Music, data, and automations that make your life easier.

Behind the workflow

• Trigger — watches Gmail for new messages matching defined senders or labels, checked on a schedule.
• Field mapping — sender, subject, date and body map directly to sheet columns, with no transformation needed.

Validated

Tested with real inbound release and tour emails — every field captured correctly, with no duplicate rows on re-runs.

Key learning

The simplest integration in the lab turned out to be the most load-bearing one — every later automation reads from the structure this one creates.

Behind the workflow

• Two-pass reasoning — classification and extraction run as separate AI steps, so a wrong extraction doesn't undo a correct classification.
• Fallback path — anything the model can't confidently classify is stored as-is rather than forced into the wrong category.

Validated

Tested with a mixed batch of release announcements, tour dates and unrelated messages — correctly classified and extracted in every case.

Key learning

Splitting "what is this" from "what does it say" into two separate AI steps made both easier to get right than asking for everything at once.

Behind the workflow

• Verify before storing — a message claiming a release doesn't get saved as one until Spotify confirms it exists — cutting down on false positives.
• Router logic — message type decides the downstream path, so only genuine releases reach the Spotify search step.

Validated

Tested with real releases from Arde Bogotá and The National — both correctly matched, verified and stored with their Spotify links.

Key learning

Adding one verification step before storage removed nearly all of the false positives that came from AI classification alone.

Behind the workflow

• Two schedules, one source — the playlist updates the moment a release is detected; the digest runs on its own weekly schedule, reading the same stored data.
• No duplicate entries — each release is checked against what's already logged before being added again.

Validated

Running for several weeks with a real tracked-artist list — every digest correctly reflected that week's releases, with no repeats or missed entries.

Key learning

Separating "catch it now" from "summarise it weekly" into two independent schedules made both simpler than trying to do both in one workflow.

Behind the workflow

• Prompt contract — five fixed sections (Key Signals, Why It Matters, Connections, Watchlist, Priority), instructed to work only from the supplied evidence and separate fact from interpretation.
• Output format — the model writes valid HTML directly, so Gmail sends it as-is with no manual formatting step.

Validated

Tested end to end with three real signal types in one run — a release, a tour, a collaboration — correctly summarised, connected and prioritised.

Key learning

Aggregation is what makes analysis possible. Without it, the AI sees one row at a time; with it, the model receives one intelligence packet and can reason about the set — the same distinction that later shaped the scoring logic in 06 — Opportunity Radar.

Behind the workflow

• Output contract — artist, opportunity_score, level, reason, recommended_action, returned as JSON and parsed into fields Make can route on.
• Contextual intelligence, by design — the model is allowed to bring background knowledge to the score rather than treating the artist as a neutral identifier — closer to how real industry judgement works. The trade-off is explicit: this is an AI-assisted assessment, not a purely mathematical index.

Validated

All three priority routes tested with real data — a High-scoring alert delivered by email, a Medium signal added to the watchlist, a Low signal recorded without action.

Debugging notes

• An unexpected High result traced back to a downstream module reading a stale sheet — a reminder to inspect module output before touching logic.
• Proving the Low route required changing exactly one variable (artist) while holding the signal constant — isolating cause from coincidence.

Behind the workflow

• Reading the API's own feedback — the first request came back as a Bad Request; correcting the query syntax produced a clean 200 — a reminder that an error response is information, not just an obstacle.
• Filter vs. AI — where the selection rule is objective (score equals 100), a plain filter is the right tool. Contextual judgement across several fields is where an AI layer earns its place — this integration deliberately uses the former.

Validated

Tested with Radiohead: correct top match returned — MBID, type Group, country GB, relevance score 100.

Key learning

Widening the result limit during testing surfaced ranked alternatives, confirming the API returns candidates to evaluate, not a single guaranteed answer.

Behind the workflow

• Three-module architecture — Job Radar and Job Opportunity Analyst run as Scheduled Tasks; Application Workspace is a Claude Project that drafts tailored documents once a decision is made.
• Human-in-the-loop, by design — the system never applies on its own — it hands over a researched, contextualised opportunity and waits for a human decision before drafting anything.

Validated

• No opportunity worth flagging — Job Radar has also correctly returned nothing to review, rather than forcing a recommendation.
• Filtered after research — Job Opportunity Analyst has found real mismatches or constraints and concluded an opportunity wasn't worth pursuing further.
• Fit and gaps, side by side — genuinely promising opportunities have been surfaced with their strengths and their gaps stated honestly, not smoothed over.
• From decision to document — once a human decides to proceed — whether the opportunity came from Job Radar or was found manually — Application Workspace drafts a CV and cover letter from real experience only, never invented.

Key learning

The system is judged by its restraint as much as its research — it surfaces gaps honestly, leaves the geographic or seniority calls to a human, and never inflates a weak opportunity into a recommendation just to have something to show.