Signal Work Management

Every growing team has a person who is everyone's first stop—not because they're the most senior, but because they're the most connected. The PM whose DMs are half other people's blockers. The ops lead who routes everything before it has a real owner. The team lead who has 40 unread messages at 9 a.m., knows maybe six of them need action today, and can't find the six without reading all 40. They're not behind because they're slow. They're behind because the work is invisible.

The problem isn't volume. It's legibility. A critical customer escalation and a low-priority docs request look identical in an email inbox. The thing that's been waiting a week looks the same as the thing that just came in. You feel behind but you don't know what you're actually behind on—because the work isn't visible as a coherent stack. It's scattered across six channels, with no shared sense of urgency, ownership, or record of what's already been decided.

Signal is the surface that makes the inbound stack visible. It collapses work from Slack, email, calendar, docs, forms, and chat into a single feed—scoped by team workspace, grouped by urgency, and triaged item by item. When you need a second opinion on how to act, Triage Assist sends the full signal context to Claude and comes back with a structured recommendation.

Teams managing multi-channel inbound typically hit the same set of structural problems:

• Everything looks equally loud — a critical enterprise renewal and a low-priority docs request compete for the same attention in an undifferentiated inbox. The urgent item doesn't announce itself.
• Source context gets lost in handoffs — knowing that a request came from a Slack DM versus a formal email versus a calendar hold matters for how you respond and who should own it. That context evaporates by the second forward.
• Triage happens mentally, not visibly — most teams triage in their heads, continuously, as new messages arrive. There's no record of "we decided to defer this" or "this is assigned to Jordan"—just an unspoken understanding that degrades over time.
• Cross-team work collides — legal requests, product decisions, design reviews, and ops tasks don't belong in the same undifferentiated list. Mixing them creates noise for everyone.
• Mistakes are hard to reverse — marking something as ignored or resolved is a lightweight action in practice, but it needs a recovery path. Triage actions that can't be undone create hesitation.

Three decisions anchored the design before any screens were drawn:

• Triage Assist — Claude API — a sparkle button in the detail panel sends the full signal context (title, urgency, workspace, why it matters, sources, related inputs) to Claude. It returns a structured recommendation: suggested urgency, recommended owner, and a 2-sentence action plan. The suggested owner pre-fills the assign dropdown so one click commits it. The result is advisory—you review and decide.
• Workspace scoping — five team workspaces (Product, Legal, Design, Operations, Sales) scope the feed so different teams don't compete in the same view. Switching workspaces is a single click; the feed, filters, and empty states all update to match the new scope.
• Urgency-grouped feed — signals are grouped by urgency (Critical → High → Medium → Low) so the stack sorts itself. You don't have to mentally re-sort a flat list. The critical items are at the top by definition.
• Triage state lifecycle — every signal moves through states: Needs Triage → Assigned, Deferred, Ignored, or Resolved. The triage view filters the feed to exactly one state at a time. "Needs Triage" is the default—the unprocessed inbound.
• Source filters — signals carry the channel they arrived from (Slack, email, calendar, doc, chat, form). The sidebar lets you toggle sources on or off so you can focus on a specific channel or silence the ones you're not managing right now.
• Detail panel — selecting a signal opens a focused panel showing the full context: why it matters, related inputs (the actual message snippets and their sources), assignee, and triage actions. The feed stays visible alongside it.
• Dark mode — system-aware dark mode with localStorage persistence and an anti-flash inline script in the document head. Toggled from the user bar in the sidebar.
• Undo via Cmd+Z — triage actions (ignore, defer, resolve, return to triage) are reversible. Cmd+Z restores the previous state without a confirm dialog. One level of undo; fast and honest about its limits.
• Import / export — the workspace can be exported as JSON and re-imported to restore a session. The import validates schema shape before overwriting anything.
• Local persistence with seed reset — state persists to localStorage between sessions. A reset button restores the original seed data, useful for demos and for recovering from a completely triaged queue.

• Next.js 16 (App Router) — client workspace component manages all signal state; the page shell is server-rendered for fast initial load. API routes handle the Claude integration server-side so the key never touches the client.
• Claude API (Anthropic) — Triage Assist calls claude-haiku-4-5 via a Next.js API route. The system prompt defines the team roster and their roles; Claude returns a typed JSON object (urgency, suggested_owner, recommended_action) that maps directly onto the UI.
• Tailwind CSS — utility-first styling with a three-level surface hierarchy (sidebar, background, card) defined in OKLCH color tokens. Full dark mode via .dark class on <html> with localStorage persistence.
• TypeScript — strict types throughout; domain types (Urgency, TriageState, SourceType, Workspace) flow from a single types.ts. Runtime type guards on import validation and Claude response parsing prevent corrupt data from silently coercing into the wrong shape.
• localStorage persistence — signals persist between sessions with schema versioning (signal-signals-v5). The service layer handles hydration, patching, and reset to seed without the signals array ever touching the component directly.
• Seed data — 30 signals across all five workspaces built around Vela, a fictional B2B SaaS company. Signals span enterprise renewals blocked on legal review (Meridian Health), mobile crash spikes, onboarding drop-offs, vendor NDAs, and API partner escalations. Mixed triage states and urgency levels so every view has something to show.

• Empty states had to explain the lens. An empty feed without context looks broken. Once workspace, triage view, source filters, and search could all combine to produce zero results, the empty state needed to tell you why: "No critical signals need triage in the Product workspace." Without that specificity, users would assume the app had failed, not that they'd triaged everything.
• Undo became essential once actions felt real. Early versions had no undo. Marking something as "ignored" felt risky — what if you mis-clicked? Adding Cmd+Z reduced hesitation on every triage action. One level is enough: you almost never need to undo two steps back, but you frequently need to undo the last one.
• The mobile sheet needed its own design. On desktop, the detail panel sits alongside the feed. On mobile, it opens as a full-height sheet — which created a problem: the sheet felt like it replaced the context you were just in rather than extending it. The fix was visual: the sheet slides up rather than replacing, and the close button returns you to exactly where you were in the feed.
• Visual hierarchy beat extra chrome. Early card designs had urgency color blocks, source icons, and badges competing for attention. Stripping back to a quiet card with strong typographic hierarchy let the why_it_matters text carry the weight. The data is the design — decoration was getting in its way.
• Schema validation on import caught silent failures. The first version of import just parsed JSON and loaded it. A malformed file would partially load or silently coerce wrong types into the state. Adding schema validation before any state mutation meant the worst case was a clear error message, not corrupted data with confusing behavior downstream.
• Claude for Desktop hijacked the API key. Claude for Desktop injects ANTHROPIC_API_KEY as an empty string into every shell it spawns, which silently overrides .env.local values. The fix was renaming the variable to SIGNAL_ANTHROPIC_KEY to avoid the collision — a good reminder that environment variable naming is part of the integration surface.

• Signal is a complete, interactive triage tool — not a concept. Every feature is wired: workspace scoping narrows the feed, triage actions update state and persist, undo works, import validates before loading, and the empty states explain exactly what lens produced them.
• The Claude integration shows where AI earns its place in a workflow tool: not replacing the decision, but reducing the cold-start cost of making it. Triage Assist reads the same context a human would—title, stakes, sources, related snippets—and returns a structured first opinion. You stay in control; it just gets you to a starting point faster.
• The design question Signal is answering is real: how do you make multi-channel inbound legible without flattening it into a generic task list? The answer is structural — urgency as the primary sort axis, workspace as the scope, triage state as the decision record, and source as the filter mode. None of those are decorative additions; each one does load-bearing work.
• Code lives here: github.com/meganleclair/signal-work-management.