Execution Blueprint — ReplyDesk
The coordinate map every other document references. ReplyDesk has two surfaces — an admin dashboard the business configures, and an embeddable widget their customers use — so the phases build accounts, then the knowledge base + key the agent needs, then the grounded agent runtime, then the embeddable widget, then insights.
P1 — Foundation & Accounts
- {P1S1} Project scaffold, CI, environments
- {P1S2} Authentication and sessions (business owner)
- {P1S3} Organization / workspace model
- {P1S4} Subscription billing (Stripe)
P2 — Knowledge Base & Keys
- {P2S1} Q&A entry data model
- {P2S2} Admin CRUD for Q&A pairs
- {P2S3} Bring-your-own AI API key (encrypted vault)
- {P2S4} Knowledge indexing for retrieval
P3 — Agent Runtime
- {P3S1} Retrieval over the Q&A (grounding)
- {P3S2} Grounded answer generation (Claude API)
- {P3S3} Streaming responses to the widget
- {P3S4} Guardrails + escalation to a human
P4 — Embed & Widget
- {P4S1} Embeddable chat widget (one-line script)
- {P4S2} Widget theming and branding
- {P4S3} Conversation logging
- {P4S4} Visitor contact capture / handoff
P5 — Insights & Launch
- {P5S1} Analytics: questions, deflection rate, gaps
- {P5S2} Settings and team access
- {P5S3} Launch hardening: rate limits, abuse controls, observability