Add PRD and phased roadmap for Firefly-style cloud platform

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UUceKW1u6RsURcq5xZwh34
2026-10-20 07:39:40 +00:00
16 changed files with 83 additions and 1406 deletions
@@ -1,163 +0,0 @@
 ---
 name: ad-creative-video
 description: This skill should be used when the user asks to "make a video ad", "create an animated ad", "build a performance/UGC-style ad", "batch-produce ad creative variations", "generate ad variants for A/B testing", "swap headline/offer/CTA across many ad versions", or "export one ad in multiple aspect ratios for Meta/TikTok/Reels". Covers ad hook structures, hook→CTA message-match, data-driven variant generation (1 template × CSV = N ads), multi-aspect export, and platform specs.
 version: 0.1.0
 ---
 # Ad Creative Video
 Build ONE motion-graphics ad template, then mass-produce variants from a data table — different hooks, offers, CTAs, and products — so a performance marketer can A/B test dozens of versions across placements and aspect ratios. The craft is making the template fully data-driven and message-matched, so 40 variants stay on-brand and only the tested variable changes.
 ## When to use
 - Performance/UGC-style video ads where the first 1.5–3s must stop the scroll.
 - Batch variant generation: one template rendered against many rows (headline/offer/CTA/product swaps) for A/B testing.
 - Multi-aspect export of the same ad for Meta Feed, Reels/Stories, TikTok, YouTube.
 Not for sale/discount countdown promos, and not for testimonial/review-driven ads — those are different structures. Stay on the batch-variation, test-everything angle.
 ## The two rules that make variants worth running
 1. **Isolate one variable per variant.** A test only teaches something if exactly one thing changes. Hold layout, motion, colors, and timing constant; swap *only* the field under test (hook, or offer, or CTA). Mixing two changes makes the winner uninterpretable.
 2. **Message-match the hook to the CTA.** The promise made in the first 3 seconds must be the promise the button pays off. A "stop wasting 2 hours a day" hook ends on "Save 2 hours — try free," not a generic "Shop now." The hook and CTA are a matched pair in every row of the data table.
 ## Ad anatomy (15–30s)
 | Beat | Job | Budget |
 |---|---|---|
 | Hook (0–3s) | Stop the scroll; state the problem or pattern-interrupt | 0–3s |
 | Context | Make the viewer feel the problem / agitate | 3–8s |
 | Payoff | Show the product as the solution, one clear benefit | 8–18s |
 | Proof | One concrete number, demo, or result | 18–25s |
 | CTA (last 3s) | Single action, message-matched to the hook | hold ≥2s |
 Land the hook's emotional trigger before the 2-second mark — judgment forms in ~1.7s and scroll speed keeps rising. Test the hook FIRST; it moves CPA more than any other element.
 ## Hook taxonomy (the field you vary most)
 | Hook type | Template | When it wins |
 |---|---|---|
 | Problem-first | "I didn't realize how much X was costing me until…" | pain is felt but unnamed |
 | Pattern interrupt | "This is going to sound controversial, but…" | crowded feed, generic category |
 | Curiosity gap | "Nobody talks about the one thing that…" | educated, skeptical audience |
 | Immediate benefit | "Cut your X in half in 7 days" | clear, quantifiable outcome |
 | Direct callout | "If you do X every morning, stop." | sharp audience segment |
 Keep a column of 5–10 hook strings and let the batch render produce one ad per hook against the same body — that is the cleanest hook test possible.
 ## Data-driven template (everything is a prop)
 Hardcode nothing the marketer might A/B test. The composition reads a single `variant` object; the renderer never edits the component. Make every animated value a pure function of `useCurrentFrame()` so each frame renders deterministically (no CSS transitions, no library timers — they desync the render).
 ```tsx
 import {useCurrentFrame, interpolate, AbsoluteFill, useVideoConfig} from "remotion";
 type Variant = {hook: string; benefit: string; proof: string; cta: string; bg: string; accent: string};
 export const AdTemplate: React.FC<{v: Variant}> = ({v}) => {
  const frame = useCurrentFrame(); const {fps} = useVideoConfig();
  const hookIn = interpolate(frame, [0, 8], [40, 0], {extrapolateRight: "clamp"});
  const ctaIn  = interpolate(frame, [fps * 22, fps * 23], [0, 1], {extrapolateLeft: "clamp", extrapolateRight: "clamp"});
  return (
    <AbsoluteFill style={{background: v.bg, fontFamily: "Inter, sans-serif"}}>
      <h1 style={{transform: `translateY(${hookIn}px)`, opacity: interpolate(frame,[0,8],[0,1])}}>{v.hook}</h1>
      {frame > fps * 8  && <p>{v.benefit}</p>}
      {frame > fps * 18 && <strong>{v.proof}</strong>}
      <button style={{opacity: ctaIn, background: v.accent}}>{v.cta}</button>
    </AbsoluteFill>
  );
 };
 ```
 See `references/batch-ad-pipeline.md` for the complete template, the CSV→props parser with validation, and the brand-lock theme object.
 ## 1 template × CSV = N variants
 The payoff: author the ad once, then let a CSV drive the matrix. Each row is one fully-formed, message-matched variant. Render once per row.
 ```bash
 # variants.csv: id,hook,benefit,proof,cta,bg,accent  → one MP4 per row
 node csv-to-props.js variants.csv ./props        # writes props/<id>.json per row
 for f in props/*.json; do
  id=$(basename "$f" .json)
  npx remotion render AdTemplate "out/${id}.mp4" --props="$f"
 done
 ```
 To build a clean test matrix, hold every column constant and vary one: 5 hooks × 1 body = 5 ads (hook test); then take the winning hook × 3 CTAs (CTA test). `references/batch-ad-pipeline.md` has a matrix generator that expands variable lists into the CSV so combinatorial tests stay one-variable-at-a-time.
 ## Multi-aspect export
 Each placement wants a native aspect. Compose against a center safe zone so one master reframes cleanly to all of them — don't letterbox, re-center.
 | Aspect | Placements | Resolution | Keep-clear safe zone |
 |---|---|---|---|
 | 9:16 | Reels, Stories, TikTok In-Feed | 1080×1920 | top ~14%, bottom ~20–35% (UI/caption), sides ~6% |
 | 4:5 | Meta Feed (best feed real estate) | 1080×1350 | center 90%; CTA out of bottom 10% |
 | 1:1 | Feed, broad reach | 1080×1080 | center square survives every crop |
 | 16:9 | YouTube, in-stream | 1920×1080 | center 90% |
 Design the hook text, product, and CTA inside the **1:1 center square** so they survive every crop. Drive aspect from a prop and render each per variant:
 ```bash
 for ar in 9x16 4x5 1x1; do
  npx remotion render AdTemplate "out/${id}-${ar}.mp4" --props="props/${id}.json" \
    --props-merge="{\"aspect\":\"${ar}\"}"
 done
 ```
 Keep critical content out of the 9:16 bottom third — that is where the platform stacks captions, the CTA button, and the username. `references/platform-specs.md` has exact pixel safe zones and per-platform duration/format limits.
 ## Output checklist
 - One variable per variant; layout/motion/timing identical across a test.
 - Hook lands its trigger before 2s; hook and CTA are message-matched in every row.
 - Every animated value is a pure function of `useCurrentFrame()` — no CSS/library timers.
 - Dataset is a prop; nothing testable is hardcoded; one template renders the whole CSV.
 - Each variant exported 9:16 + 4:5 + 1:1 (+16:9 if needed), key content in the center square.
 - Brand stays locked across all variants via a single theme object.
 ## Deliver & verify (rendered stills → MP4)
 > **Packaged helper** (`scripts/`): tile your stills with `scripts/contact-sheet.sh sheet.png f-hook.png f-mid.png f-end.png`, then assert the encode with `scripts/probe-mp4.sh out.mp4 [WxH] [fps]`. See `scripts/README.md`.
 **Output contract:**
 - A Remotion ad template registered as `<Composition>` (+ zod `schema` + `defaultProps`), every animated value frame-driven (no CSS transitions / library timers / `Date.now()` / `Math.random()`).
 - Deliverable = the rendered `out/*.mp4` per variant per aspect (plus the project + CSV, so the marketer re-renders on new rows).
 **Verify loop — render stills → inspect → encode.** Cheap PNGs first, full encode only once they're clean. Render with the **shipped** props (the real row), not just `defaultProps`.
 ```bash
 # Frame-exact stills across the hook→CTA arc, with a real variant's props
 npx remotion still AdTemplate out/f-hook.png --frame=12  --props=props/v1.json   # hook readable < 2s
 npx remotion still AdTemplate out/f-mid.png  --frame=300 --props=props/v1.json   # benefit/proof
 npx remotion still AdTemplate out/f-cta.png  --frame=689 --props=props/v1.json   # CTA, message-matched
 # inspect each: fidelity (hook / offer / proof / CTA text exact, brand bg+accent right)
 #   AND artifacts (text overflow, off-canvas, CTA inside the 9:16 bottom third, missing font, wrong row binding)
 ```
 **Multi-aspect / batch — verify one variant in EACH aspect before batch-rendering the matrix.** A layout bug repeats across every row × aspect; catch it once.
 ```bash
 for ar in 9x16 4x5 1x1; do                                  # one representative variant, every target aspect
  npx remotion still AdTemplate "out/v1-${ar}.png" --frame=300 \
    --props=props/v1.json --props-merge="{\"aspect\":\"${ar}\"}"
 done
 # stills clean in all aspects? → then batch-render every row × aspect:
 for f in props/*.json; do id=$(basename "$f" .json); for ar in 9x16 4x5 1x1; do
  npx remotion render AdTemplate "out/${id}-${ar}.mp4" --props="$f" --props-merge="{\"aspect\":\"${ar}\"}"
 done; done
 npx remotion render AdTemplate out/demo.gif --codec=gif --props=props/v1.json   # README demo
 ```
 **Before you finish:**
 1. `npx remotion still` renders cleanly at hook / mid / CTA — no errors, no missing fonts/assets.
