Things I’ve built, and why they exist.

Most agent skill libraries are a folder of prompts. In a regulated firm that is the part that gets the whole thing stopped, because nobody can say what a skill is allowed to touch, who approves it, or how you would know it had regressed.

This is the same idea built to survive that question: 173 skills across 14 functional domains — banking, capital markets, insurance, payments, risk, financial crime, model governance and the rest — each authored to a common lifecycle with an explicit risk tier (R1–R4), declared action boundaries, human-approval gates, handoffs to adjacent skills, and an eval pack that travels with it. The catalog is machine-readable, so coverage and drift are things you can compute rather than assert.

Conformant with the Agent Skills specification, which means it runs on general-purpose agent platforms rather than one vendor’s runtime.

The listening backlog is not a storage problem, it is a triage problem: you subscribe to more than you can hear, and the useful ten minutes are buried somewhere in a ninety-minute file.

Briefcast runs the whole path on hardware you control. A Go service handles subscriptions, the download queue, artwork, chapters, tags and retention against SQLite or Postgres; WhisperX does transcription on the box; and summaries go through any OpenAI-compatible endpoint — a hosted model or Ollama on the same machine, which is the difference between a tool you can point at a private feed and one you cannot. The Vue front end ships with it, and the whole thing runs from a Compose file.