THE THING WE BUILT:

Scout 2.0

Spotify keeps betting on AI DJ: a synthetic voice that “knows” your taste and picks the next track. It’s clever. It’s also sorta “mid.” You get the same safe pool of names, overused smooth transitions, none of the weird specificity that makes you actually trust a recommendation.

We used to live in simpler times: ask someone. Like your friend who listens to stuff you've never heard of. Or the record-store guy with obscure band patches. You didn’t want an algorithm’s confidence. You wanted a human who had an opinion and could explain why.

So last year, we built Scout. You name a band you like. Scout digs Reddit for what real people wrote (verbatim quotes, thread links), pulls out the band names, packs a YouTube playlist. Not “fans also like.” What someone actually said when asked.

This year, we wanted to revisit this project and do even better: same human source, sharper pipeline, and maybe even less noise.

So we built Scout again.

Scout 2.0 now with Deep Cut mode

How It Works

Scout is one agent with a short skill chain. Each step is a named capability with its own tools, not a character with a personality. The order is fixed: scout → dig → sniff → pack.

Scout runs the chain. It takes the seed band and depth mode (deep-cut or gateway), then formats the evidence pack. It never invents a recommendation. It only ships what search returned and what the filters kept.

Dig is the search skill. One Tavily call hits Reddit for human recommendations and comes back with candidate bands, verbatim quotes, and thread URLs. No paraphrase, no invented bands.

Sniff is the filter skill. It drops household defaults (Mogwai, Explosions in the Sky, Sigur Rós…), bare name-drops, weak quotes, and seed duplicates. Strict in deep-cut, looser in gateway.

Pack is the listen skill. It matches kept bands to YouTube tracks, rejects lookalike titles, and builds a playlist URL when it has hits.

Notes from Building

1. Ask humans. Process with agents.

The model never freestyles “if you like X try Y.” Every pick needs a quote and a thread. Humans supply the rec; the skill chain does the legwork.

Scout 1 had no filter: Reddit hits went straight to the playlist. In deep-cut mode every post-rock thread still names Mogwai and Explosions in the Sky, because humans do. Sniff is new: it drops household names, bare name-drops, and quotes that don’t actually support the band. Gateway stays loose. Deep cut is where Scout earns its name.

2. We switched stacks to try something new.

Scout 1 was LangChain. Scout 2.0 is on eve because it was new and we wanted to see if a skill chain (digger / sniffer / crate, each with its own instructions and tools) could beat the old single-pipeline version.

3. Coming back.

Scout v1 worked. It was fun to use. We could have left it alone.

A year later the tools moved, and so did we. Coming back is a little nostalgia, sure. It’s also the cheapest way to see how you’d ship the same idea with a year more practice. Ship fast when it’s fresh. Come back when you can do better, not just newer.

Built with ❤️ by Township

You’re receiving this because you signed up for Township’s “We Built a Thing” newsletter. We'll only send these when we, well, build things.