 2. Hook/offer/proof/CTA text exact and brand colors right; nothing in the 9:16 bottom third or outside the center safe zone.
 3. Frame-driven only — no CSS/library timers, `Date.now()`, or `Math.random()`.
 4. Shipped row's props render correctly (not just `defaultProps`); one variable per variant holds.
 5. One variant verified in 9:16 + 4:5 + 1:1 before the batch; MP4s play; (optional) GIF for the README.
 ## Reference files
 - `references/batch-ad-pipeline.md` — the full runnable Remotion ad template, the CSV→typed-props parser with validation, a one-variable test-matrix generator, the brand-lock theme object, and the complete batch + multi-aspect render script.
 - `references/platform-specs.md` — exact 2025/2026 pixel safe zones, durations, aspect ratios, and format limits for Meta Feed/Reels/Stories, TikTok In-Feed, and YouTube, plus a hook→CTA message-match worksheet.
@@ -1,289 +0,0 @@
 # Batch Ad Pipeline — full template, parser, matrix, render
 This is the complete, runnable implementation behind the SKILL.md overview: a data-driven Remotion ad template, a CSV→props parser with validation, a one-variable test-matrix generator, a brand-lock theme, and the batch + multi-aspect render scripts. Hardcode nothing a marketer might want to A/B test.
 ## 1. The brand-lock theme
 Keep every brand-controlled value in one object so 40 variants stay consistent and only the *tested* fields change per row. The variant data supplies copy and per-campaign accents; the theme supplies the locked look.
 ```ts
 // theme.ts
 export const theme = {
  fontFamily: "Inter, system-ui, sans-serif",
  ink: "#0B0B0F",
  paper: "#FFFFFF",
  radius: 20,
  hookSize: 76,    // px, on the 1080-wide master
  bodySize: 44,
  proofSize: 52,
  ctaSize: 40,
  pad: 80,         // outer padding; also the side safe margin
  durationInSeconds: 25,
 } as const;
 ```
 ## 2. Aspect config
 Drive layout from a single aspect prop so one composition reframes to every placement. Dimensions and the keep-clear bottom band differ per aspect.
 ```ts
 // aspects.ts
 export type Aspect = "9x16" | "4x5" | "1x1" | "16x9";
 export const ASPECTS: Record<Aspect, {w: number; h: number; bottomSafe: number}> = {
  "9x16": {w: 1080, h: 1920, bottomSafe: 540}, // captions + CTA + username stack here
  "4x5":  {w: 1080, h: 1350, bottomSafe: 135},
  "1x1":  {w: 1080, h: 1080, bottomSafe: 110},
  "16x9": {w: 1920, h: 1080, bottomSafe: 110},
 };
 ```
 ## 3. The data-driven template
 Every animated value is a pure function of `useCurrentFrame()` — no CSS transitions, no GSAP/library timers (they desync deterministic rendering and flicker in the export). The component reads one `variant` plus an `aspect`; the renderer never edits this file.
 ```tsx
 // AdTemplate.tsx
 import {
  AbsoluteFill, useCurrentFrame, useVideoConfig, interpolate, spring, Sequence,
 } from "remotion";
 import {theme} from "./theme";
 import {ASPECTS, Aspect} from "./aspects";
 export type Variant = {
  id: string;
  hook: string;      // 0–3s scroll-stopper
  benefit: string;   // the single payoff claim
  proof: string;     // one concrete number / result
  cta: string;       // message-matched to the hook
  accent: string;    // per-campaign accent color
  bg: string;        // background color
 };
 export const AdTemplate: React.FC<{v: Variant; aspect: Aspect}> = ({v, aspect}) => {
  const frame = useCurrentFrame();
  const {fps} = useVideoConfig();
  const {bottomSafe} = ASPECTS[aspect];
  // Hook rises + fades in over the first 8 frames, holds, then eases up and out at 8s.
  const hookY = interpolate(frame, [0, 8], [48, 0], {extrapolateRight: "clamp"});
  const hookOpacity =
    interpolate(frame, [0, 8], [0, 1], {extrapolateRight: "clamp"}) *
    interpolate(frame, [fps * 7, fps * 8], [1, 0], {extrapolateLeft: "clamp", extrapolateRight: "clamp"});
  // CTA springs in for impact in the last 3s and holds.
  const ctaProgress = spring({frame: frame - fps * 22, fps, config: {damping: 14, stiffness: 120}});
  return (
    <AbsoluteFill style={{backgroundColor: v.bg, fontFamily: theme.fontFamily, color: theme.ink}}>
      <AbsoluteFill style={{padding: theme.pad, paddingBottom: bottomSafe, justifyContent: "center"}}>
        {/* HOOK 0–8s */}
        <Sequence durationInFrames={fps * 8}>
          <h1 style={{fontSize: theme.hookSize, fontWeight: 800, lineHeight: 1.05,
                      transform: `translateY(${hookY}px)`, opacity: hookOpacity}}>
            {v.hook}
          </h1>
        </Sequence>
        {/* BENEFIT 8–18s */}
        <Sequence from={fps * 8} durationInFrames={fps * 10}>
          <FadeUp delay={0}>
            <p style={{fontSize: theme.bodySize, fontWeight: 600}}>{v.benefit}</p>
          </FadeUp>
        </Sequence>
        {/* PROOF 18–22s */}
        <Sequence from={fps * 18} durationInFrames={fps * 4}>
          <FadeUp delay={0}>
            <strong style={{fontSize: theme.proofSize, color: v.accent,
                            fontVariantNumeric: "tabular-nums"}}>{v.proof}</strong>
          </FadeUp>
        </Sequence>
        {/* CTA last 3s, message-matched to the hook */}
        <Sequence from={fps * 22}>
          <button style={{
            alignSelf: "flex-start", marginTop: 40, padding: "22px 44px",
            fontSize: theme.ctaSize, fontWeight: 700, border: "none",
            borderRadius: theme.radius, color: theme.paper, backgroundColor: v.accent,
            transform: `scale(${interpolate(ctaProgress, [0, 1], [0.85, 1])})`,
            opacity: ctaProgress,
          }}>
            {v.cta}
          </button>
        </Sequence>
      </AbsoluteFill>
    </AbsoluteFill>
  );
 };
 const FadeUp: React.FC<{children: React.ReactNode; delay: number}> = ({children, delay}) => {
  const frame = useCurrentFrame();
  const o = interpolate(frame - delay, [0, 10], [0, 1], {extrapolateLeft: "clamp", extrapolateRight: "clamp"});
  const y = interpolate(frame - delay, [0, 10], [24, 0], {extrapolateLeft: "clamp", extrapolateRight: "clamp"});
  return <div style={{opacity: o, transform: `translateY(${y}px)`}}>{children}</div>;
 };
 ```
 ## 4. Register the composition (aspect + variant as props)
 ```tsx
 // Root.tsx
 import {Composition} from "remotion";
 import {AdTemplate} from "./AdTemplate";
 import {ASPECTS} from "./aspects";
 import {theme} from "./theme";
 const defaultVariant = {
  id: "demo", hook: "Spending 2 hrs/day on reports?", benefit: "Auto-build them in one click.",
  proof: "Teams save 11 hrs a week.", cta: "Save 2 hours — try free", accent: "#4F46E5", bg: "#FFFFFF",
 };
 export const RemotionRoot = () => {
  const aspect = "9x16" as const; // overridden per-render via --props
  const {w, h} = ASPECTS[aspect];
  return (
    <Composition
      id="AdTemplate"
      component={AdTemplate}
      durationInFrames={theme.durationInSeconds * 30}
      fps={30}
      width={w}
      height={h}
      defaultProps={{v: defaultVariant, aspect}}
      // Resolve real dimensions from the incoming aspect prop at render time:
      calculateMetadata={({props}) => {
        const a = ASPECTS[(props as any).aspect ?? "9x16"];
        return {width: a.w, height: a.h};
      }}
    />
  );
 };
 ```
 `calculateMetadata` lets a single composition output any aspect from the `aspect` prop — no separate composition per ratio.
 ## 5. CSV → typed props, with validation
 The dataset is the input. Parse, validate every row (a bad hex or empty CTA must fail loudly, not render a broken ad), and write one props file per variant.
