + the docs.page composition live in `sb-inventory`
(§ "Real usage in autodocs") and `sb-setup` (docs-page composition).
- **Factory?** YES if 3+ stories share a data shape. Read `.storybook/prop-shapes.json` (candidates
flagged with `liveUsages`) — **if missing, run `scripts/extract-prop-shapes.sh` first.** Then
`scripts/scaffold-factory.sh <Type> <import-path>` and fill the deterministic stub
(`references/factory-patterns.md`); otherwise inline `args`.
- **Title taxonomy?** Match `storySort.order`; if the project has none, pick one via
`references/directory-structure.md`.
- **Does a `play` actually earn its place?** Only write one for an interaction, async data,
a portal, a CSS-driven state, or accessibility — never a bare `toBeVisible()` (anti-pattern 34).
And the **project needs exactly one `CssCheck`** (anti-pattern 33): one story asserting a real
`getComputedStyle` token value, the only proof the shared preview loaded the app's CSS. Both come
from `npx storybook ai setup`'s prompt; `validate-stories.sh` check 13 + the project tally enforce them.
- **A "playground" / showcase story is Controls-driven, not click/hover.** If a reviewer asks "what is
this playground story — can I click or hover it?": a playground is the **default story with its Controls
panel** — you change `args` (variant, size, state) in the Controls tab and the canvas re-renders. It is
NOT canvas interaction; hover/focus/active live in a `StateGrid` interaction matrix (addon-pseudo-states),
and scripted click→assert flows live in a `play` function. So: Controls = try props · pseudo-states =
hover/focus columns · `play` = a real interaction. Say which one the story is.
- **Is the Controls panel wired?** A component story must expose a usable Controls panel — it's the
reviewer's prop sandbox and powers the autodocs ArgTypes table. The react-vite default
(`react-docgen`) does NOT infer TS unions into selects, so declare `argTypes` for every enum/union
prop (`control: 'select'|'inline-radio'` + `options`), group with `table.category`, hide
escape-hatch / non-serializable props (`className`, refs, icon / callback / data props), and disable
controls on render-only showcase stories (`parameters: { controls: { disable: true } }`). Full
pattern + the docgen gotcha: `references/without-mcp.md` §13. (The panel itself must be *visible* —
`sb-setup` writes a `manager.ts` with `showPanel: true`; `sb-audit`'s `audit-controls.sh` flags any
component story missing this wiring.)
## Authoring source (mutually exclusive — load exactly one)
```bash
grep -q '@storybook/addon-mcp' package.json && test -f .mcp.json && echo WITH_MCP || echo WITHOUT_MCP
```
- `WITH_MCP` → `references/with-mcp.md` (MCP injects CSF3 conventions; you focus on judgment).
Tool names change between Storybook minors; the current names are in `references/storybook-surface.md`.
- `WITHOUT_MCP` → `references/without-mcp.md` (13 verification gaps + 4 critical SB10 patterns).
- Controlled components (Switch/Toggle/Checkbox/Tabs/Accordion/Select) start from
`templates/controlled-component-story.tsx` — the `useArgs` sync is what AI gets wrong.
**Where the file goes (ASK first — don't scatter the repo).** Read `storiesLocation` from
`.storybook/audit/status.md` (the single rule lives in `CONTEXT.md` §STORIES LOCATION).
- **If it's unset** (e.g. the repo already had Storybook so `sb-setup` never asked), **STOP and ASK
the user before writing any story** — never guess, never co-locate silently. Use `AskUserQuestion`
(Claude) / `request_user_input` (Codex), or a numbered list where no blocking tool exists:
> **Where should I save the stories? (everything else already lives under `.storybook/`.)**
> 1. **`.storybook/stories/`** *(recommended — one place, isolated; `src/` untouched, one removable folder)*
> 2. **Co-located** `src/**/<Name>.stories.tsx` *(for a project you own long-term)*
> 3. **A custom folder** *(you name it — still kept to that one place)*
Then record it in `.storybook/audit/status.md` as `storiesLocation: <isolated|colocated|PATH>`, make
sure `main.ts` `stories` includes that path, and proceed. Recommend option 1.
- **`isolated` (or `.storybook/stories/`)** → write under `.storybook/stories/` mirroring the tree
(`.storybook/stories/components/CourseCard.stories.tsx`), importing the component via the `@/` alias.
The dev server does not watch `.storybook/stories/`: a story file added, moved or deleted there
stays out of (or stale in) the index until Storybook restarts, and every MCP tool reports "No story
found" for it. Restart after changing files there (verified on 10.6.0; `src/` stories update live).
