Imported from microsoft/cat-agent-skills (submissions/scrollytelling-data/SKILL.md). Install upstream with npx skills add microsoft/cat-agent-skills --skill scrollytelling-data. Copyright stays with the author.

Skill: Data Scrollytelling

Turn an uploaded dataset (.xlsx, .csv, .json, or PDF tables) into a single self-contained scroll-driven HTML data story: animated Plotly charts, counting stat cards, and narrative chapters. Plotly.js loads from a CDN at view time.

How to Talk to the User While Working

The internal workflow uses technical steps (SQL, JSON, HTML, Plotly), but never expose that vocabulary in chat messages. Narrate like a friendly analyst explaining to a non-technical stakeholder — plain language, short sentences, a little emoji. Never say "SQL," "JSON," "HTML," "Plotly," "dataframe," or "aggregation" in a user-facing message.

Send a checkpoint per meaningful sub-step, not one per phase. Phase 1 (inspect) and Phase 3 (aggregations) can each run 30–60+ seconds silently, which feels stalled — break them into several short updates as work happens. Don't number them (no Step X/Y); lead with an emoji and a short, plain-language title.

Example checkpoints (an illustrative full run — drop any line that doesn't apply, e.g. no time or geo column, and split a slow phase into more):

• 👀 "Taking a first look at your data..."
• 📐 "Found <N> rows and <N> columns — checking what each one means..."
• 🏷️ "Spotted your key categories: <examples>..."
• 🔢 "Counting up the totals..."
• 🏆 "Ranking your top performers — your #1 is <top entity>..."
• 🥧 "Breaking things down by category..."
• 📈 "Mapping out the trend over time..." (skip if no time column)
• 🗺️ "Plotting things out by location..." (skip if no geo column)
• ✍️ "Writing the story chapters..."
• ✅ "Your story is ready! Take a look 👇"

Give ranking and chapter-writing extra beats — that's where the user most wants to feel progress. Fold a real finding into the update (the actual #1, the size of the gap, the standout insight) with real names and numbers ("🥇 Sales leads with 3,120 hours saved..."), not placeholders. Keep every line short — a few words plus one emoji — and save the substance (findings, chapter takeaways) for the delivered story itself.

Supported Input Formats

• .xlsx / .xls — via openpyxl / pandas
• .csv — via pandas
• .json — via json + pandas
• .pdf — extract tables via pdfplumber, then treat as tabular data
• Any tabular data with at least one numeric column and one categorical column

Demo Mode

If the user asks for a demo, an example, or "show me what this does" without uploading a dataset of their own, build the story from the bundled workbook at assets/Basketball_Demo_Data.xlsx (runtime path /app/skills/scrollytelling-data/assets/Basketball_Demo_Data.xlsx) and follow references/DEMO-MODE.md for the full demo workflow (sheet join, demo theme, ~5-chapter cap). If the user later uploads their own dataset, use that instead.

The Five-Phase Workflow

Phase 1 — Inspect & Understand the Data

Always start here. Never skip.

# XLSX flow only (CSV/JSON/PDF use their own inspect path, not this xlsx metadata artifact)
python /app/skills/analyzing-xlsx/scripts/preprocess.py <xlsx_file>
cat /app/workspace/_artifacts/<hash>/xlsx/1.0/metadata.json

Sandbox file access — run the preprocessor first, every time. In the Copilot Studio sandbox, uploaded files under /app/uploads/ are gated: reading one directly with pandas/openpyxl (or cd-ing into /app/uploads/ first) gets blocked before the preprocessing step has run. Always run preprocess.py against the uploaded file before any other tool touches it — that's what satisfies the gate and unlocks direct reads for the rest of the workflow. If a read gets blocked, the fix is to run the preprocessor, not to retry the same read or work around the block.

Shell execution in this sandbox. Some Python scripts in this environment need to be run through /usr/local/bin/sandbox-exec rather than invoked directly (e.g. sandbox-exec python script.py <args>). If a plain python/python3 invocation fails or is blocked, retry it wrapped in sandbox-exec before assuming the script itself is broken.

