FLY Phase 0+1 Roadmap — Internal Build Plan

Current State

What's LIVE Today

FLY Sales Bot v2 is deployed and operational on Railway. All P0/P1 features from the original spec are complete.

Conversation Engine

LIVE

4-node LangGraph StateGraph: classify (Haiku) → retrieve (4-layer hybrid) → generate (Sonnet 4.5 streaming) → validate (Haiku anti-hallucination)

Knowledge Base

LIVE

300 documents, 1,010 chunks, 71,819 graph edges. Semantic (50%) + Graph (35%) + Keyword (15%) retrieval with RRF fusion.

Chat Interface

LIVE

SSE streaming, markdown rendering, Mermaid diagrams, 4 artifact types (concept, testimonial, CTA, progress cards).

Feedback + Admin

LIVE

Thumbs up/down with comments, admin dashboard (filter, tag, status, notes), sources panel slide-out with graph edges.

The Problem: All 300 documents are in a single retrieval index — no tier filtering. The bot can surface Day 30 module content (deep program material) when a prospect asks "What is FLY?" Program content outweighs sales content 8:1 by word count.

March 13 Call

What LaMonte Told Us

Four critical feedback items that define the Phase 0→1 evolution.

"Not quite in the right way"

Bot uses generic AI comfort language instead of LaMonte's intentional NLP delivery style. Needs Socratic method — 5 questions for every 1 principle.

CRITICAL

"Kill the grooming language"

No "I hear you," no "That's completely valid." Questions should be one-liners without explanation. Ask → listen → ask again → THEN teach.

CRITICAL

"Bot goes too deep"

Pulls program-level content (Day 30, six human needs) for prospect questions. Sales bot should only access sales-tier content.

HIGH

"Guide the session"

Structured 3-module guided experience, not open-ended chatbot. Bot leads, user follows. Module 1 → Module 2 → Module 3 → CTA.

HIGH — PHASE 1

Confirmed Decisions

Decision	Detail
3-module scope	(1) Foundational principles → left/right side, (2) Clarity session exercise, (3) Neural addictions + internal vs external
Socratic method	5:1 question-to-principle ratio. Short questions, long user responses, deeper drilling, then integrated insight delivery
Clarity session = key artifact	Left/right side form as interactive worksheet — the core differentiator
20-min minimum	Set expectations upfront; possibly limit free sessions to 30 minutes
Content tiering	Learn to FLY = sales scope; 30 modules = post-sales; facilitator training = excluded
Video clips desired	Short clips of LaMonte teaching concepts, rendered as artifacts
Email campaigns	Personalized follow-up using interaction data for non-converters

Phase 0 · Mar 15-21

"Make It Sound Like LaMonte"

Before the follow-up call, the bot should FEEL different — Socratic, intentional, sales-scoped. Zero new features, maximum persona impact.

CP-1: Persona Overhaul

Phase 0HIGHEST

Complexity: Low Target: Mar 17 Blocked by: Nothing

Rewrite prompts.py — prompt-only change, zero code. Replace warmth-first persona with Socratic-first identity.

○ Rewrite PERSONA: "You are a questioner. You probe. You listen. You ask again."
○ Promote V3_INQUISITION to PRIMARY behavior (5:1 ratio explicit)
○ Add VOICE_ANTI_PATTERNS: banned comfort phrases with replacements
○ Reorder Response Guidelines: "Ask before you teach" = #1
○ Test 10 common prospect inputs, verify Socratic pattern
○ Deploy to Railway, notify LaMonte to test

CP-3: Vault Enrichment

Phase 0FOUNDATIONAL

Complexity: Medium Target: Mar 16 Blocked by: Nothing (running now)

Populate enrichment folders (concepts, exercises, patterns, voice, progression) with tier-tagged reference cards extracted from 300 source files. Each file tagged tier: sales|program|facilitator for retrieval filtering.

CP-2: Content Tier Filtering

Phase 0HIGH

Complexity: Medium Target: Mar 18 Blocked by: CP-3 (enrichment)

Add tier metadata filter to retrieval.py. Default queries filter to tier: sales only. Phase 1 modules can relax filter for deeper content when user is ready.

Decision: Metadata filter (Option A) — WHERE metadata->>'tier' = 'sales' clause on semantic retriever SQL. Hard boundary, not re-rank boost.

CP-4: Re-Ingest with Tier Tags

Phase 0ENABLES CP-2

Complexity: Low-Med Target: Mar 19-20 Blocked by: CP-2, CP-3

Propagate tier tags from enrichment → source frontmatter → chunk metadata → production database. Verify with SQL counts and end-to-end test query.

Phase 0 Pipeline

Phase 0 Success Criteria (by Mar 20-21 call):
✓ Bot asks 3-5 questions before teaching any principle
✓ Zero "I hear you" / "That's completely valid" / "Great question!" responses
✓ "What is FLY?" returns sales-tier content only
✓ LaMonte says "that sounds more like me"

Client Input Needed

Decisions Pending from LaMonte

4 items blocked on LaMonte's input. Mitigations in place for all — none are hard blockers to Phase 0.

CPI-1: First Live Event Date

Recommended: Proceed without. Bot already handles "no event confirmed" gracefully. When LaMonte confirms date + URL, integration is a 5-minute env var change.

The event CTA system is already built and deployed — FLY_EVENT_DATE and FLY_EVENT_URL environment variables on Railway.

CPI-2: Clarity Session Exercise Definition

Recommended: Jonathan drafts Module 1 questions from "2025 Learn to Fly" transcript (LaMonte's recommended source). Present draft at follow-up call for validation.

