Tiktok App Marketing

TikTok App Marketing

Automate your entire TikTok slideshow marketing pipeline: generate → overlay → post → track → iterate.

Proven results: 7 million views on the viral X article, 1M+ TikTok views, $670/month MRR — all from an AI agent running on an old gaming PC.

Prerequisites

This skill does NOT bundle any dependencies. Your AI agent will need to research and install the following based on your setup. Tell your agent what you're working with and it will figure out the rest.

Required

• Node.js (v18+) — all scripts run on Node. Your agent should verify this is installed and install it if not.
• node-canvas (npm install canvas) — used for adding text overlays to slide images. This is a native module that may need build tools (Python, make, C++ compiler) on some systems. Your agent should research the install requirements for your OS.
• Discord webhook — this is the delivery backbone in this setup. Finished slides are sent as original image files to Discord so quality stays high; user then posts manually to TikTok with a trending sound.

Image Generation (pick one)

You choose what generates your images. Your agent should research the API docs for whichever you pick:

• OpenAI — gpt-image-1.5 (ALWAYS 1.5, never 1). Needs an OpenAI API key. Best for realistic photo-style images. This is what Larry uses and what we strongly recommend.
• Stability AI — Stable Diffusion XL and newer. Needs a Stability AI API key. Good for stylized/artistic images.
• Replicate — run any open-source model (Flux, SDXL, etc.). Needs a Replicate API token. Most flexible.
• Local — bring your own images. No API needed. Place images in the output directory and the script skips generation.

Conversion Tracking (optional but recommended for mobile apps)

• RevenueCat — this is what completes the intelligence loop. TikTok profile momentum tells you what gets attention. RevenueCat tells you which content windows drive paying users. Combined, the agent can distinguish between growth that looks good and growth that monetizes. Install the RevenueCat skill from ClaWHub (clawhub install revenuecat) for full API access to subscribers, MRR, trials, churn, and revenue. There's also a RevenueCat MCP for programmatic control over products and offerings from your agent/IDE.

Cross-Posting (optional, recommended)

Manual cross-posting to Instagram Reels, YouTube Shorts, Threads, Facebook, LinkedIn, and other platforms is still recommended. Same content, different algorithms, more reach.

First Run — Onboarding

When this skill is first loaded, IMMEDIATELY start a conversation with the user. Don't dump a checklist — talk to them like a human marketing partner would. The flow below is a guide, not a script. Be natural. Ask one or two things at a time. React to what they say. Build on their answers.

Important: Use scripts/onboarding.js --validate at the end to confirm the config is complete.

Phase 0: TikTok Account Warmup (CRITICAL — Don't Skip This)

Before anything else, check if the user already has a TikTok account with posting history. If they're creating a fresh account, they MUST warm it up first or TikTok will treat them like a bot and throttle their reach from day one.

Explain this naturally:

"Quick question before we dive in — do you already have a TikTok account you've been using, or are we starting fresh? If it's new, we need to warm it up first. TikTok's algorithm watches how new accounts behave, and if you go straight from creating an account to posting AI slideshows, it flags you as a bot and kills your reach."

If the account is new or barely used, walk them through this:

The goal is to use TikTok like a normal person for 7-14 days before posting anything. Spend 30-60 minutes a day on the app:

• Scroll the For You page naturally. Watch some videos all the way through. Skip others halfway. Don't watch every single one to the end — that's not how real people scroll.
• Like sparingly. Maybe 1 in 10 videos. Don't like everything — that's bot behaviour. Only like things you'd genuinely engage with in your niche.
• Follow accounts in your niche. If they're promoting a fitness app, follow fitness creators. Room design? Interior design accounts. This trains the algorithm to understand what the account is about.
• Watch niche content intentionally. This is the most important part. TikTok learns what you engage with and starts showing you more of it. You want the For You page dominated by content similar to what you'll be posting.
• Leave a few genuine comments. Not spam. Real reactions. A few per session.
• Maybe post 1-2 casual videos. Nothing promotional. Just normal content that shows TikTok there's a real person behind the account.

The signal to look for: When they open TikTok and almost every video on their For You page is in their niche, the account is warmed up. The algorithm understands them. NOW they can start posting.

Tell the user: "I know two weeks feels like wasted time, but accounts that skip warmup consistently get 80-90% less reach on their first posts. Do the warmup. It's the difference between your first post getting 200 views and 20,000."

If the account is already active and established, skip this entirely and move to Phase 1.

