app-growth-design

A Claude Code skill for the decisions that actually move an app: onboarding, paywalls, pricing, retention, gamification, growth loops and screen-level craft.

It is not another "be a helpful product expert" prompt. It is a knowledge base — 5.7 hours of teardowns from Tim Gabe's YouTube channel compressed into six playbooks where every non-obvious claim carries a [NN] marker back to the video it came from, and every case study keeps its numbers.

npx app-growth-design

That's it. Pick user-wide or project-only, restart Claude Code, and the skill loads itself whenever you're working on one of those problems.

tl;dv

Too long; didn't view — which is rather the point.

• What. 30 app teardowns → six playbooks your agent reads instead of guessing: onboarding, paywalls, retention, psychology, growth, craft.
• Why it's not slop. Every non-obvious claim carries a [NN] marker back to the video. 243 named products with their real numbers. Where the source contradicts itself, the playbooks say so instead of picking a side.
• The trick. A design channel keeps its meaning on screen — "look at what they did here" and the captions carry here and nothing else. So the pipeline reads 1,161 screenshots too, and the notes come out 2.4× bigger than the speech.
• Install. npx app-growth-design → restart Claude Code → describe your problem. It loads itself.
• Credit. All the insight is Tim Gabe's. This repo ships synthesized notes and the pipeline — no captions, no video, no screenshots.
• Caveat. One practitioner's opinionated view, not settled fact. Cases with real numbers are evidence; the rest are hypotheses to test.

What changes when it's installed

Without it, ask Claude "my paywall converts at 2%, what do I test first?" and you get sensible, generic advice: try annual vs monthly, test your copy, add social proof.

With it, the same question routes into monetization-paywalls.md and comes back with an ordering that has evidence attached:

Structure out-tests price, always. Ranked LTV uplift across the corpus: locale 62.3% > trial structure 59.6% > plan duration 58.7% > plan count 57.1% > price 45.5% > visuals 34.6%. Adding a free trial to a weekly plan took 12-month LTV from $7.40 to $54.50 (+636%) — no price test does that. [06]

But first check which layer you're on: 57.7% of new apps never cross $1,000 in revenue. Below the top decile this isn't a monetization problem. [06]

Same for the rest:

| You're working on | It routes to | Sample of what's in there | |---|---|---| | First-run flow, quiz vs. short flow, time-to-value, activation | onboarding-activation.md | Granola ships value in 2 screens ($1.5B); Cal AI runs a 20+ step quiz ($35M year one); Noom runs 113 screens. Length is only justified if it buys visible personalization or a buyer filter. | | Paywall layout, trials, plan count, price tiers | monetization-paywalls.md | Weekly plans convert 1.7–7.4× better than annual at every price tier. 82–89% of trial starts happen on install day. | | Streaks, leaderboards, badges, churn, reward ceremonies | retention-gamification.md | A 35-intervention meta-analysis: gamification lifts autonomy (g=0.638) and relatedness (g=1.776) but barely touches competence (g=0.277). Depth is non-monotonic — past a point the fix is cutting. | | Why users behave this way — identity, dopamine, loss aversion | behavioral-psychology.md | Build Mirror → Label → Flex. Exit question: if a user left today, what would they rebuild from scratch? "Nothing much" = no moat. | | Positioning vs. an incumbent, moats, share loops, landing pages | growth-marketing.md | Design the share artifact, not the share button. Fuse share with invite at a 1-friend threshold. Attachment before virality. | | Type scale, spacing, color budget, motion, microcopy | design-craft.md | The countable craft checklist, the subtraction pass, the five-criterion screen audit. |

243 named products are indexed with what each one exemplifies, so the model can reach for a real citation instead of inventing a plausible one.