- **`colocated`** → `src/components/<X>/<X>.stories.tsx`. **A custom path** → write there, every story.
Whatever the answer, **all stories go to that one location** — never a mix.
Title (the in-Storybook path, separate from the file path): match `.storybook/preview.ts`
`storySort.order`; else `Components/<Domain>/<Name>`. An **Overview / Spec / hub** story is an entry
point — pin it to the **top of its root** via `storySort.order` (a root-level hub first; a per-feature
`Overview`/`Spec` first in its group's sub-order with a `'*'` tail), never let it sort alphabetically into
the middle of the content it summarizes. See `references/directory-structure.md` → "Overview / spec / hub".
## Pages — real-page capture (Mode A)
For a **page** (a route/view under `pages/` · `app/` · `routes/` · `views/`) do NOT recreate it.
Run `scripts/page-patterns.py <root>` first — per page it reports `importable`, `component`,
`layout`, `dataHook` + **`dataType`** (the mock signal), `sections` (render order, `ui/` excluded),
`gridHint`; plus `sharedSections[]`. Then pick the mode off `importable`:
- **`importable: true`** (page has a default-export component) → **import the real page as-is and
mock ONLY its data layer.** Scaffold with `scripts/scaffold-page-story.py <root> <page-suffix>`:
it emits a `Pages/<Name>` story that imports the real page + wires the detected provider
(Inertia `usePage` / router / store), with props **seeded from a factory keyed on `dataType`**
(`scripts/scaffold-factory.sh <dataType> <import-path>`, per `references/factory-patterns.md`).
Add one story per materially-different **data** state (empty / populated / error) — different
factory inputs, never different markup. **Never re-author the page's JSX** (anti-pattern 27):
the layout, columns, and components are the real page's, not yours.
- **`importable: false`** (assembled inline / no single component) → fall back to **Page Composition**
(`references/composition-patterns.md` Pattern 4): assemble from the real `sections` —
still factory-backed, still real components.
`sharedSections[]` (a section rendered by ≥2 pages) are the reusable page-pattern pieces — give each
its own `Components/*` story so pages **compose** them, not duplicate them.
## Overlays (Dialog / Modal / Sheet / Drawer / Popover) in autodocs
An open overlay portals a `position:fixed inset-0` overlay to `document.body`. Rendered **inline** on
the autodocs page (as the `Primary` block does), that overlay escapes over the *whole Docs page* —
Title, the "Real usage" section, Controls all vanish behind a blank backdrop. So for any overlay
component with `autodocs`, scope the story to its own iframe:
```ts
parameters: {
layout: 'fullscreen',
docs: { story: { inline: false, height: '640px' } }, // portal stays inside the frame; Docs prose stays readable
}
```
The story view (one story, full canvas) is unaffected — this is only for the Docs page. Don't reach
for it on non-overlay components (inline rendering is lighter).
**The page's own `dataHook` is just its data; the preview must also supply the provider TREE + root
CSS the page renders under.** Read those from `.storybook/runtime.json` (`scripts/discover-runtime.py . --out .storybook/runtime.json` if it's missing) —
`providers[]`, `rootCss`, `portals[]`, `network.needsMsw` — they're set up once in the shared preview
by `sb-setup`, so a page story rarely re-wires them. **Never re-derive by shell scan what a script
already wrote to `.storybook/*.json`** — cite the field.
## Batch (several components)
Write one story per component (each covering only its real states), then gate each with
`scripts/validate-stories.sh`. On Claude Code you can speed a batch up by writing components in
parallel with the Agent tool, but it's the same work — no special sub-agent needed.
## Gate before done
```bash
SKILL=${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}
CORE=${CLAUDE_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT}}
"$CORE/scripts/validate-stories.sh" path/to/Foo.stories.tsx
# --strict adds tsc + eslint · --diff lints only changed stories
# For the "give me confidence" verdict (setup readiness preflight + conformance in one),
# run "$SKILL/scripts/check-story-ready.sh" path/to/Foo.stories.tsx instead — CONFIDENT when discovery JSONs are present.
```
Exits non-zero on any FAIL — fix before continuing. If you wrote a `play`, also dispatch the
judgment sub-agent (see `references/validate-workflow.md`); bash can't verify a `play`
is meaningful. To make stories an agent-runnable CLI gate (headless vitest + a11y), see
`references/test-wiring.md`.