Agentic Boarding System — Architecture PRD

01

First Principles & Second-Order Effects

▼

What Are We Actually Building?

▼

An automated pipeline that takes a subject-matter expert's raw intellectual property (transcripts, documents, frameworks, brand assets) and converts it into a fully functioning AI coaching Operating System — complete with system prompts, knowledge files, tool configurations, lead magnets, onboarding flows, and quality rubrics.

The pipeline is modular, skill-based, and orchestrated by a Chief of Staff agent that manages a kanban board. Each step has explicit quality gates to prevent "AI collapse" — the gradual degradation of fidelity to the expert's actual thinking patterns when an LLM generates without grounding.

Core Insight

The value of a coaching OS is NOT the chat interface. It's the structured frameworks + ethical governance + persistent memory + expert voice fidelity that make the AI act as the expert's digital twin, not a generic assistant. Every extraction step must preserve these qualities.

The "Clone" Pattern

Athio builds "cognitive infrastructure" — AI that thinks like a specific expert. The boarding system is the factory that produces these clones. Each clone follows the MasteryOS architecture: expert IP nested inside Athio's technology stack.

First Principles

▼

Fidelity over speed. A clone that doesn't sound like the expert is worse than no clone. Extraction quality is the primary constraint, not pipeline velocity.
Governance is non-negotiable. Every expert has ethical boundaries, anti-patterns, and "Laws of Babylon" they explicitly resist. These must be extracted FIRST and applied as constraints on all downstream generation.
Rubrics prevent collapse. Every generated artifact (system prompt, knowledge file, tool config) must be scored against extracted patterns before advancing. No artifact moves from "Build" to "Review" without passing its rubric.
Voice is separate from soul. An expert's thinking patterns (soul) and speaking patterns (voice) are distinct extraction targets. Both must be captured independently and merged at compile time.
Frameworks are the product. Tools, sequences, taxonomies, and dependencies are what users actually run. The chat wrapper is delivery; the frameworks are the value.
Lazy-load, not eager-load. The orchestrator should only activate skills when needed. Skills remain dormant until their kanban card is claimed.
State is serializable. The entire boarding process for an expert must be captured in a single JSON kanban board that can be paused, resumed, audited, and transferred between sessions.

Second-Order Effects

▼

Desired (Explicit)

New partners can be onboarded in days, not months
Quality floor is enforced by rubrics, preventing the "bad clone" problem
LLM sessions remain productive across context boundaries via serialized state
Each boarding produces reusable artifacts (lead magnets, onboarding flows) alongside the core OS

Desired (Implied)

The pipeline itself becomes Athio's moat — competitors can build a chatbot, but not a quality-gated extraction-to-clone factory
Rubric data accumulates across partners, improving extraction heuristics over time
The kanban system doubles as an audit trail for partner reporting

Risks to Monitor

Over-extraction: Pulling patterns that aren't real from insufficient source material
Voice flattening: Reducing a unique voice to generic "coaching speak"
Rubric gaming: Generating artifacts that score well but miss the expert's actual intent
Pipeline rigidity: Building so much process that small partners feel over-served

02

Complete Architecture

▼

Phase 1: CLI Skills + File-Based State (Now)

▼

The immediate implementation uses Claude CLI skills and file-based state. This is the proven pattern from the design-system-extractor skill already built.

Component	Location	Purpose
Skills	`~/.claude/skills/{skill-name}/SKILL.md`	14 specialized recon/extraction/build skills
Skill References	`~/.claude/skills/{skill-name}/references/`	Supporting docs, templates, rubric definitions
Expert Workspace	`{project-root}/_workspaces/{expert-slug}/`	All extraction outputs per expert
Kanban Board	`{workspace}/kanban.json`	Board state per expert
Extraction Outputs	`{workspace}/extractions/`	soul.json, voice.json, frameworks.json, etc.
Build Artifacts	`{workspace}/artifacts/`	System prompt, knowledge files, tool configs
Rubric Scores	`{workspace}/rubrics/`	Scoring data per artifact
Source Files	`{workspace}/sources/`	Raw input files (PDFs, transcripts, etc.)

