| name | configure |
| description | Set up Watt in Claude, or fix it when it stops working. A guided check that confirms you're on a supported surface, verifies the Signal Graph connection and file downloads for you, and walks you through turning on automatic updates — fixing anything that's off, one step at a time. Run it the first time you install Watt, or any time something's not connecting — "set up Watt", "where do I start", "connect Watt", "Watt isn't working", "my export is blocked", "the connector's grayed out" — or type /watt:configure. |
| compatibility | Designed for Claude Cowork, Claude Code, or Agent SDK. Requires Signal Graph. |
Configure Watt
Purpose
/watt:configure is the user's setup and connectivity aide. It walks a new user through configuring Watt the first time — and it's the same place a returning user comes back to when something stops connecting. It settles their account type to tailor the path, checks the four things Watt needs, verifies the two it can verify, fixes whatever's off one step at a time, and leaves them able to use Watt.
Every beat ends able to use Watt — or holding the one exact next step to get there. When a check fails, hand over the precise move (the click, the admin link, or /watt:help) and wait on the user's signal; never retry blindly or guess past a step that isn't done — that just loops on a connection that isn't ready.
Works with
- Called by: the user (
/watt:configure), or any skill that hits a connection or authentication wall — this is the connect/recovery door the rest of the plugin points to.
- Hands off to:
/watt:audience, /watt:explore, and /watt:help — via the Interactive-input choices that close the run (and /watt:help earlier, when a check stays stuck and they want a human).
The flow
One beat per turn. Throughout, keep a setup checklist on screen and mark each item with a ✓ the moment it clears — so the user always sees where they are and what's left:
- Account type
- Supported platform
- Signal Graph connected
- File downloads on
- Automatic updates on
Show the checklist however reads clearest on this surface — the model and host decide the shape. Decisions, though, are Interactive inputs — Claude's native clickable options — robust even when a render surface is part of what's misconfigured, and a typed reply always wins.
One path, not all of them. Steps 5–7 each have a personal-account path and an organization-account path. Step 2 settles which one is theirs — so from there on, show only that path. Never paste both branches, every doc link, and every what-if into the chat; that's noise. Give the user the one move that's theirs right now, plus a single link to the guide if they want to follow along — and let the tappable choices funnel them.
1 — Introduce, and show the checklist
Open with a genuine, upbeat welcome, then say plainly what configure is: a quick guided check that gets Watt set up and working in Claude — and the place to come back to any time something stops connecting. Remind them what Watt is, briefly, in its own framing: it indexes the world's signals and makes them consumable — especially strong at turning a plain-English description of people into a real, measured audience.
Then point them at the right surface and model, since it saves a dead end: Watt runs in Claude Cowork, Claude Code, or the Agent SDK, not the Claude Chat app. Cowork is the easiest place to run it — it handles Watt's longer exports and the working files it produces — and Opus is the model to use, since its larger context and sharper reasoning earn their price on the signal engineering that drives audience quality.
Render the checklist (nothing ticked yet) so they see the shape of what's ahead. Make no tool calls yet — the checks start once we know their account type and platform.
2 — Personal or organization account?
This shapes the rest of setup: an organization account hands some settings to an Owner, a personal one keeps them all in the user's own hands. Settling it now lets every later step show only the path that's actually theirs.
Most people don't know their account type off-hand, so tell them where to look: click your account at the bottom-left of Claude — under your email it shows your plan. Then map it for them: Pro or Max → personal account; Team or Enterprise → part of an organization. (Screenshots in the guide: https://wattdata.ai/docs/integrate/claude)
Offer the answer as Interactive inputs — they pick one; the plan under their email tells them which, so don't add an "unsure" escape:
- Personal (Pro/Max) → tick Account type ✓; steps 5–7 show the personal path only.
- Organization (Team/Enterprise) → tick Account type ✓; steps 5–7 show the organization path, where an Owner sets the org-wide pieces.
Make no tool calls here — it's a single question. Move to step 3.
3 — Confirm the platform
Read the cues available to you and infer the most likely surface, then always confirm it — a wrong guess sends the rest of setup down the wrong path. Pre-state your read ("Looks like you're in Claude Code — is that right?"); if nothing points clearly, don't force a guess — just ask. Offer the choice as Interactive inputs: Claude Chat, Claude Cowork, Claude Code, Agent SDK, Something else.
- Claude Cowork, Claude Code, or Agent SDK → supported; tick Supported platform ✓ and move to step 4.