Identify:

• Sheet names and row/column counts
• Which columns are categorical (names, types, states, categories)
• Which columns are numeric (costs, counts, scores, durations)
• Which column is the primary entity (the "who" or "what" — e.g. player, product, customer)
• Which column is the primary metric (the "how much" — e.g. revenue, quantity, score)
• Whether there is a time column (enables timeline chart)
• Whether there is a geography column (state, country → enables map)
• If the file has multiple sheets, whether they relate via a shared key (e.g. a foreign-key-style ID column) — this determines whether they need to be joined in SQLite before aggregating (see Phase 3)

Run SQL via SQLite to compute:

• Total count of records
• Sum/avg of primary metric
• Top N entities by metric
• Distribution of categorical columns (outcomes, types, categories)
• Monthly/weekly/daily aggregation if time column exists
• Geographic aggregation if geo column exists

Phase 2 — Find the Story

Before writing a single line of HTML, answer these questions from the data:

Each answer becomes a chapter. Aim for 6–8 chapters. Too few = shallow. Too many = boring.

Plan once, then commit. Before touching any code, write your final chapter list — for each chapter: its title, its chart type, and the one-line insight it delivers. Make every skip / merge / split / reorder decision here. Aim to settle this in a single planning pass; allow yourself at most one revision if something is clearly wrong, then the list is locked. Repeatedly re-deriving the story — re-counting chapters, reshuffling order, second-guessing chart choices — is the single biggest time sink, so do the thinking now and stop.

Phase 3 — Prepare the Data as JSON

Extract all data needed for charts and inject it as a JavaScript constant in the HTML. This is what makes the file self-contained.

import json, sqlite3, pandas as pd

# Load data
conn = sqlite3.connect(":memory:")

# Single-sheet / CSV / JSON: one dataframe, one table
df = pd.read_excel(path) # or pd.read_csv, etc.
df.to_sql("data", conn, index=False)

# Multi-sheet Excel: preferred live-workflow approach is to load EVERY sheet
# into its OWN table (named after the sheet), then JOIN them in SQL. This is
# how sheet-to-sheet relationships (e.g. a "transactions" sheet referencing a
# "customers" sheet by ID) are normally resolved. (A reference script may use
# a pragmatic pre-merge, but the runtime workflow should still favor SQL joins.)
sheets = pd.read_excel(path, sheet_name=None)  # dict of {sheet_name: df}
for sheet_name, sheet_df in sheets.items():
    sheet_df.to_sql(sheet_name, conn, index=False)
# Example: SELECT * FROM transactions JOIN customers USING (customer_id)

# Run all aggregations
summary = pd.read_sql_query("SELECT ...", conn)
monthly = pd.read_sql_query("SELECT ...", conn)
# etc.

# Serialize everything to one dict
data = {
    "total_count": int(...),
    "total_value": float(...),
    "top10_names": [...],
    "top10_values": [...],
    "monthly_x": [...],
    "monthly_y": [...],
    "categories": [...],
    "category_values": [...],
    "state_codes": [...],
    "state_values": [...],
    "worst_records": df_worst.to_dict(orient='records'),
}
with open("/app/workspace/story_data.json", "w") as f:
    json.dump(data, f)

Save to /app/workspace/story_data.json. Then build the HTML in a separate script that reads this file, so the two phases are independent and the HTML builder can be re-run without re-querying.

Phase 4 — Build the HTML Story

Write the HTML to /app/workspace/build_story.py using the create tool, then run it with python3. Always write to a file — never use a heredoc with nested quotes.

Build the chapter list you locked in Phase 2 — don't re-plan while coding. Write the script in one pass that renders exactly that committed plan. Don't re-order, re-count, or reconsider chart choices mid-build; the scaffold below is a rendering reference for a plan you've already decided, not a prompt to decide again.

Build every chart and animation through the shared helpers in this file — renderCategoryBar/renderDonut/BG/CFG in the Chart Selection Guide, and the counter/leaderboard/record-card code in Animated Components — rather than re-typing traces per chapter. Each carries bug-fix rules (length asserts, single-fire counters) that keep charts and counters from silently dying. Only the conditional chart types (maps, scatter/bubble, dual-axis yaxis2) live in references/CHART-PATTERNS.md — open it only when the data actually shows one of those shapes.