Skill Structure Pattern

~/.claude/skills/{skill-name}/
  SKILL.md                    # Main skill definition (YAML frontmatter + instructions)
  references/
    templates.md              # Output templates
    rubric-definitions.md     # Scoring heuristics (for extractors/builders)
    examples.md               # Reference examples

Phase 2: Vercel Service + Supabase (3+ Partners)

▼

When Athio scales to 3+ partners, the same skill logic wraps into MCP tools served from a Vercel edge function, with state moving to Supabase.

CLI (Phase 1)	MCP (Phase 2)
SKILL.md instructions	MCP tool handler with same logic
kanban.json on filesystem	Supabase table with realtime subscriptions
Extraction JSON files	Supabase JSONB columns + Vercel Blob for large outputs
Task tool spawns agents	Webhook-triggered workflows via n8n or inngest
hc-publish.js for output	NowPage API direct from MCP handler

The migration path is additive — Phase 1 skills continue working, Phase 2 wraps them in API surfaces.

5-Layer Architecture

▼

Layer 0: Recon → Layer 1: Extractors → Layer 2: Synthesizers → Layer 3: Builders ← Layer 4: Orchestrator

Layer	Purpose	Skills	Input	Output
0. Recon	Pre-meeting deep research, web scraping, IP assessment & MasteryBook sync	deep-research, expert-recon, masterybook-sync	Expert name, known URLs, brand	deep-research-report.md, recon.json, raw-sources/, recon-summary.md, MasteryBook notebook
1. Extractors	Pull patterns from raw IP	soul, voice, framework, resource, design-system	Transcripts, PDFs, docs, URLs + recon data	Structured JSON extractions
2. Synthesizers	Generate quality assurance	rubric-builder, gap-analyzer	Extraction outputs	Rubrics + gap reports
3. Builders	Generate artifacts scored by rubrics	clone-compiler, lead-magnet, onboarding	Extractions + rubrics	System prompt, KF, lead magnet, onboarding flow
4. Orchestrator	Coordinates everything	boarding-orchestrator	Kanban board + workspace	Completed expert clone

03

Source Files Analyzed

▼

This section catalogs every source file analyzed to build this architecture, ensuring LLM continuity.

Reference Implementation: Align360

▼

File	What It Contains	Extraction Relevance
`Align360_System_Prompt_v6.1.md`	685-line system prompt with 12 sections: Identity, FLC Wisdom Framework (5 governing layers + Clarity Path + Tri-Filter + 7 absolute rules), Personality/Tone/Character, Mode System, Tool Activation Protocol, Phase Menus, Pathfinder, Cross-Phase Intelligence, Guardrails, Canonical Statements, Recommended Pathways, Background Systems (8 invisible layers)	KEYSTONE. This is the target output of the entire pipeline. Every extraction skill exists to produce a document like this. The soul-extractor captures the FLC Wisdom Framework equivalent; voice-extractor captures Section 3; framework-extractor captures Sections 4-6.
`Align360_Knowledge_File_Part1.md`	936-line knowledge file covering Phase 0 (7 stacks) + Phase 1 (8 stacks) + Supporting Content (Seven Redemptive Gifts detailed profiles + Guardrails). Each stack has: Purpose, Key Inputs, Framework, UX Outputs, Prompt Template.	PRIMARY. This is the reference for what framework-extractor produces. Each "stack" maps to a tool config with inputs, process, and outputs.
`Align360_Knowledge_File_Part2.md`	654-line knowledge file covering Phase 2 (7 stacks), Phase 3 (7 stacks), Phase 4 (7 stacks) + Cross-Phase Integration + Formation Resources. Total: 36 stacks across 5 phases.	REFERENCE. Shows full scope of what a mature clone looks like. Framework-extractor must handle extracting this many tools from raw IP.
`Align360_Background_Tools_Overview_v2.md.pdf`	PDF overview of 14 background tools (invisible layer systems)	IMPORTANT. Soul-extractor must identify which expert behaviors become background systems vs. user-facing tools.
`Align360 Governance Document.docx`	Governance values: Clarity, Integrity, Empathy, Balance, Growth, Contribution, Joy. Anti-patterns ("Laws of Babylon").	CRITICAL for soul-extractor. Every expert has equivalent governance values that must be extracted first.
`Branding/`	Logo (A360logo.jpg) + Branding doc with colors (#2e3c45, #edefe8, #7aa49c, #e09b67, #e45742) and font (Quicksand)	design-system-extractor target. Already built and proven.