 ```js
 // csv-to-props.js   usage: node csv-to-props.js variants.csv ./props [aspect]
 const fs = require("fs");
 const path = require("path");
 const [, , csvPath, outDir, aspect = "9x16"] = process.argv;
 const REQUIRED = ["id", "hook", "benefit", "proof", "cta", "accent", "bg"];
 const isHex = (s) => /^#([0-9a-f]{3}|[0-9a-f]{6})$/i.test(s);
 const parseCSV = (text) => {
  const [head, ...lines] = text.trim().split(/\r?\n/);
  const cols = head.split(",").map((c) => c.trim());
  return lines.filter(Boolean).map((line) => {
    // simple split; wrap fields containing commas in double quotes in the CSV
    const cells = line.match(/("([^"]|"")*"|[^,]*)/g).filter((_, i, a) => i < a.length - 1);
    const row = {};
    cols.forEach((c, i) => (row[c] = (cells[i] ?? "").replace(/^"|"$/g, "").replace(/""/g, '"').trim()));
    return row;
  });
 };
 const rows = parseCSV(fs.readFileSync(csvPath, "utf8"));
 fs.mkdirSync(outDir, {recursive: true});
 const seen = new Set();
 let errors = 0;
 rows.forEach((row, n) => {
  const where = `row ${n + 2} (${row.id || "no-id"})`;
  REQUIRED.forEach((k) => { if (!row[k]) { console.error(`✗ ${where}: missing "${k}"`); errors++; } });
  if (seen.has(row.id)) { console.error(`✗ ${where}: duplicate id`); errors++; }
  seen.add(row.id);
  if (row.hook && row.hook.length > 60) console.warn(`⚠ ${where}: hook >60 chars may clip in 9:16`);
  ["accent", "bg"].forEach((k) => { if (row[k] && !isHex(row[k])) { console.error(`✗ ${where}: "${k}" not a hex color`); errors++; } });
  if (errors) return;
  fs.writeFileSync(path.join(outDir, `${row.id}.json`), JSON.stringify({v: row, aspect}, null, 2));
 });
 if (errors) { console.error(`\n${errors} error(s) — no broken ads written.`); process.exit(1); }
 console.log(`✓ ${rows.length} variant prop files written to ${outDir}`);
 ```
 Example `variants.csv` (one fully message-matched ad per row):
 ```csv
 id,hook,benefit,proof,cta,accent,bg
 v01_problem,Spending 2 hrs/day on reports?,Auto-build them in one click.,Teams save 11 hrs a week.,Save 2 hours — try free,#4F46E5,#FFFFFF
 v02_interrupt,Your reporting tool is lying to you.,See the real numbers instantly.,Teams save 11 hrs a week.,See the real numbers — free,#DC2626,#0B0B0F
 v03_curiosity,Nobody talks about this reporting trick.,One click builds the whole report.,Teams save 11 hrs a week.,Try the trick — free,#059669,#FFFFFF
 ```
 Note the CTA mirrors the hook in every row — that is the message-match discipline encoded into the data.
 ## 6. Test-matrix generator — one variable at a time
 A test only teaches something if exactly one field changes. This expands a base variant plus a list of values for ONE field into a CSV, so a hook test or a CTA test stays clean and combinatorial mistakes are impossible.
 ```js
 // make-matrix.js   usage: node make-matrix.js > variants.csv
 const base = {benefit: "Auto-build them in one click.", proof: "Teams save 11 hrs a week.",
              cta: "Save 2 hours — try free", accent: "#4F46E5", bg: "#FFFFFF"};
 // Vary ONLY this field; everything else is held identical across the test.
 const FIELD = "hook";
 const values = [
  "Spending 2 hrs/day on reports?",
  "Your reporting tool is lying to you.",
  "Nobody talks about this reporting trick.",
  "Cut report time in half this week.",
  "If you build reports by hand, stop.",
 ];
 const esc = (s) => (/[",]/.test(s) ? `"${s.replace(/"/g, '""')}"` : s);
 const cols = ["id", "hook", "benefit", "proof", "cta", "accent", "bg"];
 console.log(cols.join(","));
 values.forEach((val, i) => {
  const row = {...base, [FIELD]: val, id: `${FIELD}_${String(i + 1).padStart(2, "0")}`};
  console.log(cols.map((c) => esc(row[c])).join(","));
 });
 ```
 To run a CTA test next, set `FIELD = "cta"`, paste the *winning* hook into `base.hook`, and list CTA variants. Same one-variable rule, next stage of the funnel.
 ## 7. Batch + multi-aspect render
 Render every variant in every aspect a campaign needs. Concurrency keeps large batches fast.
 ```bash
 #!/usr/bin/env bash
 # render-all.sh   one MP4 per (variant × aspect)
 set -euo pipefail
 node make-matrix.js > variants.csv          # 1) build the one-variable test matrix
 ASPECTS=("9x16" "4x5" "1x1")                 # placements this campaign needs
 for ar in "${ASPECTS[@]}"; do
  node csv-to-props.js variants.csv "props/$ar" "$ar"   # 2) typed, validated props per aspect
  for f in "props/$ar"/*.json; do
    id=$(basename "$f" .json)
    npx remotion render AdTemplate "out/${id}_${ar}.mp4" \
      --props="$f" --concurrency=4 --log=error              # 3) render
  done
 done
 echo "Done. $(ls out/*.mp4 | wc -l) ad variants in ./out"
 ```
 5 hooks × 3 aspects = 15 MP4s from one template and one command. Add a CTA stage and the same template produces the next test wave with zero new component code.
 ## Common pitfalls
 - **Two variables changed at once** — the winner is uninterpretable. The matrix generator prevents this; keep using it.
 - **Hook clips in 9:16** — long hooks overflow; the parser warns past 60 chars. Keep hooks short and high in the frame.
 - **CTA hidden behind platform UI** — anything in the 9:16 bottom third gets covered by captions and the CTA button. The `bottomSafe` padding reserves that band.
 - **Drift across variants** — accent/copy live in the data, everything else in `theme`. If a variant looks off-brand, the fix is the theme, not the row.
 - **Animation flicker in export** — caused by CSS transitions or library timers. Every value must derive from `useCurrentFrame()`.
 ---
 ## Built by the team behind iart.ai
 This skill is part of an open motion-graphics collection from iart.ai — the AI motion agent that turns data, scripts, and designs into editable motion graphics (Remotion → MP4). If you'd rather not hand-build this, iart.ai can mass-produce ad-creative variants from one template × a data table — change the text/data and re-export. → [iart.ai](https://iart.ai/?utm_source=github&utm_medium=reference&utm_campaign=ad-video-skills&utm_content=ref_footer&utm_term=ad-creative-video)
@@ -1,72 +0,0 @@
 # Platform Specs & Message-Match Worksheet
 Exact pixel safe zones, durations, aspect ratios, and format limits per placement (2025/2026), plus a worksheet for building message-matched hook→CTA pairs before they go into the CSV. Verify against each platform's current Ads Manager docs before a big spend — placements and safe zones shift.
 ## Aspect ratio → placement map
 | Aspect | Resolution | Native placements |
 |---|---|---|
 | 9:16 | 1080×1920 | TikTok In-Feed, Instagram/Facebook Reels, Stories, YouTube Shorts |
 | 4:5 | 1080×1350 | Meta Feed (largest feed real estate, recommended for Feed video) |
 | 1:1 | 1080×1080 | Meta Feed, broad/automatic placements, X |
 | 16:9 | 1920×1080 | YouTube in-stream, landscape feed, in-article |
 Master rule: design the hook, product, and CTA inside the **1:1 center square** so a single composition reframes to every aspect by re-centering, not letterboxing. Black bars on a vertical feed read as "ad" and underperform.
 ## Safe zones (keep-clear bands)
 Platform UI — captions, CTA chips, usernames, profile icons, progress bars — overlaps the frame. Keep all critical content out of these bands.
 | Aspect / placement | Top keep-clear | Bottom keep-clear | Sides |
 |---|---|---|---|
 | 9:16 Reels/Stories (Meta) | ~14% (~270px on 1920h) | ~20–35% (~384–672px) | ~6% (~65px) |
 | 9:16 TikTok In-Feed | ~130px | ~484px (right-side icon rail + caption) | right ~140px |
 | 4:5 Meta Feed | ~5% | ~10% (keep CTA above it) | ~5% |
 | 1:1 / 16:9 | center 90% safe | center 90% safe | center 90% safe |
 The 9:16 bottom third is the single most violated zone — captions, the CTA button, and the handle stack there. The `bottomSafe` padding in `batch-ad-pipeline.md` reserves it.
 ## Duration & format limits
 | Placement | Sweet spot | Hard limits | Format |
 |---|---|---|---|
 | Meta Feed (4:5/1:1) | 6–15s | up to 240 min, ≤4GB | MP4/MOV, H.264 |
 | Meta Reels/Stories (9:16) | 8–15s | Reels ≤90s, Stories ≤2 min | MP4/MOV |
 | TikTok In-Feed (9:16) | 15–30s (TikTok cites 21–34s) | up to 10 min (non-Spark) | MP4/MOV, ≥540×960 |
 | YouTube in-stream (16:9) | 15–30s | skippable after 5s | MP4 |
 Practical default for performance video: **15–30s, with the hook trigger landing before 2s.** Most ad-account data shows attention falls off a cliff after ~15s, so front-load everything that matters.
 ## Hook → CTA message-match worksheet
 Fill one row per ad *before* writing the CSV. The CTA must pay off the exact promise the hook made; if it doesn't, the click is wasted and the test is noisy. Each completed row becomes one record in `variants.csv`.
 | Hook (0–3s promise) | Hook type | Single benefit | Proof point | CTA (pays off the hook) |
 |---|---|---|---|---|
 | "Spending 2 hrs/day on reports?" | problem-first | one-click report build | "save 11 hrs/week" | "Save 2 hours — try free" |
 | "Your reporting tool is lying to you." | pattern interrupt | shows real numbers instantly | "save 11 hrs/week" | "See the real numbers — free" |
 | "Nobody talks about this reporting trick." | curiosity gap | one click builds it | "save 11 hrs/week" | "Try the trick — free" |
 | "Cut report time in half this week." | immediate benefit | auto-built reports | "save 11 hrs/week" | "Cut it in half — free" |
 | "If you build reports by hand, stop." | direct callout | automate it now | "save 11 hrs/week" | "Automate yours — free" |
 Bad pairing to avoid: a "stop wasting 2 hours" hook ending on a generic "Shop now" — the promise and the payoff don't connect, so even a great hook converts poorly.
 ## Testing order (which field to vary first)
 Run tests in funnel order; isolate one field per wave using the matrix generator:
 1. **Hook** — biggest lever on CPA; test 5–8 hooks against one identical body.
 2. **Offer / benefit** — take the winning hook, vary the core promise.
 3. **CTA** — winning hook + offer, vary the closing action.
 4. **Format / aspect** — same winning ad, compare placements.
 Budget guide many teams use: ~60% to proven winners, ~30% to variations of winners, ~10% to fresh concepts. Run each test long enough for a real sample (commonly several hundred impressions per variant, ~1 week to smooth daily swings) before declaring a winner.