Default story scaffold (a starting point, not a fixed recipe):

[HERO]          Full-screen dramatic title + subtitle + scroll cue
[STAT SPLASH]   Animated counting stat cards (5–6 key numbers)
[CHAPTER]       Geography / where (choropleth map)
[CHAPTER]       Method / tool / type (bar chart)
[CHAPTER]       Category / what (horizontal bar)
[CHAPTER]       Outcomes / distribution (pie or donut)
[CHAPTER]       Timeline / when (dual-axis line chart)
[DIVIDER]
[CHAPTER]       Leaderboard — top N entities (animated rows + bars)
[DIVIDER]
[CHAPTER]       Worst/best 5 individual records (animated cards)
[DIVIDER]
[CHAPTER]       Closing chapter — takeaways, recommendations, news/context, or benchmark (AI decides which fits)
[EPILOGUE]      One-paragraph verdict + badge

This scaffold keeps you off a blank page. Keep the hero → stat splash → epilogue bookends almost always — they make it feel like a complete story. The chapters in between are the plan you already committed to in Phase 2 — you decided there which to skip, merge, split, reorder, or add based on this dataset's strengths. Render that plan now in that order; don't re-open those decisions while writing.

Chapter layout pattern (two-column):

<section class="chapter" id="ch1">
  <div class="c-inner">          <!-- text LEFT, chart RIGHT -->
    <div class="reveal">
      <div class="c-num">Chapter 01</div>
      <h2 class="c-head">Narrative <em>headline here.</em></h2>
      <p class="c-body">2–3 sentences. <strong>Bold the key insight.</strong></p>
    </div>
    <div class="chart-box reveal" style="transition-delay:.2s">
      <div id="cChart1" style="height:460px"></div>
    </div>
  </div>
</section>

Use .c-inner.flip to alternate text left/right between chapters for visual rhythm. Use .c-inner.full for full-width chapters (timeline, leaderboard, worst records).

Rendering contract — two rules that keep charts and counters from silently dying:

1. One renderSection(id) function, driven by both the observer and a safety-net. Put all lazy chart/section rendering in a single named renderSection(id) function. Drive it from the IntersectionObserver and from a 2.5s setTimeout fallback. The observer alone is not enough — if it never fires (no scroll, reduced-motion, layout quirk), sections stay blank. The safety-net must do all three things the observers do — reveal, count, and render:

   function renderSection(id) {
     if (rendered[id]) return; rendered[id] = true;
     if (id==='ch-map')  { /* Plotly.newPlot('mapChart', ...) */ }
     // ...one branch per section...
   }
   const sectionObs = new IntersectionObserver(es =>
     es.forEach(e => { if (e.isIntersecting) renderSection(e.target.id); }), {threshold:0.15});
   document.querySelectorAll('section[id]').forEach(s => sectionObs.observe(s));
   
   setTimeout(() => {                       // guaranteed fallback
     document.querySelectorAll('.reveal').forEach(el => el.classList.add('vis'));
     document.querySelectorAll('.num[data-target]').forEach(n => counter(n)); // counters too!
     document.querySelectorAll('section[id]').forEach(s => renderSection(s.id)); // charts too!
   }, 2500);

   Dispatch on the SECTION id, not the chart-div id. The observer watches section[id] elements, so e.target.id is the section's id (e.g. ch-map), not the inner chart div (mapChart). Match renderSection branches against the section ids or every branch silently no-ops.

2. Syntax-check the generated JS with node --check before shipping. One stray character — most commonly an unescaped apostrophe inside a single-quoted string ('Chapter 06 · Who's Buying') — throws a SyntaxError that kills the entire <script> block, so nothing runs: blank charts and every counter frozen at 0. Escape apostrophes (Who\'s) and run node --check on the script before delivery. scripts/build_story.py does this automatically after writing the file.

Phase 5 — Deliver

Write the file to /app/created/ (per the File Naming Convention below) so it's returned as an attachment:

with open("/app/created/<dataset-name>-story-YYYY-MM-DD.html", "w") as f:
    f.write(html)

Chart Selection Guide

Default to a small set of fast, robust charts — these cover almost every dataset and rarely need bespoke math. Reach for them first and reuse them freely; repeating a bar or donut across chapters is completely fine.

