For 18 target accounts, a Claude agent scores buying signals and writes a cited, ready-to-work brief straight into Sanity's Content Lake — no spreadsheet, no separate tool. A rep opens Studio and works the list top‑down.
One brief, exactly as a rep receives it in Studio — scored, sourced, and ready to work.
Fit is icp × signal strength, scored 1–5. The top few:
Three accounts were held rather than published — a coded rubric checks grounding, score consistency, and evidence behind every competitor claim before anything goes live. A rep never opens a brief the system can't defend.
The role is an internal builder who ships GTM tooling on Sanity's own platform. This is that, in miniature: Studio and Content Lake as the application (no external database), a Claude tool-use agent with a versioned prompt and a coded eval gate, and a governed queue — every score, signal, and source traceable on the document itself.
Next, in the seat: move the batch script into a Sanity Function triggered on new account docs so briefs generate continuously; add a Studio component that shows live signal sources inline; wire real feeds (tech-stack detection, job-post ingestion) through MCP behind the same eval gate; and write back which briefs converted, feeding that into the scoring weights.