LaMonte confirmed 3-module structure but hasn't formally defined exact exercise sequences. Needed for Phase 1 module architecture (DP-1) and clarity form artifact (DP-2). Mitigation unblocks Phase 1 design.

CPI-3: Bot Response Style Calibration

Recommended: After CP-1 deploys, send LaMonte 5 specific test prompts with direct link. Request 20 minutes of testing with thumbs up/down feedback.

5 test prompts: "Tell me about FLY" · "I've tried therapy and it didn't work" · "How much does it cost?" · "I don't have time" · "What makes this different from mindset coaching?"

CPI-4: Client Access Model

Recommended: Option A (direct link) for now. Add QR code (D) for events. Custom domain (B) and embed widget (C) are Phase 1 polish.

Current direct URL works fine for testing. Custom domain is a Railway config change. Embed widget requires IIFE build — medium complexity, low urgency.

Phase 1 · Mar 22 – May 2

"Guided Experience MVP"

Transform from open-ended chatbot to structured 3-module guided journey with form artifacts, session management, and user profiles.

DP-1: 3-Module StateGraph Architecture

Phase 1HIGHEST

Complexity: High Target: Mar 28 Blocked by: Phase 0 + CPI-2

Restructure the linear 4-node graph into a module-aware routed architecture. Each module has its own system prompt variation, retrieval filter, and exit conditions.

Module	Scope	Style	Exit Condition
1: Foundations	NAAP principles, free/confined state, towards/away-from	Socratic questioning	User demonstrates concept understanding
2: Clarity Session	Left/right side exercise	Guided interview + form artifact	User completes left/right form
3: Neural Addictions	Why patterns repeat, internal vs external focus	Integrated insight + CTA	Readiness signals → book session

Module-Aware StateGraph

DP-2: Clarity Session Form Artifact

Phase 1HIGH

Complexity: High Target: Apr 4 Blocked by: DP-1

Interactive two-column form artifact embedded in the chat. Left column = "Away from" (fears, patterns, pain). Right column = "Towards" (goals, aspirations). Submits to POST /api/profile/clarity-session, saved to user_profiles table, referenced in Module 3 for personalized insight.

Clarity Session Data Flow

DP-3: Session Structure

Apr 7

Welcome flow with 3-module overview. 20-min commitment framing. Progress indicators via progress_card artifacts at transitions. Optional 30-min session limit.

MEDIUM

DP-4: User Profile Persistence

Apr 7

Persist Socratic responses, engagement depth, readiness signals beyond clarity form. Feeds email campaigns (DP-7). New user_profiles PostgreSQL table.

HIGH

DP-5: Video Clip Artifacts

Apr 14

New video_card artifact type. Start with YouTube embeds mapped to concept timestamps. LaMonte's face + voice builds trust faster than text.

MEDIUM

DP-6: Dashboard UI Redesign

Apr 21

Three-pane layout: modules (left), chat (center), history/artifacts (right). Deferred to late Phase 1 — current UI is functionally sufficient.

MEDIUM

DP-7: Email Campaigns

Apr 28

Personalized follow-up emails using interaction data. Trigger: X hours after incomplete session. Requires email service (SendGrid/Resend) + DP-4 profile data.

HIGH

Architecture

Dependency Chain

How Phase 0 gates unlock Phase 1 work.

CP-3 Enrichment

→

CP-2 Tier Filter

→

CP-4 Re-Ingest

→

DP-1 3-Module Graph

CP-1 Persona

→

CPI-3 Voice Calibration

→

DP-1 3-Module Graph

DP-1 3-Module

→

DP-2 Clarity Form

→

DP-4 Profiles

→

DP-7 Email

Phase 0 + Phase 1 Timeline

Phase Transition Gate: Phase 0 → Phase 1

○ Bot asks 3-5 questions before any principle (Socratic verification)
○ Zero AI comfort phrases in responses (anti-pattern verification)
○ "What is FLY?" returns sales-tier content only (tier filter verification)
○ LaMonte has tested and provided feedback (voice calibration)
○ All enrichment files have tier tags (content inventory complete)

Risk Management

Risk Register

Risk	Likelihood	Impact	Mitigation
Persona overhaul makes bot too terse / robotic	Medium	High	Test with feedback system before follow-up call. Keep V5_SALES_VOICE as warmth anchor. Iterate rapidly.
Tier filtering too aggressive — legitimate questions go unanswered	Medium	Medium	Start with tier filter + fallback: if no sales-tier results, expand to program-tier with disclaimer. Monitor via sources panel.
3-module graph breaks existing functionality	Low	High	Feature branch. Test suite (1,693 lines). Keep linear flow as fallback path.
Clarity session form UX too complex	Medium	Medium	Start with plain text inputs, not fancy interactive form. Iterate on feedback.
LaMonte doesn't test between calls	High	Medium	Send 5 specific test prompts + direct link. Set up Railway notification on feedback submission.
Module transition detection unreliable	Medium	Medium	Start with explicit transitions ("Ready for the next step?") rather than implicit detection.

Timeline

Content Tier Inventory

What the sales bot SHOULD access vs what it should NOT.

Sales Tier (75 files, ~55K words)

Three NAAP Principles · Free/Confined State · Neural Addiction (overview) · Towards vs Away-From · Three Stages of Transformation · Neuroplasticity · The Big Lie · Testimonials · Differentiation list · Social proof stats · Clarity session exercise · FLY Micro Experience

Program Tier (195 files, ~420K words) — EXCLUDE

Six Human Needs (Days 19-26) · Belief System Audit · RAS · Morning Routine · Evening Journal · Pattern Worksheet · 30-day progression · Story Rewriting · Observation exercises