Advanced charts (heatmap, treemap/sunburst, funnel, radar, sankey, waterfall, box/histogram) are opt-in only. They cost far more per-run reasoning (custom normalization, pivoted matrices, per-axis scaling) and are a top cause of slow runs. Use one only when its exact precondition is met and a default chart genuinely can't tell the story — never just to avoid repeating a bar or donut.

Scatter/bubble charts: plot at most ~10 entities, always. A scatter of 20–30+ points is an unreadable clump of overlapping, mislabeled bubbles — pick the top ~10 by the size metric (or the story's focus) and plot only those, even if you hand-roll the trace instead of using renderScatterBubble() (which caps at 10 automatically). Keep the chapter copy consistent ("top 10", not "top 30").

Keep every chart readable on the dark theme (spread ...BG, set explicit label/tick colors) and render it through renderSection(id) with a real height on the div. Build all bars, donuts, and rankings through the shared helpers below so the bug-fix rules (length asserts, ordering, low-variance zoom, no legend collision) apply consistently instead of being re-typed per chapter.

Base Layout (BG / CFG)

Spread ...BG into every chart's layout for dark-theme consistency:

const BG = {
  paper_bgcolor:'#111', plot_bgcolor:'#111',
  font:{color:'#bbb', size:11},
  margin:{t:16, b:52, l:58, r:18},
  xaxis:{gridcolor:'#1e1e1e', linecolor:'#2a2a2a'},
  yaxis:{gridcolor:'#1e1e1e', linecolor:'#2a2a2a'}
};
const CFG = {responsive:true, displayModeBar:false};

Category Bar — renderCategoryBar()

Charts like "resolution time by priority" (Low/Medium/High/Critical vs. hours) are a classic failure point: vertical columns misaligned to the category axis, or a category silently missing its bar — usually because each chapter's bar trace is hand-typed slightly differently. Use one shared helper for every "category compared to a metric" bar chart, never a bespoke Plotly.newPlot per chapter:

// Generic helper for ANY category-vs-metric bar chart (priority, category, channel, region, etc.)
function renderCategoryBar(divId, categories, values, opts = {}) {
  if (categories.length !== values.length) {
    throw new Error(`${divId}: categories/values length mismatch (${categories.length} vs ${values.length})`);
  }
  // Pair, then sort/order once from a single source of truth — never build x/y as two separately-derived arrays.
  const order = opts.categoryOrder;                 // optional fixed order, e.g. ['Critical','High','Medium','Low']
  const pairs = categories.map((c, i) => [c, values[i]]);
  if (order) pairs.sort((a, b) => order.indexOf(a[0]) - order.indexOf(b[0]));
  else pairs.sort((a, b) => b[1] - a[1]);            // default: descending by value

  const cats = pairs.map(p => p[0]);
  const vals = pairs.map(p => p[1]);

  // Low-variance guard: zoom the axis instead of always starting at 0 (see "Low-Variance Rankings")
  const spread = (Math.max(...vals) - Math.min(...vals)) / Math.max(...vals);
  const floor  = spread < 0.15 ? Math.min(...vals) - (Math.max(...vals) - Math.min(...vals)) * 0.5 : 0;

  Plotly.newPlot(divId, [{
    type: 'bar', orientation: 'h', y: cats, x: vals,   // always horizontal: category on y, metric on x
    marker: { color: opts.color || '#ff5fa2', opacity: .9 },
    text: vals.map(v => opts.fmt ? opts.fmt(v) : v), textposition: 'outside', textfont: { color: '#aab' },
    hovertemplate: `%{y}: %{x}${opts.unit || ''}<extra></extra>`
  }], {
    ...BG,
    xaxis: { ...BG.xaxis, title: opts.xTitle || '', range: [floor, Math.max(...vals) * 1.15] },
    yaxis: { ...BG.yaxis, type: 'category', categoryorder: order ? 'array' : 'total ascending', categoryarray: order, automargin: true },
    margin: { t: 20, b: 50, l: 140, r: 60 }
  }, CFG);
}

Enforced rules: orientation:'h' with categories on y/values on x; x/y derived from one paired sorted list (never two arrays that can drift); categories never filtered independently of values; explicit categoryorder/categoryarray; and a length assert that fails loudly instead of rendering a phantom label with no bar.

