Golden Path UXWired prototype · simulated AI (no live model)

Broadcasts — AI Compose

A clickable prototype of the Onboard "Broadcasts" admin tab, designed against the AI UX Golden Path. One natural-language input → channel-ready drafts → action-gated send/stage. Try the buttons.
Principle 1 · Capabilities
One sentence, not six dropdowns.

The model infers type, tone, channels & formatting from free text. The UI surface stays tiny.

Principle 2 · Action-based
It drafts & stages — doesn't describe.

Output is a per-channel artifact you send or copy, with the action gated to its blast radius.

Principle 3 · Zero-permission
Constraints pre-loaded.

Voice, guardrails, roster, live standings & schedule are known up front — the AI never asks.

Onboard · Admin

Gate TypesBroadcastsHealth
3Loaded context — the AI already knows this, so it won't ask
✓Kramer Advantage brand voice ✓Compliance guardrails (no income guarantees · disclaimers) ✓Privacy: first name + last initial only Live leaderboard data (issued & paid) Agent roster · 30 Zoom schedule · synced ✓Your role: Manager
Examples drop you onto the Golden Path — but any plain sentence works.
Drafts every enabled channel from this one input.
Reflecting — who's the audience and what outcome serves them best?
Predicting confusion — will the time zone / join link / standings be clear?
Checking guardrails — no income guarantees, disclaimer on standings, names masked.
Self-critique — drafted, reviewed against success criteria, revised once.
3Suggested guardrails for this broadcast — toggle optional ones; required ones are locked on
Suggestions are driven by the detected intent. The same library is encoded in the system prompt (see Broadcasts-Admin-Build-Spec.md §10) so the model self-applies them — these toggles just let an admin override the optional ones.
✓ Guardrails applied (model self-check)
    2Channel-ready drafts — each action gated to its blast radius
    GGroupMeExternal · confirm to send
    Also draft Email (SendGrid) Also draft SMS (Twilio) ↻ Syndication: one input → all enabled channels
    Why GroupMe asks but WhatsApp doesn't: sending to GroupMe is an external, outbound action (real people, hard to unsend) → confirm gate. Staging the WhatsApp message to the copy queue is internal and reversible → it auto-stages with an undo path, no permission prompt. Permission matches blast radius — nothing more.

    WhatsApp review queue — what the VA copies & posts

    Nothing staged yet. Generate a broadcast above — the WhatsApp draft auto-stages here.
    Capabilities, not crutches: notice there's no "message type" dropdown, no "tone" selector, no "pick channels" matrix on the input. The model reads one sentence and infers all of it — and the system prompt is where new behavior gets added, not new form fields. The detected-intent tags above are the model showing its work, editable if it guessed wrong.
    Zero-permission: the top bar is the whole point — brand voice, compliance guardrails, the live leaderboard feed, the agent roster, the Zoom schedule, and your role are all loaded before you type. Run 10 typical inputs through this and the model should never need to ask a clarifying question. If it does, that's a constraint we forgot to load — not a question for you.

    The Golden Path test (this prototype)

    1. A first-time manager types a normal sentence — ✓ no syntax to learn.
    2. The model produces channel-ready drafts accepted on the first try — ✓ (regenerate/adjust-tone are escape hatches, not the path).
    3. No special prompt format or "tip" required — ✓.
    4. No clarifying question — ✓ (constraints pre-loaded).
    5. Output is an action (send / staged draft), not a description — ✓.

    Send to GroupMe now?

    This posts to the live GroupMe group immediately — an external, outbound action. (WhatsApp drafts don't need this; they only stage internally.)