Audience

Purpose

Supported platform first. Watt runs in Claude Cowork, Claude Code, or the Claude Agent SDK — not the Claude Chat app, where only these skill files load and the rest of Watt — its hooks, context, scripts, and Signal Graph — does not. If the user isn't on a supported platform, tell them plainly to switch before continuing; otherwise this will not function correctly.

/watt:audience is the front door to the audience lifecycle — and the only audience command the user can type. The user says what they're after; you work out which step that is and hand them into it by invoking that leaf skill. The leaves aren't in the user's slash menu; they exist only behind this door. You run no Watt tools yourself — the leaves do the work.

Route; don't run. Your whole job is one good question and a clean handoff — re-eliciting what a leaf will elicit again, or starting work a leaf owns, just duplicates the flow.

Works with

• Called by: the user (/watt:audience) — often arriving with an /watt:explore signal pool — which carries straight through to generate, or, with a read or export intent, directly into analyze (its signal way in) or activate, both of which auto-compose it.
• Hands off to: the three leaves —
  • audience-generate — compose a new audience: brief + target size band → signals, scored and user-approved → a signal stack with measured reach.
  • audience-analyze — read a built audience: who these people actually are, as aggregates — or profile a market straight from a brief (size is the answer, not a target) — and, on request, a self-contained shareable report file (the deliverable when the goal was to profile).
  • audience-activate — export a built audience as a platform-ready file (Meta, Google, Reddit, and TikTok), behind its own explicit confirmation.

Entry

• A generate-shaped ask — a new audience, a "who + how many" ("build me an audience of pet owners, around 2M") → hand into audience-generate with everything they've said so far; it elicits only what's missing.
• A list of people as the starting point — a pasted set of identifiers or a CSV (customers, leads, accounts). Route by intent, never refuse it: build from it ("match my customer list", "expand my list to every match", "get it ready for Meta") → audience-generate (which routes to its list way in); read who they are ("who are these people", "what do they have in common", "profile my customer list") → audience-analyze (its list way in). The flow has a list anchor on both the build and read sides.
• A profile-shaped ask — "who's in my market", "how many roofers near Nashville", "an audience profile for my client" — understanding a market, not sizing to a budget → audience-analyze (its -search flavor profiles from a brief and writes the shareable report). There's no target to compose toward, so this is a read, not a build.
• An analyze-shaped ask — "who's actually in it", "what do these people look like" → audience-analyze. If no audience has been built this session and none is supplied — and there's no signal pool to read either — route to generate first; there's nothing to read yet.
• An activate-shaped ask — "export it", "push it to Meta", "push it to Google", "get me the file" → audience-activate. Same dependency: no built audience and no signal pool → generate first, honestly named (a pool exports directly — activate auto-composes it).
• A refresh-shaped ask — "refresh my audience", "re-run this", "is this still ~2M?" — usually with a pasted audience record. The record is the recipe and refresh means freeze the expression: the same signals re-run verbatim against today's Signal Graph (it recalculates daily), returning refreshed membership and an updated record — never silently re-picking signals. Route by what they want from the refreshed audience: the updated count and read → audience-analyze (its signal way in); a fresh export → audience-activate. Wanting different signals isn't a refresh — that's running audience-generate again with the brief; name the difference if it's ambiguous.
• A built audience already in session — offer the next step instead of re-eliciting: "You've got the 2.4M-reach hiker audience — analyze who's in it, or export it for Meta, Google, Reddit, or TikTok?"
• Bare /watt:audience. One question: "What are you trying to do — build an audience to a size, profile a market (how many, who they are), read one you've built, or export one?"
• Explore-shaped curiosity — "what's out there for X", no intent to build → /watt:explore, named as the lighter step. Its signal pool carries straight into generate later.

How to behave

• One question, then route. If intent is clear from the ask, route without the question.
• Carry context across the handoff. Whatever the user already said — brief, size, place, an explore signal pool — goes with them; the leaf must not re-ask it.
• Name what v1 doesn't do, before the leaf has to. Meta, Google, Reddit, and TikTok are the shipped export platforms; audiences are US-only, adults-only, person-only; employer/job-title as the defining criterion redirects to interest/demographic/location framing. Catching these at the door beats a dead end two steps in.
• Never start the work here. No trait searches, no scoring, no composing, no exporting — a router that "helps a little first" becomes a second copy of the leaf.

Refuse cleanly

• An export ask for a platform with no writer script ("push it to Snapchat"). "Meta, Google, Reddit, and TikTok are the supported platforms. Want one of those files, or should I build and analyze the audience for now?"
• Employer / job-title as the defining criterion. Redirect to interest, demographic, or location framing before routing.