Phase 1: Get to Know Their App (Conversational)

Start casual. Something like:

"Hey! Let's get your TikTok marketing set up. First — tell me about your app. What's it called, what does it do?"

Then FOLLOW UP based on what they say. Don't ask all 9 questions at once. Pull the thread:

• They mention what it does → ask who it's for ("Who's your ideal user?")
• They describe the audience → ask about the pain point ("What's the main problem it solves for them?")
• They explain the problem → ask what makes them different ("What makes yours stand out vs alternatives?")
• Get the App Store / website link naturally ("Can you drop me the link?")
• Determine category (home/beauty/fitness/productivity/food/other) — often inferable

Don't ask for "brand guidelines" robotically. Instead: "Do you have any existing content or a vibe you're going for? Or are we starting fresh?"

Then ask about their app and monetization:

"Is this a mobile app? And do you use RevenueCat (or any subscription/in-app purchase system) to handle payments?"

This is critical because it determines whether we can close the full feedback loop. If they have a mobile app with RevenueCat:

• Tell them about the RevenueCat skill on ClawHub (clawhub install revenuecat). It gives full API access to subscribers, MRR, trials, churn, revenue, and transactions. Don't auto-install — just let them know it exists and what it unlocks, and they can install it if they want.
• Explain why it matters: Without RevenueCat data, the skill can only optimize for views (vanity metrics). With it, the skill optimizes for actual paying users. The difference is massive. A post with 200K views and zero conversions is worthless. A post with 5K views and 10 paid subscribers is gold. You can only tell the difference with RevenueCat connected.

If they don't use RevenueCat but have another subscription system, note it and work with what's available. If it's not a mobile app (e.g. physical product, SaaS, service), skip RevenueCat but still track whatever conversion metric they have (website signups, purchases, leads).

Store everything in tiktok-marketing/app-profile.json.

Phase 2: Competitor Research (Requires Browser Permission)

Before building any content strategy, research what competitors are doing on TikTok. This is critical — you need to know the landscape.

Ask the user:

"Before we start creating content, I want to research what your competitors are doing on TikTok — what's getting views in your niche, what hooks they're using, what's working and what's not. Can I use the browser to look around TikTok and the App Store?"

Wait for permission. Then:

1. Search TikTok for the app's niche (e.g. "interior design app", "lip filler filter", "fitness transformation app")
2. Find 3-5 competitor accounts posting similar content
3. Analyze their top-performing content:
   • What hooks are they using?
   • What slide format? (before/after, listicle, POV, tutorial)
   • How many views on their best vs average posts?
   • What's their posting frequency?
   • What CTAs are they using?
   • What music/sounds are trending in the niche?
4. Check the App Store for the app's category — look at competitor apps, their screenshots, descriptions, ratings
5. Compile findings into tiktok-marketing/competitor-research.json:

{
  "researchDate": "2026-02-16",
  "competitors": [
    {
      "name": "CompetitorApp",
      "tiktokHandle": "@competitor",
      "followers": 50000,
      "topHooks": ["hook 1", "hook 2"],
      "avgViews": 15000,
      "bestVideo": { "views": 500000, "hook": "..." },
      "format": "before-after slideshows",
      "postingFrequency": "daily",
      "cta": "link in bio",
      "notes": "Strong at X, weak at Y"
    }
  ],
  "nicheInsights": {
    "trendingSounds": [],
    "commonFormats": [],
    "gapOpportunities": "What competitors AREN'T doing that we could",
    "avoidPatterns": "What's clearly not working"
  }
}

6. Share findings with the user conversationally:

"So I looked at what's out there. [Competitor A] is doing well with [format] — their best post got [X] views using [hook type]. But I noticed nobody's really doing [gap]. That's our angle."

This research directly informs hook generation and content strategy. Reference it when creating posts.

Phase 3: Content Format & Image Generation

First, ask about format:

"Do you want to do slideshows (photo carousels) or video? Slideshows are what Larry uses and what this skill is built around — TikTok's data shows they get 2.9x more comments and 2.6x more shares than video, and they're much easier for AI to generate consistently. That said, if you want to try video, the skill supports it but it hasn't been battle-tested like slideshows have. Your call."

Store their choice as format: "slideshow" or format: "video" in config. If they pick video, note that the text overlay, 6-slide structure, and prompt templates are designed for slideshows. Video will require more experimentation and the agent should be upfront about that.

