Spotify reporting,
on autopilot.

The numbers were there.
Collecting them didn't scale.

A company that manages and promotes a roster of artists wanted sharper tracking of how their advertising was actually performing on Spotify. The performance data existed, artist by artist, inside the Spotify for Artists view. Getting it out, into one place they could actually report on, was the hard part.

Every month, someone opened the platform and worked through it by hand, one artist at a time: navigate to the right screen, read the figures, copy them into a spreadsheet, move to the next artist, repeat. Across a full roster, hours disappeared before any analysis even began. And because it was manual, it was inconsistent and easy to skip, which made trending anything reliably over time close to impossible.

That was the real cost. Advertising decisions were being made on numbers that were partial, out of date, or slightly different depending on who collected them that month. When the goal is to see which promotion moved streams, saves and listeners, the data underneath has to be complete and consistent first. It wasn't.

One source of truth,
trended over time.

What a good outcome looked like was simple to describe and tedious to do by hand: every artist's Spotify numbers landing in one database automatically, in the same shape, on a schedule. With that in place, the team could line up advertising spend against streams, saves and listeners, compare artists, and watch the trend move month over month, across the whole roster, without anyone touching a cell.

So the brief was not a dashboard to look at. It was a dependable pipeline that fills their own database with clean Spotify data they can report on however they like, and keeps filling it, on its own.

Guided, then skilled,
then deterministic.

We didn't jump straight to code. We hardened the workflow in three deliberate stages, each one taking more of the human out of the loop, until what was left ran on its own and ran the same way every time.

Here's the path we took, from clicking by hand to a script that runs itself:

flowchart TD
  A["Stage 1: guided with MCP tools, a person drives the browser step by step"] --> B["We learn the exact steps the workflow needs"]
  B --> C["Stage 2: capture the steps as skills, an LLM runs the whole sequence on its own"]
  C --> D["Verify the workflow runs clean, end to end"]
  D --> E["Stage 3: compile the skills into local scripts, deterministic software you own"]

Why bother with all three stages? Because each one buys something the next needs. Guiding by hand finds the real steps, not the ones you assume. Skills prove those steps run without a human babysitting them. Scripts make the whole thing repeatable and cheap to run forever. Skip a stage and you either ship something brittle, or you spend far longer getting to something you can trust.

One command,
the whole roster.

The three stages above are how we built it, not three things running side by side. Each version replaced the one before it, so what runs day to day is just the last one: the script. A non-technical person just asks Claude to run it. From there it's hands-off: the scripts collect each artist's data from Spotify across the whole roster, put the numbers into one consistent shape, and write them straight into the company's database, ready for reporting and statistical tracking.

flowchart TD
  A["Ask Claude to run the scripts"] --> B["Scripts collect the data from Spotify, artist by artist across the roster"]
  B --> C["Normalize the numbers into one consistent shape"]
  C --> D["Write straight into the company's database"]
  D --> E["Reporting and statistical tracking, trended month over month"]

Because the collection is deterministic, the reports finally sit on data that's complete, consistent, and current, not a hand-copied sample that changes with whoever did it that month. Run it this month, run it next month, and the only thing that differs is the numbers themselves.

Reporting the old way
couldn't reach.

Once the data lands in one database on a schedule, the questions the team can answer change. The same numbers that used to take an afternoon to gather now sit ready to query, so reporting moves from "what did we manage to collect" to "what is actually happening". They can now ask:

• how each artist is performing on Spotify, month over month, without waiting on a manual pull;
• which advertising and promotion actually moved streams, saves and listeners, and which didn't;
• how the whole roster compares at a glance, on the same consistent numbers;
• where the advertising budget is working hardest, so the next spend goes where it pays;
• and how a campaign's effect holds up over time, not just in the week it launched.

None of that is new data. It was always there in Spotify. What changed is that it now arrives complete, consistent, and on time, which is the difference between a hunch and a report.

Hours a month,
back in the diary.

The team stopped repeating the same afternoon every month. Collection runs across every artist and lands in their database on its own, so ad-performance tracking sits on numbers that are actually trustworthy. And because it's scripts they own, they run it whenever they want, and extend it themselves when they add an artist or start tracking a new metric. No black box, no lock-in, no waiting on us.

Common questions.