Donut / Pie — renderDonut()

Use a donut/pie at most once per story. If another column also tells a share/composition story, reach for a different shape (treemap, 100% stacked bar, funnel, or a plain bar) rather than a second donut. A donut's slices are already labeled directly, so a legend is redundant and collides (Plotly places it right, overlapping slice labels into tangled text like "In DevelopmentPiloting"). Never render a donut with both on-slice labels and a legend:

// Generic donut for ANY category share/composition (status, type, segment, ...).
function renderDonut(divId, labels, values, opts = {}) {
  if (labels.length !== values.length)
    throw new Error(`${divId}: labels/values length mismatch (${labels.length} vs ${values.length})`);
  Plotly.newPlot(divId, [{
    type: 'pie', hole: 0.55,
    labels, values,
    sort: true, direction: 'clockwise',
    textinfo: 'label+percent',        // label the slice itself...
    textposition: 'inside',           // ...inside the ring
    insidetextorientation: 'horizontal',
    textfont: { color: '#fff', size: 12 },
    marker: { line: { color: '#111', width: 2 } },  // thin gap between slices
    hovertemplate: '%{label}: %{value} (%{percent})<extra></extra>'
  }], {
    ...BG,
    showlegend: false,                // ← the fix: no redundant, colliding legend
    margin: { t: 20, b: 20, l: 20, r: 20 }
  }, CFG);
}

If a slice is too thin for an inside label, don't turn the legend back on — merge tiny tail categories into an "Other" slice, or switch to a horizontal bar for many small categories.

Low-Variance Rankings — When Every Value Looks the Same

Before building any ranked bar list (leaderboards, top-N charts), check the spread: (max - min) / max. If it's small (under ~15–20%, e.g. CSAT 3.98–4.08, or 94%–97%), a 0-based bar renders every row visually identical. Fix — pick one:

1. Zoom the axis to the actual data range, not 0. Compute pct as (value - floor) / (ceil - floor) * 100 where floor is slightly below the min (or a meaningful baseline like a group average) and ceil slightly above the max. Default fix for leaderboard rows.
2. Show delta-from-baseline — plot value − average as a diverging bar (positive in --fire, negative in muted gray) instead of the absolute score.
3. Switch chart type (dot/lollipop on a zoomed axis, or a table with a per-row sparkline/tint) when a bar list still isn't informative.
4. Never leave a 0-based bar as the only visual for a low-variance metric.

State the actual spread in the chapter copy (e.g. "the gap between #1 and #10 is just 0.10 points") so the visual and narrative agree.

Stat Card Rules

Short numbers (≤6 chars, e.g. 2,000, 60, 156): use .s-card with flex:1; min-width:220px

Long numbers (≥7 chars, e.g. $419,917, 1,284,330): use .s-card-wide — a standalone centered card on its own row with more horizontal space and larger font.

Rule of thumb: if the number has a $ prefix AND is 6+ digits, give it .s-card-wide.

data-suffix/data-prefix must be short and purely a unit (e.g. "M", "%", "B", "h") — never embed prose, arrows, ranges, or multi-word labels in them (e.g. never data-suffix="M tons" or data-suffix="→$4.17/kg"). The 6-vs-7 char length check applies to the full rendered string — prefix + number + suffix combined — not just the numeric part alone, since that's what actually has to fit in the card.

If a value isn't a clean countable integer or round decimal, don't force it into a counter. A counter animates one number ticking up to a target — it can't meaningfully animate a range ($3.61→$4.17/kg) or a compound unit. When the raw stat is a range, ratio, or multi-part value, either display it as static (non-animated) text, or redesign it as something genuinely countable — e.g. replace "$3.61→$4.17/kg" with a "16% price increase" stat instead.