For slideshows (recommended):

Ask naturally:

"For the slideshows, we need images. I'd strongly recommend OpenAI's gpt-image-1.5 — it's what Larry uses and it produces images that genuinely look like someone took them on their phone. It's the difference between 'obviously AI' and 'wait, is that real?' You can also use Stability AI, Replicate, or bring your own images if you prefer."

⚠️ If they pick OpenAI, make sure the model is set to gpt-image-1.5 — NEVER gpt-image-1. The difference in quality is massive. gpt-image-1 produces noticeably AI-looking images that people scroll past. gpt-image-1.5 produces photorealistic results that stop the scroll. This one setting can be the difference between 1K and 100K views.

If they're unsure, always recommend gpt-image-1.5. It's the proven choice.

Store in config as imageGen with provider, apiKey, and model.

If they pick OpenAI, mention the Batch API:

"One thing worth knowing — OpenAI has a Batch API that's 50% cheaper than real-time generation. Instead of generating slides on the spot, you submit them as a batch job and get results within 24 hours (usually much faster). It's perfect for pre-generating tomorrow's slides overnight. Same quality, half the cost. Want me to set that up?"

If they're interested, store "useBatchAPI": true in imageGen config. The generate script supports both modes — real-time for quick iterations, batch for scheduled daily content.

Then — and this is critical — work through the image style with them. Don't just use a generic prompt. Bad images = nobody watches. Ask these naturally, one or two at a time:

"Now let's figure out what these images should actually look like. Do you want them to look like real photos someone took on their phone, or more like polished graphics or illustrations?"

Then based on their answer, dig deeper:

• What's the subject? "What are we actually showing? Rooms? Faces? Products? Before/after comparisons?"
• What vibe? "Cozy and warm? Clean and minimal? Luxurious? Think about what your audience relates to or aspires to."
• Consistency: "Should all 6 slides look like the same place or person? If yes — I need to lock down specific details so each slide doesn't look totally different."
• Must-have elements? "Anything that HAS to be in every image? A specific product? Certain furniture? A pet?"

Build the base prompt WITH them. A good base prompt looks like:

iPhone photo of a [specific room/scene], [specific style], [specific details].
Realistic lighting, natural colors, taken on iPhone 15 Pro.
No text, no watermarks, no logos.
[Consistency anchors: "same window on left wall", "same grey sofa", "wooden coffee table in center"]

Save the agreed prompt style to config as imageGen.basePrompt so every future post uses it.

Key prompt rules (explain these as they come up, don't lecture):

• "iPhone photo" + "realistic lighting" = looks real, not AI-generated
• Lock architecture/layout in EVERY slide prompt or each slide looks like a different place
• Include everyday objects (mugs, remotes, magazines) for lived-in feel
• For before/after: "before" = modern but tired, NOT ancient
• Portrait orientation (1024x1536) always — this is TikTok
• Extremely specific > vague ("small galley kitchen with white cabinets and a window above the sink" > "a kitchen")

NEVER use generic prompts like "a nice living room" or "a beautiful face" — they produce generic images that get scrolled past.

Phase 4: Discord Delivery Setup (ESSENTIAL — Powers the Entire Workflow)

Discord is the delivery layer in this setup. Without it, distribution becomes messy and quality often drops.

This setup gives you:

• High-quality delivery — final slides are sent as original image files
• Manual publish control — you choose trend sound right before posting
• Simple operations — no posting API dependency required
• Reliable handoff — every post package lands in one place

This flow still feeds the daily feedback loop (see Phase 8), but analytics are now profile-level.

Frame it naturally to the user:

"So here's the key piece — we use Discord for delivery. I send you the finished slide files, then you post manually to TikTok with a trending sound. That gives us better quality control and avoids API brittleness."

"This is free and open source. We keep the workflow lightweight: generate, deliver, post, analyze, iterate."

Walk them through connecting step by step:

1. Create a Discord channel for content delivery
2. Create a webhook in that channel
3. Add discord.webhookUrl to config
4. Run a delivery test with scripts/send-to-discord.js
5. Confirm files arrive correctly before moving on

Explain the publish workflow:

"One important thing — delivery is automatic, publishing is manual. Before you publish each post on TikTok, add a trending sound from TikTok's sound library. Music is one of the biggest reach levers."

Don't move on until Discord delivery works end-to-end.

Phase 5: Conversion Tracking (THE Intelligence Loop)

If they have a mobile app with RevenueCat (you should already know this from Phase 1), this is where the skill goes from "content automation" to "intelligent marketing system." This is the most important integration in the entire skill. Don't treat it as optional.