Architecture & Strategy Files

▼

File	What It Contains	Key Takeaways
`first-principles.md`	OS build strategy + dependency map. Reference architecture (FreedomLife). Module mapping (FreedomLife → Align360). Dependencies (Must-Have for MVP/Retention/Growth/Future).	The three-column layout pattern, module mapping approach, and dependency chains inform how clone-compiler structures its output.
`align360-details.md`	Phase details (all 5 phases with stack counts), background tools breakdown, partnerships (B3lieve, Africa, YM), pricing model, anti-patterns, governance values.	Pricing and partnership patterns will inform onboarding-builder's monetization section. Anti-patterns are soul-extraction targets.
`freedomlife-os-analysis.md`	Detailed analysis of FreedomLife OS screenshots — the reference MasteryOS implementation.	The three-column layout, chat interface patterns, and tool activation UX patterns inform clone-compiler's output structure.

Existing Skill: design-system-extractor

▼

File	What It Contains
`SKILL.md`	320-line skill definition with: YAML frontmatter (name, description), 4-step workflow (Source Analysis → Token Extraction → System Generation → Publication), Color Derivation Rules, LLM Instruction Block format, Code Block format, Output Requirements, Cross-Platform Notes.
`references/architecture.md`	1628-line complete HTML template with placeholder tokens, CSS architecture, JS interactivity, section content templates.
`references/color-science.md`	Color derivation algorithms for HSL manipulation, contrast checking, palette generation.

This skill establishes the pattern all other skills follow: YAML frontmatter → Overview → Step-by-step Workflow → Output format → References folder with templates.

04

Extraction Pipeline Design

▼

Pipeline Flow

▼

Expert Raw IP (transcripts, docs, PDFs, URLs, brand assets)
    |
    v
[soul-extractor]     -- thinking loops, values, governance, anti-patterns
[voice-extractor]    -- tone, vocabulary, energy, sentence patterns
[framework-extractor] -- tools, sequences, taxonomy, dependencies
[resource-extractor]  -- books, videos, articles, courses, mentors
[design-system-extractor] -- colors, typography, spacing, components
    |
    v  (all extractions complete)
[rubric-builder]     -- scoring heuristics from extracted patterns
[gap-analyzer]       -- completeness audit, missing info questions
    |
    v  (rubrics + gaps resolved)
[clone-compiler]     -- system prompt + knowledge files + tool configs
[lead-magnet-builder] -- interactive assessment from primary framework
[onboarding-builder]  -- gamified card flow + pathfinder routing
    |
    v  (all artifacts scored by rubrics)
REVIEW GATE          -- human review of compiled clone
    |
    v
DEPLOYED CLONE       -- published to MasteryOS

