Prompt & Configuration Registry  ·  View 03 of 21  ·  People and journeys

Actors and Their Core Journeys

Nine actors, four of them machines, and the thing each of them is actually trying to get done.

Editable source SVG draw.io All views
People who change what the model says Prompt engineer ~40 across product teams Goal — Change the assistant's wording today, not in next week's release, and know before a customer sees it that I have not broken something. Core journeys Ship a prompt change ~200 / day Compare two wordings Pin a model version Product owner 1 per feature Goal — Approve what goes live without having to read a diff of a service I do not own, and see what it did afterwards. Core journeys Approve a promotion Read an experiment out Data scientist experiment owner Goal — Run two prompts against real traffic and get a number I am willing to defend, not a screenshot of three good answers. Core journeys Define an experiment Hold a variant out People who keep it standing On-call engineer 24×7 rota Goal — Make it stop, before I understand why — and be certain afterwards what was live when it went wrong. Core journeys Roll back at 3am target Pull a kill switch Platform engineer owns the registry Goal — Be able to take my own control plane down for an afternoon without a single product noticing. Core journeys Upgrade the control plane Rotate the signing key People who ask afterwards Compliance reviewer quarterly, plus incidents Goal — Be shown exactly what configuration produced a specific answer on a specific day, signed, without taking anybody's word for it. Core journeys Recover a past configuration Export signed evidence Support lead first to notice Goal — Tell whether the answer a customer is complaining about came from a change we made this morning. Core journeys Trace an answer to a digest Machines in the cast AI feature service ~300 deployments Goal — Get an answer to "what am I supposed to say?" in under a millisecond, and keep getting one when the registry is down. Core journeys Resolve a key Serve on a stale bundle Evaluation service gate input Goal — Be asked for a verdict on a specific digest, not on "the latest prompt". Core journeys Return a gate verdict CI pipeline per product repo Goal — Publish a version from the repo that owns the prompt, and fail the build when it will not validate. Core journeys Publish from a merge Who the Registry Is For, and What They Get To Do Person or role Journey / task External / third party Security / platform v 1.0 · owner Platform Architecture · d 2026-09

What the cast says about the design

  • Three distinct populations change behaviour — the person who writes the words, the person accountable for them, and the person measuring them — and they need different permissions on the same key. That is why authoring, approval and promotion are separate grants (ADR-11).
  • The on-call engineer does not own the prompt. Every rollback affordance is designed for someone who has never read the configuration they are reverting.
  • The compliance reviewer's goal — "without taking anybody's word for it" — is why bundles are signed rather than merely stored.

Assumptions

  • ≈ 40 prompt engineers across product teams, issuing ≈ 200 promotions/day at peak (stated assumption).
  • A 24×7 on-call rota already exists and the registry hooks into it rather than creating one.
  • Support leads have access to a conversation record that carries the digest. If they do not, the trace-an-answer journey has no starting point.

What this view rules out

  • A design where only the owning team can roll back. Two of the nine actors need to act on a key they do not own, at a time when nobody who owns it is awake.
  • Approval flows that assume the approver can read a diff of the service. Product owners approve wording, not code.