Explain WHY it matters:

"So right now, we can track profile momentum — follower growth, total likes growth, and video count changes. That's top-of-funnel direction. But momentum alone doesn't pay the bills."

"This is where RevenueCat comes in. It tracks your subscribers, trials, MRR, churn — the actual revenue. When I combine TikTok momentum signals with conversion data from RevenueCat, I can make genuinely intelligent decisions:"

"If momentum spikes but conversions stay flat, the hook works but CTA/funnel needs work. If momentum is modest but subscribers rise, content quality is strong and we should scale similar angles."

"Without RevenueCat, I'm optimizing for vanity metrics. With it, I'm optimizing for revenue."

Walk them through setup step by step:

1. Install the RevenueCat skill from ClaWHub:

   clawhub install revenuecat
   

   This installs the revenuecat skill (v1.0.2+) which gives full API access to your RevenueCat project — metrics overview, customers, subscriptions, offerings, entitlements, transactions, and more. It includes reference docs for every API endpoint and a helper script (scripts/rc-api.sh) for direct API calls.

2. Get your V2 secret API key from the RevenueCat dashboard:

   • Go to your RC project → Settings → API Keys
   • Generate a V2 secret key (starts with sk_)
   • ⚠️ This is a SECRET key — don't commit it to public repos

3. Set the environment variable:

   export RC_API_KEY=sk_your_key_here
   

4. Verify it works: Run ./skills/revenuecat/scripts/rc-api.sh /projects — should return your project details.

5. Optional: RevenueCat MCP — for programmatic control over products, offerings, and entitlements from your agent or IDE. Ask your agent to research setting this up.

What RevenueCat gives the daily report:

• GET /projects/{id}/metrics/overview → MRR, active subscribers, active trials, churn rate
• GET /projects/{id}/transactions → individual purchases with timestamps (for conversion attribution)
• The daily cron compares conversion movement with profile momentum windows (directional attribution across 24-72h)

The intelligence this unlocks:

• "Momentum spiked but conversions stayed flat" → hook is great, CTA/funnel needs work
• "Momentum was moderate but paid subscribers rose" → content converts well, scale similar angles
• "Conversions are consistently poor across all posts" → might be an app issue (onboarding, paywall, pricing) not a content issue — the skill flags this for investigation

Without RevenueCat: The loop still works on profile analytics (follower/like/video deltas). You can optimize for momentum. But you're flying blind on revenue.

With RevenueCat: You optimize for actual paying users. You can tell the difference between a viral post that makes nothing and a quiet post that drives $50 in subscriptions. This is the entire point of the feedback loop. Every decision the daily report makes is better with RevenueCat data.

If they don't use RevenueCat or don't have subscriptions, the skill still works but the feedback loop is limited to profile-level optimization only.

Phase 6: Content Strategy (Built from Research)

Using the competitor research AND the app profile, build an initial content strategy:

"Based on what I found and what your app does, here's my plan for the first week..."

Present:

1. 3-5 hook ideas tailored to their niche + competitor gaps
2. Posting schedule recommendation (default: 7:30am, 4:30pm, 9pm — their timezone)
3. Which hook categories to test first (reference what worked for competitors)
4. Cross-posting plan (which platforms, same or adapted content)

Save the strategy to tiktok-marketing/strategy.json.

Phase 7: Set Up the Daily Analytics Cron

This is what makes the whole system self-improving. Set up a daily cron job that:

1. Collects daily profile snapshots (scripts/scrape-tiktok-analytics.js)
2. Pulls conversion data from RevenueCat (if connected)
3. Cross-references profile momentum with conversions to diagnose what's working
4. Generates a report with specific recommendations
5. Suggests new hooks based on performance patterns

Explain to the user:

"I'm going to set up a daily check that runs every morning. It logs your TikTok profile momentum over the last few days and, if RevenueCat is connected, compares that with actual conversions. Then it tells you exactly what's working and what to change."

"Because this is profile-level analytics, attribution is directional. We use consistent daily snapshots and compare 3-7 day windows."