- Claude Chat or Something else → tell them plainly Watt supports only Claude Cowork, Claude Code, and the Agent SDK — in the Chat app only these skill files load, not the connector, hooks, and scripts the rest of Watt needs. They'll need to switch and re-run
/watt:configure. Don't try to connect from here — stop on that note.
4 — Show what's installed (Cowork only)
Before the connection checks, show the user their Watt setup at a glance — it reassures them the pieces are in place, and it tells apart the two failure shapes the next step can't distinguish on its own: a connector that was never installed vs. one that's installed but not connecting.
This step is Cowork-only — the listing tools ride Cowork's connector- and plugin-management and aren't present on Claude Code or the Agent SDK. If they aren't available, skip this step silently and go straight to step 5.
Call Cowork's two listing tools — list_connectors (installed connectors) and list_plugins (installed plugins) — and read just two things from what comes back: whether the Watt plugin is installed (its entry also lists the skills it loaded), and whether Signal Graph is installed. If you can't find those tools on this surface, that's exactly the silent-skip case above — move to step 5. Show a small presence panel with the two lines below, however reads clearest on this surface:
- Watt plugin — installed ✓ (with its skills)
- Signal Graph — installed ✓ (or: not found)
Be honest about what this shows: presence, not health. The list reports that a connector is set up, not that it's authenticated or working right now — a connector can list as installed and still fail to connect. So don't claim Signal Graph is connected from this list, and don't tick a checklist item here — the connection is proven by the probe in step 5, never by this panel. This is orientation; it advances nothing on its own. Read the connector result forward into step 5, then move on:
- Signal Graph installed → carry that in; step 5 runs its probe as the real test.
- Signal Graph not found → it was never installed, so a probe would only fail — take them straight to step 5's connect path, skipping the probe.
Make no decision here — it's a display beat: render the panel, end the turn, and open step 5 on the next one (don't stack the panel and the connection probe in a single turn). Move to step 5.
5 — Check the Signal Graph connection
This check verifies itself. If step 4 just showed Signal Graph not installed, skip the probe — it would only fail — and go straight to the connect path below. Otherwise (and on any surface where step 4 didn't run), narrate it in a warm line, then run a single trait_search (e.g. "data") as a liveness probe.
-
Comes back with results → the connector's live; tick Signal Graph connected ✓ and move to step 6.
-
Fails on a connection or authentication problem → this is exactly what configure is here to fix, so don't bounce it elsewhere — show how to connect. Assume the person isn't technical and isn't a Claude expert; name exactly what to click one move at a time, and hand clean clickable links without explaining how a link works. Show only the path for their account type:
Want the full walkthrough with screenshots? One link: https://wattdata.ai/docs/integrate/claude
End on a tappable confirmation, not an open-ended wait — you can't detect the connection yourself, and probing mid-setup just fails and tempts a retry loop. Offer as Interactive inputs:
- Connected → re-run the
trait_search probe to confirm. Passes → tick ✓ and move to step 6. Still failing → say so plainly and put them back at the instructions above with the same choices.
- Install button is grayed out (organization only) → their org Owner has to approve Watt before they can continue; give them the org-configuration link above to send to the Owner, and offer to re-check once it's approved.
- Contact support → hand to
/watt:help to reach a human, carrying the symptom (the connector won't connect, and what they've tried) so help opens with a human-help draft ready instead of a blank ask.
6 — Check that exports can download
This check verifies itself too. Watt's exports come back as download links from a cloud-storage domain, and reaching that domain is a Claude setting — so the honest test is to fetch one. Narrate it, then run a tiny entity_find (entity type person, the age-30 signal 648c0733c3c2b995d8f4e99a3f8ba2a4, audience_limit 10, format csv, domains ["name"]) to mint a download link, and try to fetch that link. That fetch rides Claude's network egress and domain allowlist — the entity_find call itself rides the connector and succeeds regardless — so whether the file comes back is the real test.
Don't mistake a tool limit for a blocked download. These download links are long, and an over-length URL can make Claude's default fetch tool fail for reasons that have nothing to do with the user's settings. So if that first fetch fails, don't conclude anything yet — retry it from the code sandbox (a curl or wget in bash, or a tiny Python/JS fetch), where the URL-length limit doesn't apply. Only when the sandbox fetch also fails is the download genuinely blocked. (If you can't attempt a fetch at all on this surface, don't assert a result — ask the user to open the link and tell you whether it downloads, and treat their answer as the check.)
-
The file comes back (on either fetch) → downloads work; tick File downloads on ✓ and move to step 7.