 ## Pre-flight checklist per variant
 - Hook trigger lands before 2s and fits the frame (≤~60 chars in 9:16).
 - CTA visibly pays off the hook's specific promise.
 - Critical content inside the center square and outside every keep-clear band.
 - Duration within the placement's sweet spot.
 - Exported in every aspect the campaign's placements require.
 - Exactly one field differs from the other variants in this test wave.
@@ -1,137 +0,0 @@
 ---
 name: launch-video
 description: This skill should be used when the user asks to "make a launch video", "create a product hype/sizzle film", "build a teaser/trailer for a drop", "do a reveal montage with an end-card CTA", "cut a hero video to the beat", or "export a launch film in 16:9/9:16/1:1". Covers the hook→tease→reveal→feature-montage→end-card arc and multi-aspect export.
 version: 0.1.0
 ---
 # Launch Video
 Make a premium, high-energy 15–60s launch film that stops the scroll and converts. Build the classic arc — hook → tease → reveal → feature montage → end-card CTA — cut to the music, and export for every platform aspect.
 ## When to use
 - Product drops, feature announcements, trailers, hype reels (15–60s).
 - "Sizzle" hero videos for a landing page.
 - Beat-synced reveals where the brand moment lands on the drop.
 ## The arc
 | Beat | Job | Typical share (of 30s) |
 |---|---|---|
 | Hook (0–3s) | A striking frame/motion that stops the scroll | 0–3s |
 | Tease | Hint the product; build curiosity, beat-synced | 3–9s |
 | Reveal | Product/logo hits on the music drop | 9–13s |
 | Feature montage | Fast, kinetic, one benefit per shot | 13–25s |
 | End card | Logo + tagline + CTA, clean hold | 25–30s |
 Scale the same proportions for 15s (tighter) or 60s (longer montage, never a longer hook).
 ## Two non-negotiable rules
 1. **Sound design leads picture.** Lock the track first, mark the beats and the drop, then cut visuals to those marks. Never score a finished cut — the reveal lands on the drop, not near it.
 2. **Quality over quantity.** A few flawless shots beat many mediocre ones. Cut any shot that isn't premium.
 ## Beat-synced reveal on the drop
 Find the drop's exact timecode in the audio, then time the reveal so the brand mark snaps to full at that frame, with a micro-overshoot for impact.
 ```js
 import gsap from "gsap";
 const DROP = 9.0; // seconds, measured from the track
 const tl = gsap.timeline({ paused: true });
 // tease builds tension up to the drop
 tl.to(".tease", { opacity: 1, scale: 1.05, duration: DROP, ease: "power1.in" })
 // reveal SNAPS on the drop frame
  .set(".logo", { opacity: 1 }, DROP)
  .fromTo(".logo", { scale: 1.18 }, { scale: 1.0, duration: .18, ease: "power3.out" }, DROP)
  .fromTo(".flash", { opacity: .9 }, { opacity: 0, duration: .25 }, DROP); // white hit
 audio.addEventListener("play", () => tl.play());
 ```
 Cut on transients, not on a fixed grid: the hardest cuts go on kick/snare hits; ramp speed (time-remap) into the drop, then hard-cut out of it.
 ## Kinetic feature montage
 One benefit per shot, ~0.6–1.0s each, hard cuts on the beat. Each shot = a bold keyword + a single supporting visual.
 ```css
 .feature { opacity:0; }
 .feature.in .kw   { animation: pop .35s cubic-bezier(.22,1,.36,1) forwards; }
 .feature.in .word { display:inline-block; }
@keyframes pop { from { opacity:0; transform:translateY(24px) } to { opacity:1; transform:none } }
 ```
 ```js
 // drive shots off beat marks measured from the track
 const beats = [13.0, 13.8, 14.6, 15.4, 16.2]; // seconds
 beats.forEach((t, i) => scheduleAt(t, () => showFeature(i)));
 ```
 Keep one motion language across the montage (same enter curve, same exit) so speed reads as confidence, not chaos.
 ## End card
 Hold the logo + tagline + CTA still and clean for ≥2s. No busy motion competing with the CTA; let the brand settle. Land the brand moment on the strongest remaining beat, then a crisp button/URL.
 ## Multi-aspect export
 Compose with a center safe zone so one master crops cleanly to all aspects.
 | Aspect | Use | Resolution | Safe zone |
 |---|---|---|---|
 | 16:9 | YouTube, landing hero, X | 1920×1080 | keep key content within center 90% |
 | 9:16 | Reels, TikTok, Shorts, Stories | 1080×1920 | text within center 80% h; avoid top 12% / bottom 18% (UI) |
 | 1:1 | Feed posts | 1080×1080 | center square of the 16:9 frame |
 Design the hero/logo/CTA inside the **1:1 center square** so it survives every crop. Render the 16:9 master, then reframe (don't just letterbox) 9:16 and 1:1 from the same project.
 ## Output checklist
 - Track locked first; reveal lands exactly on the drop.
 - Hook earns the first 3 seconds.
 - Montage: one benefit per shot, hard cuts on beats, consistent motion language.
 - End card holds ≥2s with one clear CTA.
 - 16:9 + 9:16 + 1:1 exports, key content in the center safe zone.
 - Premium-only shots — nothing mediocre survives the cut.
 ## Deliver & verify (rendered stills → MP4)
 > **Packaged helper** (`scripts/`): tile your stills with `scripts/contact-sheet.sh sheet.png f-hook.png f-mid.png f-end.png`, then assert the encode with `scripts/probe-mp4.sh out.mp4 [WxH] [fps]`. See `scripts/README.md`.
 If the launch film is built/rendered in Remotion, treat it as the heavy tier: register a `<Composition>` (+ zod `schema` + `defaultProps`), drive every value off `useCurrentFrame()` — no GSAP timelines / audio-event callbacks / `Date.now()` / `Math.random()` at render time (bake beat + drop timecodes into props). Deliverable = the rendered `out/*.mp4` per aspect (plus the project).
 **Verify loop — render stills → inspect → encode.** Sample the arc, with the **shipped** props (real headline/tagline/CTA), not just `defaultProps`.
 ```bash
 # Frame-exact stills across hook → reveal → end-card
 npx remotion still Launch out/f-hook.png   --frame=24  --props=props/launch.json   # hook earns first 3s
 npx remotion still Launch out/f-reveal.png --frame=270 --props=props/launch.json   # logo SNAPS on the drop frame
 npx remotion still Launch out/f-end.png    --frame=899 --props=props/launch.json   # end card: logo+tagline+CTA, clean hold
 # inspect each: fidelity (product/logo, tagline, CTA text exact, brand colors right, reveal at the drop frame)
 #   AND artifacts (text overflow, off-canvas, CTA in the 9:16 bottom third, missing font, flash frame stuck on)
 ```
 **Multi-aspect — verify the master in EACH aspect before encoding all of them.** A reframe bug (hero pushed off the safe zone) repeats across every aspect; catch it once.
 ```bash
 for ar in 16x9 9x16 1x1; do                                 # check reveal + end card per aspect
  npx remotion still Launch "out/reveal-${ar}.png" --frame=270 \
    --props=props/launch.json --props-merge="{\"aspect\":\"${ar}\"}"
 done
 # stills clean in all aspects? → then encode each:
 for ar in 16x9 9x16 1x1; do
  npx remotion render Launch "out/launch-${ar}.mp4" --props=props/launch.json --props-merge="{\"aspect\":\"${ar}\"}"
 done
 npx remotion render Launch out/demo.gif --codec=gif --props=props/launch.json   # README demo
 ```
 **Before you finish:**
 1. `npx remotion still` renders cleanly at hook / reveal / end card — no errors, no missing fonts/assets.
 2. Logo/tagline/CTA text exact, brand colors right; reveal lands on the drop frame; key content inside the center safe zone, clear of the 9:16 bottom third.
 3. Frame-driven only — no GSAP/audio-callback timing, `Date.now()`, or `Math.random()` (beats/drop baked into props).
 4. Shipped props render correctly (not just `defaultProps`); the end card holds ≥2s with one CTA.
 5. Master verified in 16:9 + 9:16 + 1:1 before encoding; MP4s play; (optional) GIF for the README.
 ## Reference files
 - `references/launch-structure.md` — a full shot-by-shot template with timecodes for a 30s film (scalable to 15–60s), beat-synced reveal timing, kinetic-montage shot recipes, the sound-design-leads-picture workflow, and detailed multi-aspect safe-area maps for 16:9 / 9:16 / 1:1.
@@ -1,142 +0,0 @@
 # Launch Video Structure — Cookbook
 Shot-by-shot template, beat-sync timing, montage recipes, and multi-aspect maps for a premium 15–60s launch film.
 ## Sound-design-leads-picture workflow
 1. **Pick/commission the track first.** The track defines the film's length and energy.
 2. **Mark the structure in the audio**: intro, build, the **drop** (the single biggest moment), and the kick/snare transients through the montage. Note exact timecodes.
 3. **Map beats → cuts.** Hard cuts land on transients; the reveal lands on the drop frame.
 4. **Build picture to the marks.** Never finish a cut and then look for music. Score-first guarantees the reveal hits.
 5. **Mix last**: add risers into the drop, an impact/whoosh on the reveal, and let the end-card breathe (drop music low, or a clean tag).