Set up the cron:

Use the agent's cron system to schedule a daily analytics job. Run it every morning before the first post of the day (e.g. 7:00 AM in the user's timezone) so the report informs that day's content:

Schedule: daily at 07:00 (user's timezone)
Task: Run scripts/daily-report.js --config tiktok-marketing/config.json --days 3
Output: tiktok-marketing/reports/YYYY-MM-DD.md + message to user with summary

The daily report uses the diagnostic framework:

• High momentum + High conversions → Scale it — more of the same, test posting times
• High momentum + Low conversions → Hook works, CTA is broken — test new CTAs on slide 6, check app landing page
• Low momentum + High conversions → Content converts but nobody sees it — test radically different hooks, keep the CTA
• Low momentum + Low conversions → Full reset — new format, new audience angle, new hook categories

This is the intelligence layer. Without it, you're just posting and hoping. With it, every day's content is informed by data.

Phase 8: Save Config & First Post

Store everything in tiktok-marketing/config.json (this is the source of truth for the entire pipeline):

{
  "app": {
    "name": "AppName",
    "description": "Detailed description",
    "audience": "Target demographic",
    "problem": "Pain point it solves",
    "differentiator": "What makes it unique",
    "appStoreUrl": "https://...",
    "category": "home|beauty|fitness|productivity|food|other",
    "isMobileApp": true
  },
  "imageGen": {
    "provider": "openai",
    "apiKey": "sk-...",
    "model": "gpt-image-1.5"
  },
  "discord": {
    "webhookUrl": "https://discord.com/api/webhooks/...",
    "username": "Larry Marketing Bot"
  },
  "tiktok": {
    "username": "your_handle",
    "profileUrl": "https://www.tiktok.com/@your_handle",
    "secUid": "",
    "region": "US",
    "language": "en"
  },
  "revenuecat": {
    "enabled": false,
    "v2SecretKey": "sk_...",
    "projectId": "proj..."
  },
  "posting": {
    "schedule": ["07:30", "16:30", "21:00"],
    "manualPostingNotes": "Slides are delivered to Discord as image files. You post manually to TikTok."
  },
  "competitors": "tiktok-marketing/competitor-research.json",
  "strategy": "tiktok-marketing/strategy.json"
}

Then generate the first test slideshow — but set expectations:

"Let's create our first slideshow. This is a TEST — we're dialing in the image style, not posting yet. I'll generate 6 slides and we'll look at them together. If the images look off, we tweak the prompts and try again. The goal is to get the look nailed down BEFORE we start posting."

⚠️ THE REFINEMENT PROCESS IS PART OF THE SKILL:

Getting the images right takes iteration. This is normal and expected. Walk the user through it:

1. Generate a test set of 6 images using the prompts you built together
2. Show them the results and ask: "How do these look? Too polished? Too dark? Wrong vibe? Wrong furniture?"
3. Tweak based on feedback — adjust the base prompt, regenerate
4. Repeat until they're happy — this might take 2-5 rounds, that's fine
5. Lock the prompt style once it looks right — save to config

Things to watch for and ask about:

• "Are these realistic enough or do they look AI-generated?"
• "Is the lighting right? Too bright? Too moody?"
• "Does this match what your users would actually relate to?"
• "Are the everyday details right? (furniture style, objects, layout)"

You do NOT have to post anything you don't like. The first few generations are purely for refining the prompt. Only start posting once the images consistently look good. The agent learns from each round — what works, what doesn't, what to emphasise in the prompt.

Once the style is locked in, THEN use the hook strategy from competitor research and their category (see references/slide-structure.md) and start the posting schedule.

Core Workflow

1. Generate Slideshow Images

Use scripts/generate-slides.js:

node scripts/generate-slides.js --config tiktok-marketing/config.json --output tiktok-marketing/posts/YYYY-MM-DD-HHmm/ --prompts prompts.json

The script auto-routes to the correct provider based on config.imageGen.provider. Supports OpenAI, Stability AI, Replicate, or local images.

⚠️ Timeout warning: Generating 6 images takes 3-9 minutes total (30-90 seconds each for gpt-image-1.5). Set your exec timeout to at least 600 seconds (10 minutes). If you get spawnSync ETIMEDOUT, the exec timeout is too short. The script supports resume — if it fails partway, re-run it and completed slides will be skipped.

Critical image rules (all providers):

• ALWAYS portrait aspect ratio (1024x1536 or 9:16 equivalent) — fills TikTok screen
• Include "iPhone photo" and "realistic lighting" in prompts (for AI providers)
• ALL 6 slides share the EXACT same base description (only style/feature changes)
• Lock key elements across all slides (architecture, face shape, camera angle)
• See references/slide-structure.md for the 6-slide formula

2. Add Text Overlays

This step uses node-canvas to render text directly onto your slide images. This is how Larry produces slides that have hit 1M+ views on TikTok — the text sizing, positioning, and styling are dialled in from hundreds of posts.

