← Multimodal Agent Interface · Prompt Deck · Pictures · Plan

How do we make conversation the default and forms the escape hatch?

Intent Contract​

Agent autonomy boundary. Not parsed by engineering — governance instrument for humans and agents.

Story Contract​

Stories are test contracts. Each row = 1+ test file. Tests must be RED before implementation starts. Tests going GREEN = value delivered.

Build Contract​

The deliverable. Engineering builds from this. Commissioning reads this. Every row has an acceptance test. FAVV v2.1 format.

Job 1: Talk to the Agent​

Job 2: Agent Routes to the Right WorkChart​

Job 3: See Progress and Results​

Cross-Cutting​

Screen Contracts​

Screen: Chat Interface (/chat)​

Serves: S1-S5, Build Contract #1-4, #8-10

Flow and States​

Elements​

Navigation Architecture​

How does the user move from pain to effortless performance?

The Screen Contract above defines WHAT /chat contains. This section defines WHERE it lives in the app, HOW users discover it, and WHY the information architecture changes.

Pain-to-Perform Journey​

Every navigation decision maps to a stage in the user's journey from friction to flow. The UI team builds the navigation that carries the user through each stage — not just the destination screen.

Navigation States (Feature Flag Controlled)​

Cross-Cutting Navigation​

Chat is not a module — it is a layer across all modules. The navigation must reflect this.

Sidebar Hierarchy​

[Chat]          ← NEW: first item, conversation icon
─────────────── ← visual separator
Proposals       ← existing
Pipeline        ← existing
Insights        ← existing
Plans           ← existing
Agents          ← existing
───────────────
Settings        ← existing, bottom-anchored

Chat goes first because the PRD's own principle states: "The agent IS the product. Forms are the escape hatch." If chat is buried below existing sections, the navigation contradicts the intent.

Wiring Coordinates​

Principles​

What truths constrain the design?

The Job​

The hidden objection: "I'll spend more time explaining what I want to the AI than it would take to just click through the forms." The answer: context + initiative. The agent remembers your pipeline, suggests the next action, and only asks when genuinely ambiguous.

Why Now​

8 WorkCharts x 0 conversational entry = 0 agent value to users
1x1 modality coverage (text-to-text) out of 49 possible (7x7)
Vercel AI SDK (@ai-sdk/react, useChat) in stack but unused in UI
GPT-4o voice, Claude voice, Gemini Live = multimodal input feasible now
HubSpot AI, Salesforce Einstein, Clay = competitors shipping conversational CRM
Without this, drmg-sales is a form-based SaaS in an agent-first world

Design Constraints​

Performance​

How do we know it's working?

Priority Score​

PRIORITY = Pain x Demand x Edge x Trend x Conversion

North Star: Conversational task completion > 50% (tasks completed via conversation / total tasks).

Quality Targets​

Failure Budget​

Kill signal: Chat widget ships but <10% of tasks go through conversation after 30 days. Users bypass agent and navigate directly to forms.

Platform​

What do we control?

Current State​

Build Ratio​

~40% composition (Vercel AI SDK, WorkCharts, skill router, agent registry), ~60% new code (chat UI, modality router, input normaliser, session memory, streaming progress).

Protocols​

How do we coordinate?

Build Order​

Total: ~22 days T0-T4. Kill date: 2026-05-01.

Commissioning​

Agent-Facing Spec​

Commands:

pnpm nx serve stackmates        # Dev server
pnpm nx test stackmates          # Unit tests
pnpm nx e2e stackmates-e2e       # E2E tests

Boundaries: Always: UI layout, streaming strategy, model selection. Ask first: schema changes, new entity types, external API keys. Never: delete user data, send external communications, modify billing.

Players​

Who creates harmony?

Demand-Side Jobs​

Job 1: Get Answers Without Navigation​

Situation: Sales rep needs pipeline status before a meeting in 10 minutes. Currently: open app, navigate to pipeline page, scan kanban, filter by date, mentally summarise.

Features that serve this job: #1, #4, #9, #10

Job 2: Process Documents Without Copy-Paste​

Situation: RFP PDF arrives. Currently: download, open in viewer, copy-paste questions into textarea one by one, wait for AI to generate answers, manually review.

Features that serve this job: #2, #5, #6, #7, #8

Job 3: Draft Outreach Without Context-Switching​

Situation: Three prospects need personalised outreach. Currently: open Sales Dev, navigate to prospect, review profile, click generate, wait, review, repeat x3.

Features that serve this job: #1, #4, #6, #9

ICP: SMB Sales Team Without IT​

Psycho-logic: "We don't need AI" means "we tried a chatbot that couldn't find our data." The stated objection is AI skepticism. The real objection is broken promises from previous AI tools. The unlock: the agent knows your pipeline because it shares the database. First interaction must demonstrate real CRM awareness, not generic responses.

Role Definitions​

Relationship to Other PRDs​

Context​

• PRD Index — Multimodal Agent Interface
• Prompt Deck — 5-card pitch
• Pictures — Pre-flight maps
• Plan — Implementation coordinates

Questions​

What happens when the agent knows your pipeline better than you do — and acts on it?

• If S1 FORBIDDEN fires (hallucinated RFP questions), does the user lose trust in the agent permanently or just for that session?
• Which Story Contract row will be hardest to make GREEN — and does that story block the others?
• When the modality router misclassifies (S5), is the cost a wasted WorkChart run or a wrong proposal sent to a client?