Extraction Output Schemas

▼

soul.json

{
  "expert": "Samuel Ngu",
  "governing_framework": {
    "name": "FLC Wisdom Framework",
    "layers": [...],
    "processing_path": [...],
    "output_filter": {...}
  },
  "values": ["Clarity", "Integrity", "Empathy", ...],
  "anti_patterns": ["manufactured urgency", "identity erosion", ...],
  "governance_rules": [...],
  "background_behaviors": [
    { "name": "Epistemic Drift Detection", "triggers": [...], "responses": [...] },
    { "name": "Pastoral Discernment", "activation_signals": [...] }
  ],
  "canonical_statements": {
    "completion": "This is complete for your stated goal...",
    "no_pressure": "We'll never rush you, track you, or pressure you..."
  },
  "thinking_loops": {
    "before_responding": ["Pause", "Understand", "Simplify", "Guide", "Reflect"],
    "before_recommending": ["check workload", "check season", "check energy"]
  }
}

voice.json

{
  "expert": "Samuel Ngu",
  "character": "digital mentor — steady, kind, objective",
  "tone_principles": [
    { "name": "Warm, not sentimental", "definition": "...", "example": "..." },
    { "name": "Direct, not harsh", "definition": "...", "example": "..." }
  ],
  "language_rules": {
    "avg_sentence_length": "12-18 words",
    "reading_level": "8th grade",
    "vocabulary_style": "universal, no jargon, no corporate speak",
    "spiritual_framing": "present where it fits, never forced"
  },
  "never_say": ["You should do this...", "This is the right decision...", ...],
  "opening_patterns": {...},
  "closing_patterns": {...},
  "energy_spectrum": {
    "high": "engaged, forward-moving, tool execution",
    "low": "reflective, rest-offering, completion-naming"
  }
}

frameworks.json

{
  "expert": "Samuel Ngu",
  "phases": [
    {
      "id": 0, "name": "DesignSuite", "promise": "Discover how you're wired before you build",
      "stacks": [
        {
          "id": 1, "name": "Wiring for Impact",
          "purpose": "Identify primary and secondary Redemptive Gifts",
          "inputs": ["assessment_responses", "current_role", "career_context"],
          "framework": { "structure": "10 adaptive questions", "scoring": "multi-dimensional" },
          "ux_outputs": ["Gift Profile Card", "Design Summary", "Strengths Breakdown", ...],
          "prompt_template": "..."
        }
      ]
    }
  ],
  "cross_phase_bridges": [
    { "from": "DesignSuite", "to": "Career Navigator", "data": ["spiritual_gift", "orientation_profile"] }
  ],
  "taxonomy": { "phases": 5, "stacks": 36, "background_tools": 14 }
}

resources.json

{
  "expert": "Samuel Ngu",
  "recommended_reading": [
    { "title": "The Go-Giver", "author": "Bob Burg", "season": "Stabilize", "relevance": "..." }
  ],
  "recommended_podcasts": [...],
  "external_frameworks": ["Romans 12:6-8", "Kiyosaki Cashflow Quadrants", "Dave Ramsey Baby Steps"],
  "governance_rules_for_resources": [
    "Always present as companions, not requirements",
    "Never recommend resources that promote shame, urgency, or fear"
  ]
}

Quality Gates (Anti-Collapse Mechanism)

▼

Every artifact must pass rubric scoring before advancing on the kanban board. This prevents "AI collapse" where generated content gradually loses fidelity to the expert's actual patterns.

Gate	What's Checked	Pass Threshold
Extraction → Scoring	Completeness (all fields populated), Source fidelity (citations to raw material), Consistency (no contradictions between extractions)	All fields populated + at least 3 source citations per major section
Scoring → Compilation	Rubric coverage (heuristics exist for every extraction category), Gap resolution (all critical gaps addressed)	No critical gaps remaining + rubric covers 80%+ of extraction fields
Build → Review	Voice fidelity score, Framework accuracy score, Governance compliance score, Completeness score	All scores ≥ 7/10 + zero governance violations

