Objectives
- Build the
/coach/sessions
workspace where students engage with their selected AI coaches in real time. - Integrate the Coach Reasoner, Knowledge Retrieval, and Action Item Planner agents to produce tailored guidance.
- Capture transcripts, decisions, and session summaries in a durable timeline that compounds across meetings.
- Finalize the action item handoff to TickTick lists with student confirmation and metadata tagging.
Functional Scope
Session Workspace UX
- Provide a session lobby where users pick a wealth area (pre-filled based on recent activity) and AI coach combination.
- Implement a chat-style interface supporting streaming AI responses, markdown rendering, and attachments (e.g., reference notes).
- Display contextual sidebars: selected AI coaches, relevant knowledge snippets, prior commitments, and upcoming deadlines.
- Offer quick actions to inject favorite blog posts or knowledge cards into the conversation.
- Allow multi-AI-coach mode where insights from multiple personas appear as separate threads before being synthesized.
Knowledge Retrieval & Reasoning
- Build retrieval pipeline that fetches embeddings based on wealth area + AI coach combination with fallback heuristics.
- Parameterize Coach Reasoner prompts using persona data, student profile, and prior session history.
- Support model routing (cheap vs premium) with guardrails to stay within token budgets.
- Log prompts/responses with redaction of personal data for debugging.
Action Item Planning & Confirmation
- Generate structured tasks from the conversation, including title, description, due date, and tags.
- Present tasks to the student for review, edit, and approval before pushing to TickTick.
- Persist approved tasks to
action_items
table with references to session ID and TickTick list. - Provide bulk confirmation UI so multiple tasks can be accepted or deferred in a single action.
- Handle offline scenarios by queueing tasks for later sync and notifying the student.
Session Summaries & Memory
- Summarize key takeaways, commitments, and AI coach feedback at the end of each session using the Session Summarizer agent.
- Append summary to
session_summaries
with ability to tag wealth areas, highlight blockers, and link to favorite posts. - Expose timeline view within
/coach/history
with filters by wealth area, AI coach persona, or completion status. - Enable export/share functionality (PDF or markdown) for students and optional email digest to administrators or human supporters.
Technical Considerations
- Use streaming server-sent events or WebSockets to deliver AI responses in near real time.
- Implement optimistic UI for action item edits while ensuring eventual consistency with TickTick updates.
- Cache knowledge snippets for the duration of a session to avoid repeated embedding lookups.
- Add token usage metering per session and throttle or warn if budgets are exceeded.
- Record telemetry for latency (prompt-to-response), task creation time, and student approvals.
Multi-Agent Workstream
| Agent | Responsibilities | Deliverables | | --- | --- | --- | | AI Coach Reasoner | Generate tailored advice using persona prompts and retrieved knowledge. | Prompt templates, response evaluator, tests. | | Knowledge Curator | Maintain embeddings, retrieval ranking, and fallback heuristics. | Embedding jobs, retrieval API, monitoring dashboards. | | Session Summarizer | Craft structured summaries and update history timeline. | Summary prompts, storage adapters, UI integration. | | Action Item Planner | Transform guidance into tasks, manage student approvals, and sync to TickTick. | Task schema, approval UI, sync hooks. | | Frontend Lead | Build session workspace UI, streaming chat, and context sidebars. | React components, state management, UX polish. | | Analytics Lead | Instrument session metrics, token usage, and task conversion rates. | Event taxonomy, dashboards, alert thresholds. |
Exit Criteria
- Students can start a session, converse with AI coaches, and see contextual knowledge alongside the chat.
- Session transcripts, decisions, and summaries persist and display in the history timeline.
- Approved action items sync to TickTick with appropriate tags and appear in the student dashboard.
- Multi-AI-coach mode successfully synthesizes advice from more than one persona without UI degradation.
- Observability dashboards report prompt latency, task approval rate, and summary completion per session.
Risks & Mitigations
| Risk | Mitigation | | --- | --- | | AI hallucinations or misaligned advice. | Ground prompts in retrieved knowledge, add guardrail checks, allow human-in-the-loop override. | | Latency impacting conversation flow. | Preload embeddings, use streaming responses, and degrade gracefully with “processing” indicators. | | Task overload for students. | Require explicit approval, allow snooze/decline, and cap default number of tasks per session. | | Summary quality variance. | Evaluate summaries with automated QA, provide edit interface, and enable follow-up iterations. |
Dependencies & Notes
- Requires Phase 2 data: AI coach assignments, wealth areas, TickTick mappings, and MCP toolset.
- Coordinate with security to ensure transcripts and summaries comply with privacy policies (PII redaction, retention limits).
- Provide API endpoints for future mobile clients to start sessions and fetch history.
- Share telemetry schema with Phase 4 team to leverage for accountability dashboards.