diff --git a/docs/superpowers/specs/2026-07-11-phase-3-ai-dashboard-design.md b/docs/superpowers/specs/2026-07-11-phase-3-ai-dashboard-design.md
new file mode 100644
index 0000000..380a780
--- /dev/null
+++ b/docs/superpowers/specs/2026-07-11-phase-3-ai-dashboard-design.md
@@ -0,0 +1,175 @@
+# Phase 3 — AI Features & Dashboard Upgrade
+**Date:** 2026-07-11
+**Status:** Approved
+
+---
+
+## Scope
+
+Phase 3 adds AI-assisted capabilities via the Claude API and upgrades all role dashboards from placeholders to functional views. Similar-incident retrieval (pgvector) is deferred to Phase 4.
+
+**Deliverables:**
+1. Upgraded dashboards for all 5 roles
+2. Report quality check (AI)
+3. Severity/category suggestion at triage (AI)
+4. RCA/CAPA drafting assistant (AI)
+
+---
+
+## 1. Dashboard Upgrades
+
+### Approach
+Replace placeholder home pages for each role with data-driven dashboards scoped to what that role needs to act on. All dashboards are server-rendered (Next.js App Router, `force-dynamic`). Site/zone heatmap is a plain HTML table with Tailwind bg-opacity intensity — no canvas library.
+
+### Role Dashboards
+
+**HSE Officer** — `/hse/dashboard` (upgrade existing)
+- Leading vs lagging split: near_miss + hazard = leading indicators; injury + lti = lagging
+- Site/zone heatmap: incident count per site × zone, color-coded by intensity
+- CAPA on-time completion rate: `(verified CAPAs where completed_at <= due_date) / total verified * 100`
+- Top 3 incident categories by count (current month)
+- Top 3 root causes from `investigations.root_cause_summary` (simple text frequency, no NLP)
+- DOSH-reportable filing status: counts of `dosh_reports` rows by status (not_required / pending / submitted)
+- CSV export endpoint: `GET /api/dashboard/export?role=hse` — all incidents with key fields
+
+**Management** — `/management` (upgrade placeholder)
+- This month vs last month incident count (delta + arrow)
+- LTI count (incidents where `medical_status = 'lti'`)
+- Severity distribution bar (count per severity 1-5)
+- Leading/lagging split (same calc as HSE)
+- CAPA overdue count (status = 'overdue')
+- Site comparison table: incident count per site, sorted descending
+- CSV export: `GET /api/dashboard/export?role=management`
+
+**Supervisor** — `/supervisor` (upgrade placeholder)
+- Scoped to `user.site_id` — only their site's data
+- Open incidents count + list (status != closed), linked to incident detail
+- CAPA status summary for their site: open / in_progress / overdue counts
+- Overdue CAPAs: list with owner name + due date
+
+**CAPA Owner** — `/capa-owner` (upgrade placeholder)
+- Scoped to `capa_actions.owner_user_id = current_user`
+- My open CAPAs: list with incident ref, description, due date, status
+- Overdue count highlighted in red
+- No export needed
+
+**Reporter** — `/reporter` (upgrade placeholder)
+- Scoped to `incidents.reported_by = current_user`
+- My submissions: list with reference_no, date, type, current status
+- Status shown as colored badge (reported / triaged / investigating / capa_pending / verification / closed)
+- No export needed
+
+---
+
+## 2. AI Features
+
+### Architecture
+- All Claude API calls: server-side only, `/app/api/...` routes
+- Model: `claude-sonnet-4-6`
+- API key: `ANTHROPIC_API_KEY` env var (never client-side)
+- Every AI suggestion logged to `audit_log` with: table=`ai_suggestion`, action=`suggested`, old_value=null, new_value=JSON of suggestion, changed_by=current user
+
+### 2a. Report Quality Check
+
+**Trigger:** Incident report form submit — before DB write, client POSTs to quality-check route first.