05

Agentic Kanban System

▼

Board Structure

▼

Backlog

Future work items

Intake

Source files uploaded

Extraction

Soul extraction

Voice extraction

Framework extraction

Scoring

Build rubrics

Gap analysis

Compilation

Clone compilation

Build

Lead magnet

Onboarding flow

Review

Human review

Done

kanban.json Schema

▼

{
  "version": "1.0.0",
  "expert": {
    "slug": "samuel-ngu",
    "name": "Samuel Ngu",
    "brand": "Feeling Like Chocolate",
    "created": "2026-03-09T00:00:00Z"
  },
  "columns": ["backlog","intake","extraction","scoring","compilation","build","review","done"],
  "cards": [
    {
      "id": "card-001",
      "title": "Upload source files",
      "column": "done",
      "skill": null,
      "assignee": "human",
      "created": "2026-03-09T00:00:00Z",
      "moved": "2026-03-09T01:00:00Z",
      "outputs": ["sources/transcripts/", "sources/docs/"],
      "score": null,
      "blocked_by": []
    },
    {
      "id": "card-002",
      "title": "Extract soul patterns",
      "column": "extraction",
      "skill": "soul-extractor",
      "assignee": "agent",
      "created": "2026-03-09T00:00:00Z",
      "moved": null,
      "outputs": [],
      "score": null,
      "blocked_by": ["card-001"]
    }
  ],
  "history": [
    { "timestamp": "...", "card_id": "card-001", "from": "intake", "to": "done", "agent": "human" }
  ]
}

18 Pre-Templated Cards (with --recon)

▼

When the orchestrator runs init --recon, all 18 cards are created. Without --recon, only cards 1-15 are created (warm boarding).

#	Card Title	Skill	Column Start	Blocked By
0a	Run deep research	deep-research	backlog	—
0b	Run expert recon	expert-recon	backlog	0a
0c	Sync to MasteryBook	masterybook-sync	backlog	0b
1	Upload source files	human	intake	0b (with --recon) or — (without)
2	Extract soul patterns	soul-extractor	extraction	1
3	Extract voice patterns	voice-extractor	extraction	1
4	Extract frameworks & tools	framework-extractor	extraction	1
5	Extract resources & references	resource-extractor	extraction	1
6	Extract design system	design-system-extractor	extraction	1
7	Build scoring rubrics	rubric-builder	scoring	2, 3, 4
8	Run gap analysis	gap-analyzer	scoring	2, 3, 4, 5
9	Resolve gaps (human + AI)	human + gap-analyzer	scoring	8
10	Compile clone (system prompt + KF)	clone-compiler	compilation	7, 9
11	Score compiled clone	rubric-builder	compilation	10
12	Build lead magnet	lead-magnet-builder	build	10
13	Build onboarding flow	onboarding-builder	build	10
14	Score all build artifacts	rubric-builder	build	12, 13
15	Human review & approval	human	review	11, 14

06

14 Skills — Complete Specification

▼

LLM Implementation Guide

Each skill below is specified at the level needed to build a SKILL.md file. The skill file should follow the pattern established by design-system-extractor: YAML frontmatter (name, description) → Overview → Step-by-step Workflow → Output format → References folder with templates/rubrics/examples.

Layer 0: Reconnaissance

DEEP RESEARCH

0a. deep-research

Purpose: Heavy-lifting research engine that wraps the Perplexity Sonar Deep Research API. Runs 3-5 sequential deep research queries to build comprehensive intelligence on an expert's public presence. Falls back to Claude's native WebSearch + WebFetch when Perplexity is unavailable.

Inputs: Expert name, brand/company name, known URLs, vertical

Outputs: deep-research-report.md (full narrative), deep-research.json (structured proto-findings with confidence scores), deep-research-sources.json (all citation URLs)

Key capabilities:

5 query categories: background/credentials, offerings/pricing, interviews/philosophy, competitors/landscape, recent activity
Perplexity Sonar Deep Research API (~$0.40-1.30 per run)
Automatic fallback to WebSearch + WebFetch if PPLX_API_KEY unavailable
Structured findings matching recon.json proto-schema with confidence scores
Source quality assessment (high/medium/low authority)

Integration: Delegated to by expert-recon as the first step. Outputs feed into supplemental scraping and pattern pre-extraction.