<!-- Row 1: short numbers -->
<div class="stat-row">
  <div class="s-card reveal">...</div>
  <div class="s-card reveal">...</div>
  <div class="s-card reveal">...</div>
</div>
<!-- Row 2: large dollar amount gets its own wide card -->
<div class="stat-row">
  <div class="s-card-wide reveal">...</div>
</div>

Animated Components

Scroll-Triggered Counter

Add data-target="12345" and optionally data-prefix="$" to any .num element; a dedicated observer calls counter(el) when its card scrolls into view.

<div class="num" data-prefix="$" data-target="419917">$0</div>

Never start counters from the same IntersectionObserver that drives generic .reveal transitions. That observer fires for every reveal element in the same batch, so a counter can start (or get skipped) while the card is still mid-opacity-transition — numbers race, stall, or never animate. Use two independent, single-purpose observers, and guard counter() so it can never fire twice:

function counter(el) {
  if (el.dataset.counted) return;   // guard: fire at most once per element
  el.dataset.counted = '1';
  // ...existing animation logic...
}

// Visual reveal only — no counter logic here.
const revealObs = new IntersectionObserver((entries) => {
  entries.forEach(e => { if (e.isIntersecting) e.target.classList.add('vis'); });
}, {threshold: 0.15});
document.querySelectorAll('.reveal').forEach(el => revealObs.observe(el));

// Isolated counter observer: higher threshold so the card is substantially visible before
// numbers move, and unobserve immediately so it can never double-trigger.
const counterObs = new IntersectionObserver((entries) => {
  entries.forEach(e => {
    if (!e.isIntersecting) return;
    e.target.querySelectorAll('.num[data-target]').forEach(n => counter(n));
    counterObs.unobserve(e.target);
  });
}, {threshold: 0.3});
document.querySelectorAll('.num[data-target]').forEach(n => {
  counterObs.observe(n.closest('.reveal') || n);
});

Leaderboard Rows

Rows slide in from the left one by one, then value bars grow right. Check variance first (see "Low-Variance Rankings"); if top10 values are tightly clustered, scale from a zoomed floor instead of 0:

const vals = D.top10_values;
const spread = (Math.max(...vals) - Math.min(...vals)) / Math.max(...vals);
const floor = spread < 0.15 ? Math.min(...vals) - (Math.max(...vals) - Math.min(...vals)) * 0.5 : 0;
const ceil = Math.max(...vals) * 1.02;
const pct = v => ((v - floor) / (ceil - floor)) * 100;

function buildLeaderboard(){
  D.top10_names.forEach((name, i) => {
    const row = document.createElement('div');
    row.className = 's-row';
    row.style.transitionDelay = (i * .07) + 's';
    row.innerHTML = `...`; // use pct(D.top10_values[i]) for bar width, not a raw 0-based ratio
    list.appendChild(row);
  });
  list.querySelectorAll('.s-row').forEach((r, i) => {
    setTimeout(() => {
      r.classList.add('vis');
      setTimeout(() => {
        r.querySelector('.s-bar').style.width = r.querySelector('.s-bar').dataset.pct + '%';
      }, 350);
    }, i * 90);
  });
}

Record Cards

5 cards fly up from below, staggered:

wrap.querySelectorAll('.i-card').forEach((c, i) => {
  setTimeout(() => c.classList.add('vis'), i * 110);
});

The Closing Chapter — Decide What the Reader Needs, Don't Default to One Format

Every story needs a payoff beyond "here's what happened" — but what kind of payoff depends on the dataset and audience, not a fixed template. Add this as the second-to-last chapter, right before the epilogue, derived from The Fix story beat (Phase 2) — but "The Fix" is a prompt to figure out the right closing move, not a mandate that it's always a takeaways/action-item list.

Decide which closing format actually serves this data. Common options:

Pick one (or blend two, e.g. takeaways + one grounding external stat) based on what would actually be useful to this reader for this dataset — don't reach for "what to do about it" by default if the data is purely descriptive, and don't reach for takeaways if the data screams for action.