+
+**API route:** `POST /api/incidents/quality-check`
+
+**Request payload:**
+```json
+{
+  "description": "string",
+  "incident_type": "injury | near_miss | ...",
+  "injury_involved": true,
+  "evidence_count": 0
+}
+```
+
+**Claude prompt strategy:** Send incident fields; ask Claude to return a JSON array of short warning strings covering: vague description (< 20 words), injury with no photo, no location detail, missing witness info for serious incidents.
+
+**Response:**
+```json
+{ "warnings": ["Description too vague — add what happened and where", "Injury reported but no photo attached"] }
+```
+
+**UI:** If warnings exist, show yellow warning panel above the submit button listing each warning. "Submit anyway" button still present — user is never blocked, only informed.
+
+**Audit log:** Not logged (no human decision to record — warnings are informational only).
+
+### 2b. Severity/Category Suggestion (Triage)
+
+**Trigger:** "Get AI suggestion" button in existing triage panel at `/hse/incidents/[id]/triage`.
+
+**API route:** `POST /api/incidents/[id]/ai-triage`
+
+**Claude prompt strategy:** Send description, incident_type, injury_involved, lost_days. Ask Claude to return suggested severity 1-5 with confidence (high/medium/low) and one-sentence reasoning. Return as JSON.
+
+**Response:**
+```json
+{
+  "suggested_severity": 3,
+  "confidence": "medium",
+  "reasoning": "Injury involved with medical treatment but no LTI reported — consistent with severity 3."
+}
+```
+
+**UI:** Suggestion card appears below the severity input in the triage form. "Accept" pre-fills the severity field. HSE can ignore and type their own value. Card shows confidence label and reasoning text.
+
+**Audit log:** On triage form save, log: suggested_severity, confidence, human's final severity value. `record_id` = incident id.
+
+### 2c. RCA/CAPA Drafting Assistant
+
+**Trigger:** "Generate suggestions" button in investigation workspace at `/hse/incidents/[id]/investigation`.
+
+**API route:** `POST /api/incidents/[id]/ai-rca`
+
+**Claude prompt strategy:** Send incident description, investigation findings_text, root_cause_summary. Ask Claude to return: 2-3 root cause categories (from standard EHS taxonomy: Human Factors, Equipment Failure, Procedure, Environment, Management System) + 2-3 draft CAPA action descriptions as JSON array.
+
+**Response:**
+```json
+{
+  "root_causes": [
+    { "category": "Procedure", "explanation": "No documented SOP for forklift loading at Dock A" }
+  ],
+  "capa_drafts": [
+    "Develop and implement SOP for forklift loading operations at all dock areas within 14 days",
+    "Conduct refresher training for all forklift operators on LOTO procedure by end of month"
+  ]
+}
+```
+
+**UI:** Collapsible "AI Suggestions" panel below the RCA form fields. Each root cause has a "Use this" button that appends text to `root_cause_summary`. Each CAPA draft has a "Create CAPA" button that opens the CAPA creation form pre-filled with that description.
+
+**Audit log:** Log on button click: which suggestion index was used, full suggestions array as old_value context. `record_id` = incident id.
+
+---
+
+## 3. Data & Schema
+
+No new DB tables required. Existing `audit_log` table handles AI suggestion logging via `action = 'ai_suggested'`.
+
+CSV export routes query existing tables — no new schema.
+
+---
+
+## 4. Implementation Order
+
+1. Dashboard — HSE (upgrade existing, most complex)
+2. Dashboard — Management
+3. Dashboard — Supervisor
+4. Dashboard — CAPA Owner
+5. Dashboard — Reporter
+6. CSV export endpoint (shared, covers HSE + management)
+7. Report quality check (AI)
+8. Severity suggestion at triage (AI)
+9. RCA/CAPA drafting assistant (AI)
+
+---
+
+## 5. Out of Scope (Phase 3)
+
+- Similar-incident retrieval via pgvector — Phase 4
+- WhatsApp notifications — Phase 4
+- Multi-language UI — Phase 4
+- CAPA effectiveness re-check automation — Phase 4
+- Risk heatmap / predictive analytics — Phase 4