BUILT

ORCHESTRATOR

0b. expert-recon

Purpose: Layer 0 orchestrator — coordinates deep research, supplemental web scraping, pattern pre-extraction, IP assessment, and MasteryBook sync. Produces the complete recon package.

Inputs: Expert name, known URLs (LinkedIn, YouTube, website), brand/company name

Outputs: recon.json (merged proto-findings from deep-research + scraping), raw-sources/ folder, recon-summary.md (human-readable brief for onboarding call)

Key capabilities:

Delegates to deep-research for Perplexity-powered intelligence
Supplemental web scraping across 9+ platforms for content deep-research can't reach
YouTube transcript extraction for interview/talk content
Merged pattern pre-extraction with confidence scoring (confirms across both methods)
IP Footprint Assessment: Volume, Depth, Consistency, Uniqueness, Extractability (1-5 each)
Boarding Readiness Score (1-10) determines how much manual upload is needed
Delegates to masterybook-sync for team-accessible RAG notebook

Integration: Pre-populates _workspaces/{expert}/sources/, pre-seeds extraction context for all Layer 1 extractors, enriches gap analysis. Orchestrator's init --recon flag triggers this before standard kanban cards.

BUILT

SYNC

0c. masterybook-sync

Purpose: Syncs expert workspace sources to a MasteryBook notebook for team-accessible RAG-based Q&A. Creates a notebook, uploads text/URL/YouTube/PDF sources, and returns a shareable notebook URL.

Inputs: Workspace path, optional notebook ID, expert name

Outputs: masterybook-sync-status.json (notebook ID, upload counts), masterybook-summary.md (RAG-generated executive summary)

Key capabilities:

MasteryBook API integration (FastAPI backend at notebooklm-api.vercel.app)
Uploads text, URLs, YouTube, and PDF sources via appropriate endpoints
Generates executive summary via RAG chat query
Graceful degradation: If MasteryBook API is unavailable, reports status and continues — never blocks the pipeline

Integration: Delegated to by expert-recon as the final step. Notebook URL included in recon-summary.md for team access.

BUILT

Layer 1: Extractors

KEYSTONE

1. soul-extractor

Purpose: Extract thinking loops, values, governance rules, anti-patterns, background behaviors, and canonical statements from expert's raw IP.

Inputs: Transcripts, governance docs, system prompts, interviews

Outputs: soul.json — governing framework, values, anti-patterns, background behaviors, thinking loops, canonical statements

Key extraction targets:

Governing framework (equivalent to FLC Wisdom Framework)
Processing path (equivalent to Clarity Path: Pause → Understand → Simplify → Guide → Reflect)
Output filter (equivalent to Tri-Filter: Truth, Clarity, Impact)
Absolute governance rules (equivalent to "Zero Coercion", "Completion is Sacred", etc.)
Background system behaviors (equivalent to Epistemic Drift Detection, Pastoral Discernment, etc.)
Anti-patterns ("Laws of Babylon" the expert explicitly resists)

2. voice-extractor

Purpose: Extract tone, vocabulary, energy spectrum, sentence patterns, and character description from expert's communications.

Inputs: Transcripts, blog posts, emails, presentations, social media

Outputs: voice.json — character description, tone principles, language rules, never-say list, opening/closing patterns, energy spectrum

Key extraction targets:

Character description (equivalent to "digital mentor — steady, kind, objective")
Tone principles as contrast pairs (equivalent to "Warm, not sentimental")
Average sentence length, reading level, vocabulary style
Phrases the expert NEVER uses (equivalent to "What You Never Say")
Signature phrases and recurring metaphors

3. framework-extractor

Purpose: Extract tools, sequences, taxonomies, phase structures, and inter-tool dependencies from expert's methodology.