Setting Up node-canvas

Before you can add text overlays, your human needs to install node-canvas. Prompt them:

"To add text overlays to the slides, I need a library called node-canvas. It renders text directly onto images with full control over sizing, positioning, and styling — this is what Larry uses for his viral TikTok slides.

Can you run this in your terminal?"

npm install canvas

"If that fails, it's because node-canvas needs some system libraries. Here's what to install first:"

macOS:

brew install pkg-config cairo pango libpng jpeg giflib librsvg
npm install canvas

Ubuntu/Debian:

sudo apt-get install build-essential libcairo2-dev libpango1.0-dev libjpeg-dev libgif-dev librsvg2-dev
npm install canvas

Windows:

# node-canvas auto-downloads prebuilt binaries on Windows
npm install canvas

"Once installed, I can handle everything else — generating the overlays, sizing the text, positioning it perfectly. You won't need to touch this again."

Don't skip this step. Without node-canvas, the text overlays won't work. If installation fails, help them troubleshoot — it's usually a missing system library. Once it's installed once, it stays.

How Larry's Text Overlay Process Works

1. Load the raw slide image into a node-canvas
2. Configure text settings based on the text length for that specific slide
3. Draw the text with white fill and thick black outline
4. Review the output — check sizing, positioning, readability
5. Adjust and re-render if anything looks off
6. Save the final image once it looks right

Exact code Larry uses:

const { createCanvas, loadImage } = require('canvas');
const fs = require('fs');

async function addOverlay(imagePath, text, outputPath) {
  const img = await loadImage(imagePath);
  const canvas = createCanvas(img.width, img.height);
  const ctx = canvas.getContext('2d');
  ctx.drawImage(img, 0, 0);

  // ─── Adjust font size based on text length ───
  const wordCount = text.split(/\s+/).length;
  let fontSizePercent;
  if (wordCount <= 5)       fontSizePercent = 0.075;  // Short: 75px on 1024w
  else if (wordCount <= 12) fontSizePercent = 0.065;  // Medium: 66px
  else                      fontSizePercent = 0.050;  // Long: 51px

  const fontSize = Math.round(img.width * fontSizePercent);
  const outlineWidth = Math.round(fontSize * 0.15);
  const maxWidth = img.width * 0.75;
  const lineHeight = fontSize * 1.3;

  ctx.font = `bold ${fontSize}px Arial`;
  ctx.textAlign = 'center';
  ctx.textBaseline = 'top';

  // ─── Word wrap ───
  const lines = [];
  const manualLines = text.split('\n');
  for (const ml of manualLines) {
    const words = ml.trim().split(/\s+/);
    let current = '';
    for (const word of words) {
      const test = current ? `${current} ${word}` : word;
      if (ctx.measureText(test).width <= maxWidth) {
        current = test;
      } else {
        if (current) lines.push(current);
        current = word;
      }
    }
    if (current) lines.push(current);
  }

  // ─── Position: centered at ~28% from top ───
  const totalHeight = lines.length * lineHeight;
  const startY = (img.height * 0.28) - (totalHeight / 2);
  const x = img.width / 2;

  // ─── Draw each line ───
  for (let i = 0; i < lines.length; i++) {
    const y = startY + (i * lineHeight);

    // Black outline
    ctx.strokeStyle = '#000000';
    ctx.lineWidth = outlineWidth;
    ctx.lineJoin = 'round';
    ctx.miterLimit = 2;
    ctx.strokeText(lines[i], x, y);

    // White fill
    ctx.fillStyle = '#FFFFFF';
    ctx.fillText(lines[i], x, y);
  }

  fs.writeFileSync(outputPath, canvas.toBuffer('image/png'));
}

Key details that make Larry's slides look professional:

• Dynamic font sizing — short text gets bigger (75px), long text gets smaller (51px). Every slide is optimized.
• Word wrap — respects manual \n breaks but also auto-wraps lines that exceed 75% width. No squashing.
• Centered at 28% from top — text block is vertically centered around this point, not pinned to it. Stays in the safe zone regardless of line count.
• Thick outline — 15% of font size. Makes text readable on ANY background.
• Manual line breaks preferred — use \n in your text for control. Keep lines to 4-6 words.

