Supercut

A supercut of what's working in your market's ads — every pattern counted, every claim with receipts.

Open a finished playbook first: coffee · protein powder — download and open in a browser, no setup or keys needed.

Supercut takes a market name and produces a self-contained HTML playbook of what that market's paid social video ads keep doing: format chapters with counted patterns, hook-frame strips from the ads themselves, and three tests to run next, each with a draft script. It's built for a creative strategist or founder who needs to brief a shoot, not read a data dump.

The point of view: longevity is an admission filter, not a verdict. Advertisers kill losing creative fast, so an ad still delivering after 30+ days — and seen in the last 30 — has earned study. The search strata include the market's heaviest spenders, so "still delivering" means real money kept backing the ad, not just that it lingered. No single ad is ever called a winner; conclusions come only from counting regularities across the admitted pool ("34 of 39 ads show the product in the first 3 seconds").

How it thinks: describe → tally → explain

1. Describe (AI): a video model watches one ad at a time and fills out a fixed questionnaire. It never sees the other ads, so it cannot invent market-level patterns.
2. Tally (code): deterministic TypeScript counts the answers across the pool. Every count keeps its contributing ad ids as receipts.
3. Explain (AI): one final call phrases the counted patterns into chapters and tests. It sees only the tallies, so it cannot cite a number that has no count behind it.

Run it

bun install
cp .env.example .env   # add PIPISPY_API_KEY and OPENROUTER_API_KEY

bun run scan coffee    # writes examples/coffee.html (about a minute, ~240 API credits)
bun run dev            # web app: browse playbooks, start scans, generate brand briefs

ffmpeg must be on PATH. Deterministic modules are tested: bun test.

Limits, stated plainly

• "Still delivering" is a proxy: the ad library has no live flag, so recency of last_put_time stands in for it.
• Survival says an ad kept earning spend, not that it was profitable.
• The video model's visual timestamps drift 2–5 seconds, so timing is reported in coarse bins (first 3s / 3–10s / later), and cut counts come from ffmpeg, not the model. Each playbook's appendix measures where the model was noisy.
• One market, one snapshot. Patterns are what's working now, not laws.

What I didn't build and why

• A whitespace finder. "Nobody does X" needs census-level coverage, and a gap can mean "tried and died" as easily as "opportunity". The playbook flags what almost nobody in the pool does — as an observation with a count, not as advice to charge in.
• Winners vs losers comparison. The library shows when ads stopped being seen, but dead ≠ failed (campaigns end, budgets move). Longevity as an admission filter uses the trustworthy half of that signal.
• A full strategist copilot (chat, scoring dashboards, asset generation). Three products in a trenchcoat. One document a strategist can act on beats all three half-done.
• Advertiser-anchored market scans. The ad library can search by advertiser and rank top spenders per category — a competitor-intelligence spine that's a different product from a market playbook.
• Meta Ad Library reach data. The provider exposes audience reach and active days for Meta ads — real delivery data, stronger than play counts — but it's a second data model whose numbers don't compare cleanly with the play-count pool a single document tallies across.

Where this goes next

Grounded in API capabilities verified during this build and cut for scope:

• A sameness score per format: submit an exemplar hook frame to the provider's image-similarity search and count visually near-identical ads across brands — the credible version of a whitespace finder.
• A living playbook: ad monitor tasks (~$1 per advertiser per month) on the market's top advertisers, regenerating the document with a diff of what changed since last month.
• Brand deep-dives: per-competitor launch cadence and workhorse assets from the store ad-schedule and longest-running-creative endpoints.

How I used AI

In the product: a video model fills a fixed per-ad questionnaire (describe); a text model orders the watch queue by caption relevance (triage), phrases counted tallies (explain), and adapts tests to a brand (brief). Every model output is schema-validated; drift is measured against ffmpeg and the provider's transcripts in each playbook's appendix. In the process: this repo was built with an AI coding agent doing the implementation, with the design, architecture rules, and quality bar specified up front and enforced by review, plus recon runs against the live APIs before the pipeline was written.

docs/architecture.md (what and where) and docs/decisions.md (why)