 ## Shot-by-shot template — 30s film (scalable)
 ```
 00:00–00:03  HOOK
  Visual : single striking frame — extreme close detail / impossible motion
  Motion : slow push-in or a fast whip that resolves; NO logo yet
  Audio  : sub-bass swell begins, sparse
  Cut    : on first audible beat at ~03s
 00:03–00:09  TEASE
  Visual : fragments/silhouettes of the product, never the full thing
  Motion : light leaks, partial reveals, parallax; tension rising
  Audio  : riser building toward the drop
  Cut    : quick cuts accelerating into the drop
 00:09–00:13  REVEAL  ← THE DROP at 00:09
  Visual : product/logo SNAPS to full on the drop frame
  Motion : scale 1.18→1.0 over 0.18s + white flash-out; settle clean
  Audio  : drop hits + impact stinger on the exact frame
  Hold   : let it sit ~1.2s before moving
 00:13–00:25  FEATURE MONTAGE
  Visual : one benefit per shot, ~0.6–1.0s each (≈8–10 shots)
  Motion : bold keyword pop + one supporting visual; consistent enter/exit
  Audio  : hard cuts on kick/snare; energy sustained
  Pattern: KEYWORD + UI/product moment, repeat with rhythm
 00:25–00:30  END CARD
  Visual : logo + tagline + CTA, centered, still
  Motion : minimal — a soft settle, then hold ≥2s
  Audio  : music resolves; clean sonic tag
  CTA    : one action (URL or button), nothing competing
 ```
 ### Re-proportioning
 - **15s**: Hook 0–2s · Tease 2–5s · Reveal 5–7s · Montage 7–12s (4–5 shots) · End card 12–15s.
 - **60s**: same hook (still ≤3s) · longer tease and a richer montage (14–16 shots, grouped in waves) · end card ≥3s.
 The hook never gets longer — only the montage scales.
 ## Beat-synced reveal (timing detail)
 ```js
 import gsap from "gsap";
 const DROP = 9.0;                 // measured drop timecode (s)
 const tl = gsap.timeline({ paused: true });
 // build tension toward the drop
 tl.fromTo(".tease", { scale: 1.0, opacity: .6 },
                    { scale: 1.08, opacity: 1, duration: DROP, ease: "power1.in" });
 // REVEAL exactly on the drop
 tl.set(".logo",  { opacity: 1 }, DROP);
 tl.fromTo(".logo",  { scale: 1.18 }, { scale: 1.0, duration: .18, ease: "power3.out" }, DROP);
 tl.fromTo(".flash", { opacity: .9 },  { opacity: 0, duration: .25, ease: "power2.out" }, DROP);
 tl.to(".tease", { opacity: 0, duration: .12 }, DROP);
 // hold the reveal
 tl.to(".logo", { scale: 1.0, duration: 1.2 }); // dwell
 audio.addEventListener("play", () => tl.play());
 ```
 Speed-ramp into the drop (Remotion time-remap idea): play the last ~0.5s of the tease faster, then hard-cut to the reveal on the drop frame for a snap.
 ## Kinetic montage shot recipe
 Each shot: a bold keyword + one product/UI visual, ~0.6–1.0s, hard cut on the beat.
 ```css
 .feature { position:absolute; inset:0; opacity:0; }
 .feature.in { opacity:1; }
 .feature .word { display:inline-block; opacity:0; transform:translateY(26px); }
 .feature.in .word { animation: pop .32s cubic-bezier(.22,1,.36,1) forwards; }
 .feature.in .word:nth-child(2){ animation-delay:.04s }
 .feature.in .word:nth-child(3){ animation-delay:.08s }
@keyframes pop { to { opacity:1; transform:none; } }
 ```
 ```js
 // schedule shots on measured beat marks
 const beats = [13.0, 13.8, 14.6, 15.4, 16.2, 17.0, 17.8, 18.6]; // seconds
 let i = 0;
 function loop(now){
  const t = (now - start) / 1000;
  if (i < beats.length && t >= beats[i]) { showFeature(i); i++; }
  if (i < beats.length) requestAnimationFrame(loop);
 }
 ```
 Montage discipline: keep ONE enter curve and ONE exit across all shots; vary content, never grammar. Alternate visual rhythm (full-frame ↔ detail) but keep cut cadence locked to the beat.
 ## Multi-aspect safe-area maps
 Compose so a single master reframes to every aspect. Place hero/logo/CTA inside the smallest common safe zone — the 1:1 center square.
 ```
 16:9  1920×1080
  Safe: center 90% (96px H / 54px V margins)
  Use : YouTube, landing hero, X timeline
 9:16  1080×1920
  Safe: text within center 80% height
  Avoid: top 12% (status/clock area), bottom 18% (caption + UI buttons)
  Use : Reels, TikTok, Shorts, Stories
 1:1   1080×1080
  Safe: this square == the design anchor; put logo/CTA here
  Use : feed posts
 ```
 Export workflow:
 1. Build the 16:9 master with all key content inside the 1:1 center square.
 2. For 9:16, **reframe** (pan/scale per shot to keep the subject centered) — do not letterbox.
 3. For 1:1, crop the center square; verify text isn't clipped.
 4. Re-check captions/CTA against each platform's UI overlay zones before export.
 ## Final QA checklist
 - [ ] Reveal lands on the exact drop frame, with overshoot + flash.
 - [ ] Hook holds attention in the first 3s without showing the product.
 - [ ] Every montage shot is one benefit, premium quality, on a beat.
 - [ ] End card holds ≥2s, single CTA, nothing competing.
 - [ ] 16:9 / 9:16 / 1:1 each pass safe-area and caption checks.
 - [ ] Audio mixed: riser → impact on drop → clean tag on end card.
 ---
 Nail the reveal beat and the sound-leads-picture cut, and a launch film lands. Built by **[iart.ai](https://iart.ai/?utm_source=github&utm_medium=readme&utm_campaign=ad-video-skills&utm_content=skill_footer&utm_term=launch-video)** — the AI motion agent for editable, on-brand motion graphics.
@@ -1,196 +0,0 @@
 ---
 name: testimonial-video
 description: This skill should be used when the user asks to "make an animated testimonial", "create a quote video", "build a review video", "turn a customer testimonial into a graphic", "animate a customer review", "make a social proof video", "add an animated star rating", or "turn a list of reviews into videos". Covers quote typography, staggered line reveals, kinetic emphasis on key phrases, animated star ratings, the author block (name/role/avatar/company), and templating one card across many quotes.
 version: 0.1.0
 ---
 # Testimonial Video
 Animate a written quote, review, or testimonial into a short, trustworthy social-proof clip. The craft is typographic, not flashy: a clean quote card where the words land in reading order, the key phrase gets emphasis exactly when it should, the stars fill to the real score, and the author signs off. One card design, driven by props, becomes a video for every review on the list.
 ## When to use
 - Customer review → a 6–12s vertical/feed clip (the headline use case).
 - A pull-quote or testimonial graphic with author name, role, avatar, company.
 - An animated star rating (including half/fractional scores like 4.5).
 - A batch: a list/CSV of reviews → one video each, same template.
 Do **not** reach here for sale/discount promos or generic ad creative — this skill is strictly the quote / social-proof angle.
 ## The one rule: trust before motion
 A testimonial earns its job by feeling *real*, not produced. Restraint reads as credibility; over-animation reads as an ad people skip.
 1. **Reveal in reading order.** The eye must finish a line before the next arrives. Stagger lines (or clauses), never all words at once, and never reorder for visual flair.
 2. **Quote the words exactly.** Do not paraphrase, trim mid-sentence, or invent a rating. If you emphasize a phrase, it must be the customer's phrase.
 3. **The author is the proof.** A quote with no attributable human is just a slogan. Name + role + company (and avatar if real) is non-negotiable.
 ## The anatomy of a quote card
 | Element | Job | Timing |
 |---|---|---|
 | Quote mark motif | Signals "this is a quote" before words read | Frame 0, scales/fades in first |
 | Quote body | The testimonial, revealed line-by-line | Staggered, ~6–10 frames apart |
 | Emphasis span | The one phrase that sells | Lands *after* its line settles |
 | Star rating | The score, filled to the real value | Fills left→right after the quote |
 | Author block | Name / role / company / avatar | Signs off last, slides up |
 Build in that order; reveal in that order. The viewer reads the quote, believes it, sees the score, then learns who said it.
 ## Staggered line reveal
 Split the quote into lines (or clauses), then offset each one's entrance. In Remotion, drive every line off `useCurrentFrame()` so it renders deterministically — never a CSS transition or a JS timer, which desync from the frame renderer.
 ```jsx
 import { useCurrentFrame, spring, useVideoConfig } from "remotion";
 const QuoteBody = ({ lines, startAt = 8, stagger = 6 }) => {
  const frame = useCurrentFrame();
  const { fps } = useVideoConfig();
  return (
    <blockquote>
      {lines.map((line, i) => {
        const enter = spring({ frame: frame - startAt - i * stagger, fps,
          config: { damping: 200 } });           // 0→1, no overshoot
        return (
          <span key={i} style={{ display: "block",
            opacity: enter,
            transform: `translateY(${(1 - enter) * 14}px)` }}>
            {line}
          </span>
        );
      })}
    </blockquote>
  );
 };
 ```
 Keep one motion language: same easing, same 14px rise, every line. Speed reads as confidence only when it's consistent. See `references/quote-card.md` for the full composition.
 ## Kinetic emphasis on the key phrase
 Pick the single phrase that carries the testimonial ("saved us 10 hours a week") and let it arrive *after* its line has settled — a highlight sweep, a weight shift, or a color pop. Emphasis competing with the reveal cancels both.
 ```jsx
 const sweep = interpolate(frame, [hlStart, hlStart + 10], [0, 100],
  { extrapolateLeft: "clamp", extrapolateRight: "clamp" });
 // a highlighter wiping in behind the phrase, left→right
 <mark style={{ background:
  `linear-gradient(90deg, var(--accent) ${sweep}%, transparent ${sweep}%)`,
  color: "inherit", padding: "0 .08em" }}>{phrase}</mark>
 ```
 Emphasize at most one span per quote. Two highlights is no highlight.