Text content rules:

• REACTIONS not labels — "Wait... this is actually nice??" not "Modern minimalist"
• 4-6 words per line — short lines are scannable at a glance
• 3-4 lines per slide is ideal
• No emoji — canvas can't render them reliably
• Safe zones: No text in bottom 20% (TikTok controls) or top 10% (status bar)

The difference between OK slides and viral slides is in these details. Larry's slides consistently hit 50K-150K+ views because the text is sized right, positioned right, and readable at a glance while scrolling.

⚠️ LINE BREAKS ARE CRITICAL — Read This:

The texts.json file must contain text with \n line breaks to control where lines wrap. If you pass a single long string without line breaks, the script will auto-wrap, but manual breaks look much better because you control the rhythm.

Good (manual breaks, 4-6 words per line):

[
  "I showed my landlord\nwhat AI thinks our\nkitchen should look like",
  "She said you can't\nchange anything\nchallenge accepted",
  "So I downloaded\nthis app and\ntook one photo",
  "Wait... is this\nactually the same\nkitchen??",
  "Okay I'm literally\nobsessed with\nthis one",
  "Snugly showed me\nwhat's possible\nlink in bio"
]

Bad (no breaks — will auto-wrap but looks worse):

[
  "I showed my landlord what AI thinks our kitchen should look like",
  ...
]

Rules for writing overlay text:

1. 4-6 words per line MAX — short lines are scannable at a glance
2. Use \n to break lines — gives you control over the rhythm
3. 3-4 lines per slide is ideal — more lines are fine, they won't overflow
4. Read it out loud — each line should feel like a natural pause
5. No emoji — canvas can't render them, they'll show as blank
6. REACTIONS not labels — "Wait... this is nice??" not "Modern minimalist"

The script auto-wraps any line that exceeds 75% width as a safety net, but always prefer manual \n breaks for the best visual result.

3. Deliver Slides to Discord

Use scripts/send-to-discord.js:

node scripts/send-to-discord.js --config tiktok-marketing/config.json --dir tiktok-marketing/posts/YYYY-MM-DD-HHmm/ --caption "caption" --title "title"

Why We Publish Manually — Best Practice

Slides are delivered automatically, but TikTok publishing is manual. This is intentional and critical:

1. Music is everything on TikTok. Trending sounds massively boost reach. You need to pick what's hot RIGHT NOW in your niche.
2. You add the music manually, then publish in TikTok. Takes ~30 seconds per post.
3. Posts without music get buried. Silent slideshows look like ads and get skipped.
4. Creative control. You can preview final pacing + caption + sound before going live.

This is still the workflow that helped us hit 1M+ TikTok views and $670/month MRR. Don't skip the music step.

Tell the user during onboarding: "Slides land in Discord as image files. Before publishing on TikTok, add a trending sound — it makes a massive reach difference."

Cross-posts to other platforms (Instagram, YouTube, etc.) are manual in this setup.

Caption rules: Long storytelling captions (3x more views). Structure: Hook → Problem → Discovery → What it does → Result → max 5 hashtags. Conversational tone.

4. Collect Profile Analytics (After Publishing)

After publishing, analytics are collected at profile level (not exact per-post metrics).

Run daily collector:

node scripts/scrape-tiktok-analytics.js --config tiktok-marketing/config.json

Inspect trend window:

node scripts/check-analytics.js --config tiktok-marketing/config.json --days 3

Generate daily report:

node scripts/daily-report.js --config tiktok-marketing/config.json --days 3

What gets tracked daily:

1. followerCount
2. followingCount
3. heartCount (total profile likes)
4. videoCount

⚠️ Known Limitation: No Exact Per-Video Attribution

This setup intentionally avoids fragile per-video API coupling. That means attribution is directional, not exact:

1. We detect momentum using daily deltas/spikes
2. We compare those windows to conversion movement (if RevenueCat is connected)
3. We infer likely winning hooks/CTAs from timing and context
4. We validate with repeated patterns over multiple days

Best practice:

1. Capture profile snapshots at consistent times every day
2. Keep a stable 3-7 day analysis window
3. Change one major variable at a time (hook or CTA)
4. Double down only after repeated positive signals
5. Use RevenueCat to confirm monetization impact

See references/analytics-loop.md for full analytics loop docs.

The Feedback Loop (CRITICAL — This is What Makes It Work)

This is what separates "posting TikToks" from "running a marketing machine." The daily cron pulls data from two sources:

1. TikTok profile analytics → follower/like/video momentum (daily deltas)
2. RevenueCat (if connected) → conversion data (trial starts, paid subscriptions, revenue)

Combined, the agent can make intelligent decisions about what to do next — not guessing, not vibes, actual data-driven optimization.