Inputs: System prompts, knowledge files, course materials, training docs

Outputs: frameworks.json — phases, stacks (with purpose/inputs/framework/outputs/prompt template), cross-phase bridges, taxonomy

Key extraction targets:

Phase/module structure (how many phases, what progression)
Individual tools/stacks with full specification
Data flow between tools (cross-phase bridges)
Recommended pathways and routing logic
Tool activation triggers (what user says → which tool runs)

4. resource-extractor

Purpose: Extract recommended books, podcasts, videos, courses, external frameworks, and mentors referenced in expert's IP.

Inputs: Knowledge files, transcripts, course materials

Outputs: resources.json — categorized resources with relevance context, external frameworks referenced, governance rules for resource recommendations

Key extraction targets:

Books/podcasts/videos organized by season/phase
External frameworks the expert builds on (e.g., Kiyosaki, Ramsey, StoryBrand)
Rules for how resources should be recommended

BUILT

5. design-system-extractor

Purpose: Extract colors, typography, spacing, components from expert's brand.

Status: Already built and working at ~/.claude/skills/design-system-extractor/

Layer 2: Synthesizers

6. rubric-builder

Purpose: Generate scoring heuristics from extracted patterns. These rubrics score all downstream artifacts for fidelity to the expert's actual IP.

Inputs: soul.json, voice.json, frameworks.json

Outputs: rubrics.json — scoring criteria for: voice fidelity, framework accuracy, governance compliance, completeness

Rubric categories:

Voice Fidelity (0-10): Does the artifact sound like the expert? Check tone principles, sentence patterns, vocabulary, never-say violations.
Framework Accuracy (0-10): Are tools, sequences, and logic correct? Check against frameworks.json for step accuracy.
Governance Compliance (0-10): Does the artifact respect ALL governance rules? Zero tolerance for anti-pattern violations.
Completeness (0-10): Are all required sections present? Check against extraction output schemas.

7. gap-analyzer

Purpose: Audit extraction completeness and generate follow-up questions for missing information.

Inputs: All extraction outputs (soul, voice, frameworks, resources)

Outputs: gaps.json — missing fields, weak areas, follow-up questions for the expert/team, priority ranking

Gap categories:

Critical: Missing governance rules, no voice samples, incomplete tool specs
Important: Thin background behaviors, few canonical statements, sparse resource list
Nice-to-have: Additional voice samples, more tool activation triggers, edge case coverage

Layer 3: Builders

8. clone-compiler

Purpose: Compile extractions into a complete AI coaching OS: system prompt + knowledge files + tool configurations. This is the core output of the entire pipeline.

Inputs: soul.json, voice.json, frameworks.json, resources.json, rubrics.json

Outputs:

system-prompt.md — Full system prompt following the v6.1 section structure
knowledge-file-part1.md — Active phase stacks with full tool specifications
knowledge-file-part2.md — Future phase stacks (if applicable)
tool-configs.json — MasteryOS tool configuration format

Compilation rules:

System prompt MUST follow the 12-section structure from v6.1
Voice from voice.json applied to ALL prose sections
Soul from soul.json becomes Sections 2, 9, 12
Frameworks from frameworks.json become Sections 4-6, 8, 11 + Knowledge Files
Every output scored against rubrics before advancing

9. lead-magnet-builder

Purpose: Build an interactive assessment (like Wiring for Impact) as a standalone lead magnet HTML page with email gate.

Inputs: frameworks.json (primary assessment tool), voice.json, design system

Outputs: lead-magnet.html — self-contained assessment page with: questions, scoring, result types, email capture, CTA to full platform

Reference: align360.asapai.net/wiring-for-impact (published example)

10. onboarding-builder

Purpose: Build a gamified, Netflix-style card-based onboarding flow that routes new users to the right starting tool.

Inputs: frameworks.json (tool menu + pathfinder logic), voice.json, design system

Outputs: onboarding-flow.html — card-based UI with: "what's urgent" + "where's the dissonance" questions, pathfinder routing logic, tool recommendations

Design principles: Netflix-style gamified cards, no linear forced path, "what's urgent" + "where's the dissonance" discovery questions.