 ## Animated star rating (fractional)
 Render the real score, including halves. Use one full row of star glyphs as a mask, then wipe a filled layer to `rating / max` width — this handles 4.5 as cleanly as 5.0, with no half-star sprite.
 ```jsx
 const StarRating = ({ rating, max = 5, fillStart = 30, fillDur = 14 }) => {
  const frame = useCurrentFrame();
  const target = (rating / max) * 100;
  const w = interpolate(frame, [fillStart, fillStart + fillDur], [0, target],
    { extrapolateLeft: "clamp", extrapolateRight: "clamp" });
  return (
    <div className="stars" role="img" aria-label={`${rating} out of ${max} stars`}>
      <div className="track">★★★★★</div>
      <div className="fill" style={{ width: `${w}%` }}>★★★★★</div>
    </div>
  );
 };
 ```
 ```css
 .stars { position: relative; display: inline-block; font-size: 40px; letter-spacing: 4px; }
 .track { color: rgba(255,255,255,.22); }
 .fill  { position: absolute; inset: 0; overflow: hidden; white-space: nowrap;
         color: var(--accent); }   /* clipped by width → fractional fill */
 ```
 Fill left→right after the quote reads. Never animate stars before the words — the score must feel earned, not pre-loaded. Always set `aria-label` to the real value.
 ## The author block
 Sign off last. The block slides up under the stars and holds still — stillness is what makes attribution read as fact.
 ```jsx
 <figcaption className="author">
  {avatar && <img src={avatar} alt="" className="avatar" />}
  <div>
    <span className="name">{name}</span>
    <span className="role">{role}{company && `, ${company}`}</span>
  </div>
 </figcaption>
 ```
 Use a real avatar or none — a generic placeholder face actively destroys trust. Name is the loudest line in the block; role/company is secondary weight.
 ## Template × reviews — one video per quote
 The payoff: make every field a prop, hardcode nothing, then render once per review.
 ```jsx
 export const Testimonial = ({ quote, highlight, rating, name, role, company, avatar }) => { /* … */ };
 ```
 ```bash
 for f in reviews/*.json; do
  npx remotion render Testimonial "out/$(basename "$f" .json).mp4" --props="$f"
 done
 ```
 Keep colors/fonts/quote-mark/layout in one theme object so 50 testimonials stay brand-locked and only the words and score change. See `references/batch-and-typography.md` for the review schema, CSV→props parsing, quote-typography craft (smart quotes, hanging punctuation, line-splitting, orphans), and the batch script.
 ## Output checklist
 - Lines reveal in reading order; one consistent enter motion, ~6–10 frame stagger.
 - The quote is verbatim; exactly one emphasized span, landing after its line settles.
 - Stars fill to the *real* score (halves supported) after the quote, with a correct `aria-label`.
 - Author block (name + role + company, real avatar or none) signs off last and holds still.
 - Every animated value derives from `useCurrentFrame()` — no CSS/JS timers.
 - Every field is a prop; one template renders every review in the folder.
 ## Deliver & verify (rendered stills → MP4)
 > **Packaged helper** (`scripts/`): tile your stills with `scripts/contact-sheet.sh sheet.png f-hook.png f-mid.png f-end.png`, then assert the encode with `scripts/probe-mp4.sh out.mp4 [WxH] [fps]`. See `scripts/README.md`.
 **Output contract:**
 - A Remotion `Testimonial` registered as `<Composition>` (+ zod `schema` + `defaultProps`), every value frame-driven (no CSS transitions / JS timers / `Date.now()` / `Math.random()`).
 - Deliverable = the rendered `out/*.mp4` per review per aspect (plus the project + review data, so the user re-renders on new quotes).
 **Verify loop — render stills → inspect → encode.** Cheap PNGs first; render with the **shipped** review's props, not just `defaultProps`.
 ```bash
 # Frame-exact stills across the reveal order: quote → stars → author
 npx remotion still Testimonial out/f-quote.png  --frame=30  --props=reviews/r1.json  # lines in reading order, emphasis span
 npx remotion still Testimonial out/f-stars.png  --frame=48  --props=reviews/r1.json  # stars filled to the REAL score (e.g. 4.5)
 npx remotion still Testimonial out/f-author.png --frame=215 --props=reviews/r1.json  # author block signed off, holding still
 # inspect each: fidelity (quote VERBATIM, exactly one emphasized span, star fill matches rating, name/role/company exact)
 #   AND artifacts (text overflow/orphans, off-canvas, author in the 9:16 bottom third, missing font, broken avatar)
 ```
 **Multi-aspect / batch — verify one review in EACH target aspect before batch-rendering the list.** A layout bug (overflowing quote, clipped author) repeats across every review × aspect; catch it once.
 ```bash
 for ar in 9x16 1x1 16x9; do                                  # one representative review, every aspect
  npx remotion still Testimonial "out/r1-${ar}.png" --frame=215 \
    --props=reviews/r1.json --props-merge="{\"aspect\":\"${ar}\"}"
 done
 # stills clean in all aspects? → then batch-render the list:
 for f in reviews/*.json; do id=$(basename "$f" .json); for ar in 9x16 1x1 16x9; do
  npx remotion render Testimonial "out/${id}-${ar}.mp4" --props="$f" --props-merge="{\"aspect\":\"${ar}\"}"
 done; done
 npx remotion render Testimonial out/demo.gif --codec=gif --props=reviews/r1.json   # README demo
 ```
 **Before you finish:**
 1. `npx remotion still` renders cleanly at quote / stars / author — no errors, no missing fonts/avatar.
 2. Quote is verbatim with one emphasized span; stars fill to the real score; name/role/company exact and inside the safe zone (clear of the 9:16 bottom third).
 3. Frame-driven only — no CSS/JS timers, `Date.now()`, or `Math.random()`.
 4. Shipped review's props render correctly (not just `defaultProps`); `aria-label` carries the real rating.
 5. One review verified in 9:16 + 1:1 + 16:9 before the batch; MP4s play; (optional) GIF for the README.
 ## Reference files
 - `references/quote-card.md` — a complete runnable Remotion `Testimonial` composition: staggered line reveal, key-phrase highlight sweep, fractional star fill, author block, theme object, and the timing map for a 9:16 card. Plus a dependency-free HTML/CSS variant.
 - `references/batch-and-typography.md` — the review prop schema, CSV/JSON→props parsing and validation, quote typography craft (smart quotes, hanging quotation marks, line/clause splitting, orphan control, legibility in motion), star-rating accessibility, and the template×reviews batch render script for N videos.
@@ -1,140 +0,0 @@
 # Batch & Typography
 How to feed many reviews into one card design, plus the quote-typography craft that separates a credible testimonial from a sloppy one.
 ## Review prop schema
 One object per testimonial. Keep it flat so a CSV row maps to it directly.
 ```ts
 type Review = {
  quote: string;        // verbatim — never paraphrased or trimmed mid-sentence
  highlight?: string;   // EXACT substring of `quote`; one phrase only
  rating: number;       // 0–max, halves allowed (4.5)
  max?: number;         // default 5
  name: string;         // required — a quote with no name is not social proof
  role?: string;
  company?: string;
  avatar?: string;      // real photo URL, or omit (never a stock face)
 };
 ```
 ### Validation (fail loud, before rendering 50 videos)
 ```ts
 function validate(r: Review): string[] {
  const errs: string[] = [];
  if (!r.quote?.trim()) errs.push("quote is empty");
  if (!r.name?.trim()) errs.push("name is required (no anonymous testimonials)");
  const max = r.max ?? 5;
  if (r.rating < 0 || r.rating > max) errs.push(`rating ${r.rating} out of 0–${max}`);
  if (r.highlight && !r.quote.includes(r.highlight))
    errs.push(`highlight "${r.highlight}" is not a substring of the quote`);
  return errs;
 }
 ```
 The `highlight ⊂ quote` check is the important one: it guarantees you only ever emphasize words the customer actually wrote. If the highlight isn't an exact substring, the emphasis silently disappears in the card — so block the render instead.
 ## CSV → props
 Reviews usually arrive as a spreadsheet export. Map columns to the schema and coerce the rating.
 ```js
 import { parse } from "csv-parse/sync";
 import { readFileSync, writeFileSync, mkdirSync } from "node:fs";
 // CSV columns: quote,highlight,rating,name,role,company,avatar
 const rows = parse(readFileSync("reviews.csv"), { columns: true, trim: true });
 mkdirSync("reviews", { recursive: true });
 rows.forEach((row, i) => {
  const r = {
    quote: row.quote,
    highlight: row.highlight || undefined,
    rating: Number(row.rating),
    name: row.name,
    role: row.role || undefined,
    company: row.company || undefined,
    avatar: row.avatar || undefined,
  };
  const errs = validate(r);
  if (errs.length) { console.error(`Row ${i + 1} (${r.name}):`, errs); process.exit(1); }
  writeFileSync(`reviews/${String(i).padStart(3, "0")}-${slug(r.name)}.json`,
    JSON.stringify(r, null, 2));
 });
 const slug = (s) => s.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-|-$/g, "");
 ```
 ## Batch render — one video per review
 ```bash
 mkdir -p out
 for f in reviews/*.json; do
  name=$(basename "$f" .json)
  npx remotion render Testimonial "out/$name.mp4" --props="$f"
 done
 ```
 Everything visual lives in the single `theme` object in the composition (colors, fonts, quote-mark glyph, spacing). Because only the props change per render, 50 testimonials stay pixel-for-pixel brand-locked while the words and score vary. To re-theme the whole set, edit one object and re-run the loop.
 For multi-aspect (9:16 feed + 1:1 post + 16:9 site), register the composition at each size with a shared center-safe layout and loop over both `reviews/*.json` and the aspect list.