Install

# interactive — asks user-wide or project-only
npx app-growth-design

# non-interactive
npx app-growth-design --user      # ~/.claude/skills/app-growth-design
npx app-growth-design --project   # ./.claude/skills/app-growth-design
npx app-growth-design --dir ~/some/skills

# later
npx app-growth-design --force       # update in place
npx app-growth-design --uninstall

Straight from the repo works too, if you'd rather skip the registry: npx github:bixxter/app-growth-design

Node 18+. No dependencies — the installer is one file that copies skill/ into your skills directory and nothing else.

Prefer to do it by hand? cp -r skill ~/.claude/skills/app-growth-design. It works in anything that reads the Agent Skills format.

How it was built

The short version. The long version — every decision and the two things that went wrong — is in docs/PROCESS.md.

The problem with turning a YouTube channel into agent knowledge is that the naive path fails twice. Dumping raw captions gives you 220k words of rolling-window duplicates, most of it filler. And a design channel is the worst possible case for captions alone: the man is pointing at a screen. "Look at how they did this" is the entire claim, and the audio carries none of it.

So the pipeline is built around one idea: the frames are not decoration, they are the missing half of the content.

Note what happens to the word count in the middle: the notes come out more than twice the size of the speech. That's the whole point — 1,161 screenshots put back what the captions lost.

Three details that turned out to matter more than expected:

• Captions lag the screen by a beat. Frames get extracted 2–4 seconds after he starts describing something, or you capture the previous slide.
• The verify pass is where quality comes from. The single most common defect was a note that still said "the example shown here" — a reference the downstream agent can't use. Verification agents re-extract the frame, look at it, and write in what was actually there.
• [NN] markers everywhere. Every claim in the playbooks points back at a source video. That's what makes the skill arguable rather than oracular — you can check it, and so can the model.

Rebuild it for a different channel

Everything needed is in pipeline/, and the whole thing is channel-agnostic — swap pipeline/manifest.tsv for your own list of videos and re-run. See pipeline/README.md for the exact commands.

pipeline/
├── manifest.tsv               num ⇥ video_id ⇥ title ⇥ duration_s — the only input you change
├── 1-fetch-subs.sh            captions, 6 parallel
├── 2-clean-vtt.py             dedupe + regroup into ~30s timestamped blocks
├── 3-fetch-video.sh           480p, 3 parallel
├── 3b-retry-video.sh          sweep for what YouTube throttled
├── 4-enrich.workflow.js       one agent per video: frames + note, then verify
├── 5-synthesize.workflow.js   six agents: notes → playbooks → SKILL.md
└── recover.workflow.js        the lean rewrite of stage 4, after the first run
                               burned itself out (see docs/PROCESS.md)

Stages 4–5 are Claude Code Workflow scripts — deterministic orchestration over parallel subagents. Stages 1–3 need yt-dlp and ffmpeg.

Honesty section

This is one practitioner's opinionated view, not settled fact. Nearly every mechanic in the corpus is asserted from survivorship-selected case studies. There are essentially no before/after deltas attributable to a specific change. Treat the cases that carry real numbers as evidence and the rest as hypotheses worth testing.

The corpus contradicts itself, and the playbooks say so rather than papering over it — short vs. long onboarding, friction-removal vs. friction-as-amplifier, design-as-moat vs. design-is-not-the-moat. Each reference file ends with Anti-patterns, Contradictions and Caveats. Read them before advising anyone.

Agency promos are flagged. Tim runs a design agency and pitches it 2–4 times per video; several case studies are his own unaudited client work. Those segments are marked as promotional. Use the mechanics, never the implied results.

Attribution & licensing

All the underlying insight belongs to Tim Gabe — if this is useful to you, go watch the source videos. Every [NN] marker in the skill links to one.

This repo ships structured, synthesized notes, not republished transcripts. No captions, no video, no screenshots are included here; the pipeline that produces them from public YouTube content is included so you can reproduce the work yourself.

The pipeline code is MIT (LICENSE). The distilled playbooks in skill/ are commentary on publicly published material — use them, and keep the attribution.

Not affiliated with, endorsed by, or reviewed by Tim Gabe.