Layer 4: Orchestrator

CHIEF OF STAFF

11. boarding-orchestrator

Purpose: The Chief of Staff agent. Manages the kanban board, lazy-loads skills on demand, enforces quality gates, tracks progress, and coordinates the entire extraction-to-clone pipeline.

Inputs: Expert workspace path, kanban.json

Capabilities:

Initialize: Create a new expert workspace with kanban.json (18 cards with --recon, or 15 cards without)
Status: Read and display current kanban state
Run next: Identify the next unblocked card, invoke its skill, write outputs, update board
Score: After a build skill completes, invoke rubric-builder to score the output
Gate: Enforce quality thresholds before advancing cards
Report: Generate a progress summary for human review

Orchestrator rules:

Never run two extraction skills simultaneously on the same source file
Always run gap-analyzer after all extractors complete
Never advance to compilation without rubric scores
The orchestrator is the ONLY skill that writes to kanban.json

07

Build Order & Dependencies

▼

Order	Skill	Layer	Depends On	Rationale
0a	deep-research	Recon	None (standalone)	Perplexity-powered intelligence. Runs first to seed all downstream research. No dependencies.
0b	masterybook-sync	Recon	None (standalone)	MasteryBook notebook sync. No dependencies on other skills. Gracefully degrades if API unavailable.
1	soul-extractor	Extractor	design-system-extractor (pattern)	KEYSTONE. Governance must exist before any other extraction makes sense. Other extractors reference soul for consistency.
2	voice-extractor	Extractor	soul-extractor (pattern)	Voice is independent of frameworks but benefits from soul context. Can be built in parallel with framework-extractor.
3	framework-extractor	Extractor	soul-extractor (pattern)	The largest extractor. Tool specs are the core product value. Can be built in parallel with voice-extractor.
4	resource-extractor	Extractor	framework-extractor (context)	Simpler extractor. Resources map to phases/tools from frameworks.json.
5	rubric-builder	Synthesizer	soul, voice, framework extractors	Can't build scoring criteria until you know what you're scoring against.
6	gap-analyzer	Synthesizer	All extractors	Needs complete extraction picture to identify what's missing.
7	clone-compiler	Builder	All extractors + rubric-builder	The core compilation step. Needs everything upstream.
8	lead-magnet-builder	Builder	framework-extractor + clone-compiler	Builds the standalone assessment. Can be parallel with onboarding-builder.
9	onboarding-builder	Builder	framework-extractor + clone-compiler	Builds the gamified onboarding. Can be parallel with lead-magnet-builder.
12	boarding-orchestrator	Orchestrator	All other skills	Built last because it needs to know all skill interfaces. Now supports 18 cards with --recon Layer 0 or 15 cards without.

deep-research [0a] → masterybook-sync [0b] → design-system-extractor [BUILT] → soul-extractor [1] → voice [2] + framework [3] → resource [4] → rubric [5] + gap [6] → clone [7] → lead-magnet [8] + onboarding [9] → orchestrator [12]

08

Verification Criteria

▼

Criterion	How to Verify	Status
PRD published	Accessible at cowork.asapai.net/boarding-architecture	Pending publish
Skills buildable from PRD	Each skill section contains enough detail to create a SKILL.md without additional context	Self-contained in Section 06
Kanban testable	boarding-orchestrator can initialize a workspace, create kanban.json, and process at least one card	Pending skill build
Align360 as reference	Running the full pipeline against Align360 source files produces outputs comparable to the existing v6.1 system prompt	Integration test (future)
LLM continuity	A new Claude session reading only this PRD + skill files can continue building from where the last session left off	Architecture verified

Session Continuity Protocol

When starting a new session to continue building the boarding system:

Read this PRD (boarding-architecture.html or the published URL)
Read the MEMORY.md file for project context
Check ~/.claude/skills/ for which skills are already built
Check any active expert workspace's kanban.json for pipeline state
Resume from the next unbuild skill or unprocessed kanban card