 ## Quote typography craft
 Motion is the easy part; legible, well-set type is what makes the quote trustworthy.
 ### Smart quotes, always
 Straight quotes (`"`) read as un-proofed. Convert to typographic quotes before rendering.
 ```js
 const smarten = (s) => s
  .replace(/(^|[\s([{])"/g, "$1“").replace(/"/g, "”")   // “ ”
  .replace(/(^|[\s([{])'/g, "$1‘").replace(/'/g, "’")   // ‘ ’
  .replace(/--/g, "—");                                       // em dash
 ```
 ### Hanging quotation marks
 A leading quote glyph indents the first line and breaks the flush-left edge. Hang it into the margin so the text block stays optically aligned — the gold standard for set quotes.
 ```css
 blockquote { hanging-punctuation: first; }            /* Safari/WebKit */
 /* Cross-engine fallback: pull the opening mark left by its own width */
 blockquote::before { content: "\201C"; margin-left: -0.55em; color: var(--muted); }
 ```
 When the card uses a large decorative quote mark (as in `quote-card.md`), drop the inline opening glyph so the mark isn't doubled.
 ### Splitting into lines / clauses
 Reveal in reading order means splitting on meaning, not on a fixed character count. Prefer clause boundaries (`. ! ? , ; —`), then pack to a comfortable measure (~32–40 chars at card size). The `toLines` helper in `quote-card.md` does exactly this. Never split a short quote into one-word-per-line — that reads as a ransom note, not a testimonial.
 ### Orphans & widows
 A single dangling word on the last line looks broken. Bind the last two words with a non-breaking space so they wrap together.
 ```js
 const noOrphan = (s) => s.replace(/ ([^ ]+)\s*$/, " $1");
 ```
 ### Legibility in motion
 - Use a font that stays readable while moving — avoid thin weights and tight tracking; a serif or a humanist sans at 500–600 weight holds up best on a rising line.
 - Animate **opacity + a small translate** only. Blur, rotation, and scale on body text hurt readability mid-reveal.
 - Contrast: body text ≥ 4.5:1 against the card; the highlight accent should pass against both the card and the ink it sits behind.
 - One emphasized span per quote. A second highlight halves the weight of the first.
 ## Star-rating accessibility & correctness
 - The visual stars are decorative; expose the value to assistive tech with `role="img"` and `aria-label="4.5 out of 5 stars"` on the container (already in `quote-card.md`).
 - For fractional scores, the width-clip method (`rating / max * 100%`) is exact for any decimal — no rounding to the nearest half-star, so 4.3 renders as 4.3.
 - Never animate the stars before the quote is read, and never show more filled stars than the data supports. A faked rating is the fastest way to lose the trust the testimonial was meant to build.
 - If a review has no rating, omit the star row entirely rather than defaulting to 5 — an invented score is worse than none.
 ## Pacing notes
 - Hold the finished card still for the back half of the clip (≥3s) so it's screenshot-able and re-shareable — shares are the whole point of social proof.
 - For a sequence of testimonials in one video, give each card a full read (~3–4s) and hard-cut between them; cross-dissolving quotes makes both unreadable.
@@ -1,241 +0,0 @@
 # Quote Card — complete runnable component
 A full testimonial card you can drop into a Remotion project, plus a zero-dependency HTML/CSS variant for non-Remotion pipelines. Both reveal in reading order, emphasize one phrase, fill stars to the real score, and sign off with the author.
 ## Timing map (9:16, 30fps, ~10s)
 | Frame | Beat |
 |---|---|
 | 0–8 | Quote mark scales in; card settles |
 | 8–48 | Quote body, line-by-line (8 + i×6) |
 | 44–54 | Key-phrase highlight sweep (after its line settles) |
 | 56–70 | Stars fill left→right to `rating/max` |
 | 72–88 | Author block slides up, holds |
 | 88–300 | Everything holds still (screenshot-able) |
 The card holds motionless for the back half — stillness is what makes a testimonial read as a fact rather than an ad.
 ## Remotion composition
 ```tsx
 import {
  AbsoluteFill, useCurrentFrame, useVideoConfig, spring, interpolate, Img,
 } from "remotion";
 export type Review = {
  quote: string;          // verbatim testimonial
  highlight?: string;     // exact substring of `quote` to emphasize
  rating: number;         // e.g. 4.5
  max?: number;           // default 5
  name: string;
  role?: string;
  company?: string;
  avatar?: string;        // real photo URL, or omit
 };
 const theme = {
  bg: "#0E1116",
  ink: "#F5F7FA",
  muted: "rgba(245,247,250,.55)",
  accent: "#FFC24B",
  font: "Inter, system-ui, sans-serif",
  quoteFont: "'Newsreader', Georgia, serif",
  quoteMark: "“",     // left double quotation mark
 };
 // Split a quote into clauses for reading-order reveal.
 // Prefer sentence/clause boundaries; fall back to ~32-char wraps.
 const toLines = (q: string): string[] => {
  const parts = q.match(/[^.!?,;]+[.!?,;]?\s*/g) ?? [q];
  const lines: string[] = [];
  let buf = "";
  for (const p of parts) {
    if ((buf + p).length > 34 && buf) { lines.push(buf.trim()); buf = p; }
    else buf += p;
  }
  if (buf.trim()) lines.push(buf.trim());
  return lines;
 };
 // Render a line, wrapping the highlight substring in <mark> if present.
 const Line: React.FC<{ text: string; highlight?: string; sweep: number }> = ({
  text, highlight, sweep,
 }) => {
  if (!highlight || !text.includes(highlight)) return <>{text}</>;
  const [before, after] = text.split(highlight);
  return (
    <>
      {before}
      <mark
        style={{
          background: `linear-gradient(90deg, ${theme.accent} ${sweep}%, transparent ${sweep}%)`,
          color: "inherit",
          padding: "0 .08em",
          borderRadius: 3,
          WebkitBoxDecorationBreak: "clone",
        }}
      >
        {highlight}
      </mark>
      {after}
    </>
  );
 };
 export const Testimonial: React.FC<Review> = ({
  quote, highlight, rating, max = 5, name, role, company, avatar,
 }) => {
  const frame = useCurrentFrame();
  const { fps } = useVideoConfig();
  const lines = toLines(quote);
  // Quote mark
  const markIn = spring({ frame, fps, config: { damping: 14 } });
  // Highlight sweep, after the body has revealed
  const hlStart = 8 + lines.length * 6 + 4;
  const sweep = interpolate(frame, [hlStart, hlStart + 10], [0, 100], {
    extrapolateLeft: "clamp", extrapolateRight: "clamp",
  });
  // Stars fill, after the quote
  const fillStart = hlStart + 14;
  const target = (rating / max) * 100;
  const starW = interpolate(frame, [fillStart, fillStart + 14], [0, target], {
    extrapolateLeft: "clamp", extrapolateRight: "clamp",
  });
  // Author block slides up last
  const authStart = fillStart + 16;
  const auth = spring({ frame: frame - authStart, fps, config: { damping: 200 } });
  return (
    <AbsoluteFill style={{ background: theme.bg, fontFamily: theme.font,
      padding: 120, justifyContent: "center", color: theme.ink }}>
      <div
        aria-hidden
        style={{ fontFamily: theme.quoteFont, fontSize: 200, lineHeight: .6,
          color: theme.accent, opacity: markIn, transform: `scale(${markIn})`,
          transformOrigin: "left top" }}
      >
        {theme.quoteMark}
      </div>
      <blockquote style={{ fontFamily: theme.quoteFont, fontSize: 64,
        lineHeight: 1.25, margin: "24px 0 48px", fontWeight: 500 }}>
        {lines.map((line, i) => {
          const enter = spring({ frame: frame - (8 + i * 6), fps,
            config: { damping: 200 } });
          return (
            <span key={i} style={{ display: "block", opacity: enter,
              transform: `translateY(${(1 - enter) * 14}px)` }}>
              <Line text={line} highlight={highlight} sweep={sweep} />
            </span>
          );
        })}
      </blockquote>
      {/* Stars */}
      <div role="img" aria-label={`${rating} out of ${max} stars`}
        style={{ position: "relative", display: "inline-block", fontSize: 52,
          letterSpacing: 6, marginBottom: 56 }}>
        <div style={{ color: "rgba(245,247,250,.20)" }}>{"★".repeat(max)}</div>
        <div style={{ position: "absolute", inset: 0, overflow: "hidden",
          whiteSpace: "nowrap", width: `${starW}%`, color: theme.accent }}>
          {"★".repeat(max)}
        </div>
      </div>
      {/* Author */}
      <figcaption style={{ display: "flex", alignItems: "center", gap: 20,
        opacity: auth, transform: `translateY(${(1 - auth) * 16}px)` }}>
        {avatar && (
          <Img src={avatar} alt=""
            style={{ width: 88, height: 88, borderRadius: "50%", objectFit: "cover" }} />
        )}
        <div style={{ display: "flex", flexDirection: "column" }}>
          <span style={{ fontSize: 38, fontWeight: 700 }}>{name}</span>
          <span style={{ fontSize: 30, color: theme.muted }}>
            {[role, company].filter(Boolean).join(", ")}
          </span>
        </div>
      </figcaption>
    </AbsoluteFill>
  );
 };
 ```
 Register it with a duration long enough for the back-half hold:
 ```tsx
 <Composition id="Testimonial" component={Testimonial}
  durationInFrames={300} fps={30} width={1080} height={1920}
  defaultProps={{
    quote: "We shipped in days, not months. Onboarding saved us 10 hours a week.",
    highlight: "saved us 10 hours a week",
    rating: 4.5, name: "Dana Reyes", role: "Head of Ops", company: "Northwind",
  }} />
 ```
 ## Why these choices
 - **`spring` with `damping: 200`** gives a clean rise with no overshoot — testimonials should feel composed, not bouncy. The quote mark uses lighter damping (14) for a touch of life on the one decorative element.