-
Both the default fetch and the sandbox fetch fail → that's the allowlist or egress setting, not the connection — show how to turn it on. Same non-technical care as step 5. Show only the path for their account type:
Full walkthrough with screenshots, one link: personal → https://wattdata.ai/docs/integrate/claude/personal-account; organization → https://wattdata.ai/docs/integrate/claude/org-user.
End on a tappable confirmation. Offer as Interactive inputs:
- Enabled it → re-run the download probe (default fetch, then the sandbox fallback). Comes back → tick ✓ and move to step 7. Still blocked → say so and put them back at the instructions.
- The option isn't available (organization only) → these settings are org-wide and their Owner hasn't enabled them yet; give them the org-configuration link (https://wattdata.ai/docs/integrate/claude/org-configuration) to send to the Owner, and offer to re-check once it's done.
- Contact support → hand to
/watt:help to reach a human, carrying the symptom (exports won't download, and what they've tried).
7 — Turn on automatic updates
This one can't be verified — Claude exposes no way to read the setting — so it's a guided step the user confirms. State plainly why it matters: Claude never updates the plugin on its own. Until this is on, they stay on the version they installed — they won't get new features or fixes as we ship them. Everyone should turn it on. Show only the path for their account type:
- Personal account → Open Customize → Plugins → Browse plugins (https://claude.ai/customize/plugins?browse), find Watt on the Personal tab, click the ⋯ menu next to it and toggle on Sync automatically, then click Check for updates once — syncing won't start on its own until you do.
- Organization account → If your organization installed Watt for you, it shows under the Your organization tab and stays current automatically — nothing for you to do. If you installed it yourself under the Personal tab, follow the personal steps above.
End on a tappable confirmation. Offer as Interactive inputs:
- Turned it on → tick Automatic updates on ✓ and move to step 8.
- My organization manages it → the org's installation keeps it current automatically — nothing for them to do. Tick ✓ and move on.
- Contact support → hand to
/watt:help to reach a human.
8 — Done — show the full checklist, then point the way
Render the checklist with all five items ticked ✓. Tell them Watt is configured and ready, and that they can return to /watt:configure any time something stops connecting.
One tip worth more than it looks:
- Describe the why, not the how. Tell Watt who you're trying to reach and why they buy — what they care about — not a spec like "30–45, $200k+ HHI, in California." That's just list-pulling. Give Watt the why and it figures out the how better than a person can: sell to bar owners — hard to find, not on LinkedIn — and Signal Graph knows who's buying bar POS systems, plus fifty other angles on them.
Then set them loose — the handoff happens right here in the chat. Offer the three ways in as Interactive inputs; when they pick one, invoke that skill's command and let it take over. Don't fire a canned task of your own — the skills are guided and will ask what they actually want. The three:
- Build an audience →
/watt:audience — describe who you want to reach; Watt engineers the signals into an audience, then you analyze who's in it or activate it on a platform.
- Explore what's out there →
/watt:explore — see what Signal Graph holds right now; the fastest way to size up an idea.
- Get help →
/watt:help — ask how something works, request a signal or feature, or reach our team.
How to behave
- Verify what you can; never claim a setting you can't see. The connection and the download check verify themselves — run them and report what actually happened. Automatic updates can't be read, so it's confirmed by the user, never asserted as done.
- The installed-list is presence, not health. The Cowork listing tools (step 4) report what's set up, not what's authenticated or working — a connector can list as installed and still fail to connect, and the list carries no version or auto-update state. So the panel orients the user and steers step 5's path; it never ticks a checklist item, never stands in for the probe, and never asserts the connection is live.
- The download check needs both attempts before it fails. A blocked download and a fetch the tool simply couldn't make are not the same thing. The default fetch tool can choke on an over-length download URL — so a first-fetch failure is never the verdict; retry from the code sandbox, and only call downloads blocked when that fails too.
- Show one path, not all of them. Step 2 settles personal vs. organization (a required pick), so every later step shows only that account type's path — one move, one link. Don't paste both branches and every what-if into the chat; let the tappable choices funnel.
- A connection or authentication failure is configure's job, not a transient. This is the one skill that owns the connect path, so work the connector beat — don't loop the probe, don't go diagnosing the connector or the MCP registry. If a later check fails on connection/auth rather than its own setting, it's the connector again: route back to step 5. But a pre-call notice ("requires authentication" / "still connecting") isn't a failure — the step-5 probe settles it; only a probe that also fails is the real thing.
- Never push past a step that isn't done. Move on a user's pick, not a guess; re-run the self-verifying probe after they say they fixed it, and say so plainly if it still fails.
- Configure never files a ticket itself. Reaching a human is
/watt:help's lane — hand off with the symptom in tow; don't call the support channel from here.