The Daily Cron (Set Up During Onboarding)

Every morning before the first post, the cron runs the analytics scripts:

1. Collects daily profile snapshots (followers, likes, videos)
2. Builds trend windows and delta/spike signals
3. If RevenueCat is connected, pulls conversion events in the same window (24-72h)
4. Cross-references momentum windows with paying-user movement
5. Applies the diagnostic framework (below) to determine what's working
6. Generates tiktok-marketing/reports/YYYY-MM-DD.md with findings
7. Messages the user with a summary + suggested hooks for today

The Diagnostic Framework

This is the core intelligence. Two axes: momentum (are people discovering/responding?) and conversions (are people paying?).

High momentum + High conversions → 🟢 SCALE IT

• This is working. Make 3 variations of the winning hook immediately
• Test different posting times to find the sweet spot
• Cross-post to more platforms for extra reach
• Don't change anything about the CTA — it's converting

High momentum + Low conversions → 🟡 FIX THE CTA

• The hook is doing its job — people are engaging. But they're not downloading/subscribing
• Try different CTAs on slide 6 (direct vs subtle, "download" vs "search on App Store")
• Check if the app landing page matches the promise in the slideshow
• Test different caption structures — maybe the CTA is buried
• The hook is gold — don't touch it. Fix everything downstream

Low momentum + High conversions → 🟡 FIX THE HOOKS

• The people who DO see it are converting — the content and CTA are great
• But not enough people are discovering it, so the hook/thumbnail isn't stopping the scroll
• Test radically different hooks (person+conflict, POV, listicle, mistakes format)
• Try different posting times and different slide 1 images
• Keep the CTA and content structure identical — just change the hook

Low momentum + Low conversions → 🔴 FULL RESET

• Neither the hook nor the conversion path is working
• Try a completely different format or approach
• Research what's trending in the niche RIGHT NOW (use browser)
• Consider a different target audience angle
• Test new hook categories from scratch
• Reference competitor research for what's working for others

High momentum + High downloads + Low paying subscribers → 🔴 APP ISSUE

• The marketing is working. People are engaging AND downloading. But they're not paying.
• This is NOT a content problem — the app onboarding, paywall, or pricing needs fixing.
• Check: Is the paywall shown at the right time? Is the free experience too generous?
• Check: Does the onboarding guide users to the "aha moment" before the paywall?
• Check: Is the pricing right? Too expensive for the perceived value?
• This is a signal to pause posting and fix the app experience first

High momentum + Low downloads → 🟡 CTA ISSUE

• People are engaging but not downloading. The hooks work, the CTAs don't.
• Rotate through different CTAs: "link in bio", "search on App Store", app name only, "free to try"
• Check the App Store page — does it match what the TikTok shows?
• Check that "link in bio" actually works and goes to the right place

The daily report automates all of this. It cross-references TikTok profile momentum with downloads and revenue (RevenueCat) and tells you which part of the funnel is likely broken. It also auto-generates new hook suggestions based on your winning patterns and flags when CTAs need rotating.

Hook Evolution

Track in tiktok-marketing/hook-performance.json:

{
  "hooks": [
    {
      "postRef": "2026-02-15-0730-bedroom-v1",
      "text": "My boyfriend said our flat looks like a catalogue",
      "app": "snugly",
      "date": "2026-02-15",
      "followerDelta": 38,
      "heartDelta": 4600,
      "videoDelta": 1,
      "conversions": 4,
      "cta": "Download Snugly — link in bio",
      "lastChecked": "2026-02-16"
    }
  ],
  "ctas": [
    {
      "text": "Download [App] — link in bio",
      "timesUsed": 5,
      "totalFollowerDelta": 110,
      "totalHeartDelta": 9200,
      "totalConversions": 8,
      "conversionRate": 0.067
    },
    {
      "text": "Search [App] on the App Store",
      "timesUsed": 3,
      "totalFollowerDelta": 62,
      "totalHeartDelta": 5100,
      "totalConversions": 12,
      "conversionRate": 0.141
    }
  ],
  "rules": {
    "doubleDown": ["person-conflict-ai"],
    "testing": ["listicle", "pov-format"],
    "dropped": ["self-complaint", "price-comparison"]
  }
}

The daily report updates this automatically. Each post gets tag

…(truncated)