 - **Highlight as a gradient `<mark>`** wipes a real highlighter behind the phrase without a second element or a mask, and survives line wraps via `box-decoration-break: clone`.
 - **Stars as two stacked glyph rows clipped by width** render any fractional score (4.5, 4.3) with no half-star asset, and the filled row inherits the brand accent.
 - **Author after stars** matches reading psychology: claim → proof → who. Reordering weakens trust.
 ## Zero-dependency HTML/CSS variant
 For pipelines that capture a headless-browser timeline (or just a web embed). Animations are CSS `@keyframes` with per-line `animation-delay`; if rendering frame-exact video, prefer the Remotion version above so frames stay deterministic.
 ```html
 <figure class="card" style="--accent:#FFC24B; --rating:4.5; --max:5">
  <div class="qmark" aria-hidden="true">&ldquo;</div>
  <blockquote>
    <span style="--i:0">We shipped in days, not months.</span>
    <span style="--i:1">Onboarding <mark>saved us 10 hours a week.</mark></span>
  </blockquote>
  <div class="stars" role="img" aria-label="4.5 out of 5 stars">
    <div class="track">★★★★★</div><div class="fill">★★★★★</div>
  </div>
  <figcaption>
    <img src="dana.jpg" alt="" class="avatar">
    <div><b>Dana Reyes</b><small>Head of Ops, Northwind</small></div>
  </figcaption>
 </figure>
 ```
 ```css
 .card { --bg:#0E1116; --ink:#F5F7FA; background:var(--bg); color:var(--ink);
  font-family:Inter,system-ui,sans-serif; padding:120px; max-width:1080px; }
 .qmark { font:200px/.6 'Newsreader',Georgia,serif; color:var(--accent);
  animation:pop .4s cubic-bezier(.22,1,.36,1) both; }
 blockquote { font:500 64px/1.25 'Newsreader',Georgia,serif; margin:24px 0 48px; }
 blockquote span { display:block; opacity:0;
  animation:rise .5s cubic-bezier(.22,1,.36,1) both;
  animation-delay:calc(.27s + var(--i) * .2s); }
 mark { background:linear-gradient(90deg,var(--accent) 100%,transparent 100%);
  color:inherit; padding:0 .08em; border-radius:3px;
  -webkit-box-decoration-break:clone; box-decoration-break:clone;
  animation:sweep .35s ease-out both; animation-delay:1s; }
@keyframes sweep { from { background-size:0 100% } to { background-size:100% 100% } }
 .stars { position:relative; display:inline-block; font-size:52px; letter-spacing:6px; }
 .track { color:rgba(245,247,250,.20); }
 .fill { position:absolute; inset:0; overflow:hidden; white-space:nowrap;
  color:var(--accent); width:0; animation:fill .5s ease-out 1.3s forwards; }
@keyframes fill { to { width:calc(var(--rating) / var(--max) * 100%) } }
 figcaption { display:flex; align-items:center; gap:20px; margin-top:56px;
  opacity:0; animation:rise .5s ease-out 1.8s both; }
 .avatar { width:88px; height:88px; border-radius:50%; object-fit:cover; }
 figcaption b { font-size:38px; } figcaption small { display:block; font-size:30px;
  color:rgba(245,247,250,.55); }
@keyframes pop  { from { opacity:0; transform:scale(.7) } to { opacity:1; transform:none } }
@keyframes rise { from { opacity:0; transform:translateY(14px) } to { opacity:1; transform:none } }
 ```
 The CSS `mark` sweep uses `background-size` (animatable) rather than a gradient color-stop, since color-stop percentages don't animate in CSS. The star `--rating/--max` math is the same fractional fill as the Remotion version.
 ---
 ## Built by the team behind iart.ai
 This skill is part of an open motion-graphics collection from iart.ai — the AI motion agent that turns data, scripts, and designs into editable motion graphics (Remotion → MP4). If you'd rather not hand-build this, iart.ai can turn a list of reviews/quotes into testimonial videos, swap text, and re-export. → [iart.ai](https://iart.ai/?utm_source=github&utm_medium=reference&utm_campaign=ad-video-skills&utm_content=ref_footer&utm_term=testimonial-video)
@@ -1 +0,0 @@
 ../../.agents/skills/ad-creative-video
@@ -1 +0,0 @@
 ../../.agents/skills/launch-video
@@ -1 +0,0 @@
 ../../.agents/skills/testimonial-video
@@ -0,0 +1,83 @@
 # PRD: "Firefly-style" Cloud Control Platform (working name: CloudLens)
 Status: DRAFT 1. Written from general knowledge of Firefly.ai (their website was
 blocked from my work environment). Please compare against firefly.ai before we build.
 ## 1. What is this, in plain words?
 Companies rent computers, storage and databases from Amazon (AWS), Google (GCP) and
 Microsoft (Azure). Over time nobody knows exactly what they have, what it costs, or
 who changed what. Firefly is like a **smart inventory + safety camera for the cloud**:
 1. It **looks** at your cloud account and lists everything you own.
 2. It spots things that were made **by hand** (nobody wrote them down as code).
 3. It **writes the code** (Terraform) so those things are tracked.
 4. It **watches for changes** ("drift") and warns you.
 5. It checks **rules** (e.g. "no public storage buckets") and **cost**.
 ## 2. Core features (the full dream list)
 | # | Feature | 4th-grade explanation |
 |---|---------|-----------------------|
 | 1 | Cloud connection | Give the app a read-only key so it can look at your cloud. |
 | 2 | Asset inventory | One big searchable list of everything you own. |
 | 3 | Unmanaged-asset finder | Highlights things made by hand, not by code. |
 | 4 | Code generator (IaC) | Turns a hand-made thing into Terraform code. |
 | 5 | Drift detection | "This changed and nobody told us!" alerts. |
 | 6 | Policy / rules engine | Checks against safety and cost rules. |
 | 7 | Cost view | Shows what each thing costs per month. |
 | 8 | Alerts | Email / Slack messages when something is wrong. |
 | 9 | Backup & restore | Saves a copy of your setup so you can rebuild it. |
 | 10 | AI helper | Ask "which servers are public?" in normal words. |
 | 11 | Teams, logins, audit log | Who can see what, and who did what. |
 ## 3. Phased roadmap (small steps, test each one)
 ### Phase 0 - Quick win TODAY (about 1 session)
 Goal: something you can open and click.
 - A web dashboard with a **sample** cloud inventory (fake data, clearly labeled).
 - Search, filter by cloud / type / region, and a red badge for "unmanaged".
 - Summary cards: total assets, unmanaged count, monthly cost.
 - No logins, no real cloud yet. Zero risk.
 - **Tests:** page loads with no errors -> search/filter buttons work.
 ### Phase 1 - Real data (read-only)
 - Connect ONE real AWS account with a read-only key.
 - Replace sample data with real resources (EC2, S3, RDS first).
 - **You must:** create a read-only AWS user and give me its keys (I will walk you through it).
 - **Tests:** real API returns data -> keys never shown in the browser or saved in code.
 ### Phase 2 - Drift + rules
 - Save a snapshot every day; show "what changed".
 - First 10 rules (public bucket, no encryption, untagged, etc.).
 - Tests: change something on purpose -> app notices.
 ### Phase 3 - Code generation
 - Click an unmanaged item -> get Terraform code. Start with S3 + EC2 only.
 - Tests: generated code passes `terraform validate`.
 ### Phase 4 - Alerts, cost, teams
 - Email/Slack alerts, real cost numbers, logins and roles, audit log.
 - Security review: passwords, permissions, secret storage, rate limits.
 ### Phase 5 - Multi-cloud + AI
 - Add GCP and Azure. Add the "ask in plain English" helper (Claude API).
 - Backup & restore of cloud setups.
 ## 4. Suggested tech (so you know, no need to learn it)
 - Frontend: Next.js + React (what Firefly-style dashboards use).
 - Backend: Node/TypeScript API, PostgreSQL database.
 - Cloud reading: official AWS / GCP / Azure SDKs.
 - Secrets: a vault / environment variables, never in code.
 ## 5. Security rules (non-negotiable)
 - Read-only cloud keys first. Write access only in a later phase, with approval.
 - Keys stored encrypted on the server only.
 - Every phase ends with a security check before the next starts.
 ## 6. Out of scope (for now)
 Copying Firefly's logo, name, text or design. We build our **own** product with the
 same *kind* of features. Their exact code/designs are theirs.
 ## 7. Open questions for you
 1. Who is it for: just you, your company, or customers you will sell to?
 2. Which cloud do you use most: AWS, Google, or Azure?
 3. Do you have a cloud account we can use for testing? (If not, Phase 0 stays on fake data.)
@@ -1,23 +0,0 @@
 {
  "version": 1,
  "skills": {
    "ad-creative-video": {
      "source": "iart-ai/ad-video-skills",
      "sourceType": "github",
      "skillPath": "skills/ad-creative-video/SKILL.md",
      "computedHash": "94c33e3435abd802773c4b7e6058f3d47456bc56354d87b1e6fc2b689757735d"
    },
    "launch-video": {
      "source": "iart-ai/ad-video-skills",
      "sourceType": "github",
      "skillPath": "skills/launch-video/SKILL.md",
      "computedHash": "a55b82d2ca76de5bb80a1217594b11d1bf1ae4485b487bcad7ab256e2f116790"
    },
    "testimonial-video": {
      "source": "iart-ai/ad-video-skills",
      "sourceType": "github",
      "skillPath": "skills/testimonial-video/SKILL.md",
      "computedHash": "0d71fa501ffc13fc979d4d93865785c16741c6b4041ccbb3926eaf0da0f486da"
    }
  }
 }