Content rules (apply regardless of format chosen):

• 3–5 items, each grounded in a specific number or finding already shown earlier in the story (don't introduce new data-derived claims here — external/news items are the one exception, and those must be clearly sourced)
• Keep each item to 1–2 sentences — this is a scannable list, not another narrative chapter
• End with a short closing line appropriate to the format (a "Next step" for recommendations, a one-line synthesis for takeaways, a source/date note for news items)

Markup pattern (numbered cards, full-width — same structure works for any format, only the copy changes):

<section class="chapter" id="ch8">
  <div class="c-inner full">
    <div class="reveal">
      <div class="c-num">Chapter 08</div>
      <h2 class="c-head"><!-- headline matches the chosen format, e.g. "What to do <em>about it.</em>" or "What to <em>remember.</em>" or "What's happening <em>right now.</em>" --></h2>
      <p class="c-body">One sentence framing the closing format you chose.</p>
    </div>
    <div class="takeaway-grid">
      <div class="t-card reveal">
        <div class="t-index">01</div>
        <h3>Insight-driven headline</h3>
        <p>One sentence citing the specific number/finding (or, for the news format, the headline + source), plus why it matters.</p>
      </div>
      <!-- repeat for 3–5 cards total -->
    </div>
    <div class="next-steps reveal">
      <strong><!-- "Next step:" / "Bottom line:" / "As of <date>:" depending on format --></strong> One concrete closing line.
    </div>
  </div>
</section>

If choosing "News & Context": confirm web access works before committing the chapter to it (a quick test fetch, not an assumption), cite the source and date for each headline/stat pulled in, and keep the dataset's own findings as the throughline — external context should support the data's story, not replace it.

.t-card styling should match .s-card (dark card, --fire accent border-left, staggered reveal like .i-card). Number each card (.t-index) so it reads as a ranked action list, not a wall of text.

Narrative Writing Guidelines

Good scrollytelling copy is short, punchy, and specific. Use this formula per chapter:

1. Setup (1 sentence): state the angle
2. Specific finding (1 sentence with a <strong> entity name): name the data winner/loser
3. Implication or tension (1 sentence): why does it matter, or what's surprising

Good example:

The West leads on volume — its top teams post the most total points by a wide margin. But efficiency tells a different story: the highest points-per-game average belongs to a single standout player who quietly outscores every roster around them.

Bad example (too vague):

There are many different teams. Some score more than others. The data shows interesting patterns.

Make headlines dramatic. Use <em> (styled with the accent color) for the key word:

<h2 class="c-head">The margin<br>of victory <em>matters.</em></h2>

The epilogue should echo the hero's opening claim and close the loop:

• Hero: "One Number Explains the Season."
• Epilogue: "The data is clear. The standings never lie."

Design System (Dark Neon Theme)

:root {
  --fire:  #ff5fa2;  /* primary accent — headlines, numbers, bars (pink) */
  --ember: #a855f7;  /* gradient end — buttons, glow (purple) */
  --text: #f0ede8;   /* body text */
  --muted: #999;     /* secondary text, labels */
}
/* Background layers */
body        { background: #0a0a0a }
.chapter    { background: #0a0a0a or #0d0d0d (alternate) }
.chart-box  { background: #111; border: 1px solid #1e1e1e }
.s-card     { background: #111; border: 1px solid #222 }

Favor green, pink, and purple accents — they contrast well with the dark background. For a different mood, replace --fire and --ember and update the radial gradient in #hero and #epilogue.

File Naming Convention

Lowercase kebab-case dataset name, the word story, then the current date. Example: sales-data-story-2025-07-21.html

Always deliver to /app/created/ so it is returned as an attachment.

Gotchas

Quick-Start Checklist

• Inspect the data — identify entity, metric, time, geo columns
• Run SQL aggregations — top 10, totals, distributions, monthly, geo
• Serialize all chart data to /app/workspace/story_data.json
• Write /app/workspace/build_story.py using the create tool
• Run python3 /app/workspace/build_story.py — verify "Done! X chars written"
• Syntax-check the embedded JS with node --check (build_story.py does this automatically) — a single bad char blanks the whole page
• Closing chapter — pick the right format (takeaways / recommendations / news-context / benchmark) for this data, 3–5 items grounded in earlier findings
• Stat cards: short numbers → .s-card, large dollar amounts → .s-card-wide
• Chart heights: minimum 460px per chart div
• Headline font: clamp(2.2rem, 4.5vw, 4rem) minimum
• Output file → /app/created/<dataset-name>-story-YYYY-MM-DD.html