Product Requirements Document

Agentic Expert Boarding
via Swarm Intelligence

How Athio will autonomously board JV expert partners from raw IP to launched Expert OS in 8 weeks -- replacing 200 hours of manual work with an orchestrated agent swarm that handles 90%+ of extraction, research, and artifact generation.

Document
PRD v1.0
Date
2026-03-09
Status
Active
Current Phase
Phase 1
Target Human Time
<10 hrs
1

Vision & Goals

▼
North Star

Board an expert partner from raw IP to launched Expert OS in 8 weeks with less than 10 hours of human time.

From: Manual Process
  • 3 team transcripts manually analyzed
  • Multiple source documents manually read and synthesized
  • Design system manually extracted from brand assets
  • Positioning brief manually researched and written
  • Boarding extraction pack manually compiled
  • Lead magnet manually coded
  • Roadmap manually built
  • ~200 hours per partner
➔
To: Agentic Swarm
  • Swarm ingests all raw inputs simultaneously
  • Parallel agent workflows extract, research, and generate
  • Design system auto-extracted via skill
  • Positioning brief auto-generated by 7-agent swarm
  • Boarding pack is a living document, auto-updated
  • All artifacts auto-built from extraction output
  • Humans review at 4 gates only
  • <10 hours per partner
Key Metric

Time-to-first-artifact: <2 hours from IP upload. The moment an expert uploads their first transcript or document, the swarm begins processing. Within 2 hours, they should see the first generated artifact (typically the preliminary positioning brief or content inventory).

200h
Current Manual Hours
<10h
Target Human Hours
95%
Automation Target
<2h
First Artifact SLA

What "Expert OS Clone" Means

Athio builds AI-powered platforms that embody a specific expert's intellectual property, frameworks, and mentorship style. Each Expert OS is a complete digital product including: a customized AI assistant trained on the expert's methodology, an offer page, lead magnets, onboarding flows, knowledge files, and a system prompt that captures the expert's voice, governance rules, and framework logic. The agentic boarding system automates the creation of all these components.

The 20x Efficiency Promise

By replacing manual research, extraction, and artifact generation with orchestrated agent swarms, Athio can onboard partners 20x faster while maintaining or improving quality. The key insight: most of the 200 hours are spent on tasks that LLMs excel at -- reading documents, extracting patterns, cross-referencing information, generating structured output, and building HTML artifacts. Humans are needed only for quality judgment and expert approval.

2

System Architecture

▼

2A: The Swarm Structure

18 Agents Total ▼

The boarding system uses a hierarchical multi-agent architecture. A single orchestrator (the Chief of Staff) decomposes incoming work into parallel workstreams, delegates to color-coded specialist swarms, tracks completion across all gates and workflows, and assembles the final boarding pack.

Chief of Staff Agent (Orchestrator)

The top-level agent that manages the entire boarding process.

  • Receives raw inputs (transcripts, documents, brand assets, URLs)
  • Decomposes inputs into parallel workstreams by type and workflow color
  • Assigns tasks to specialist agents via the Task tool or Agent SDK
  • Tracks progress across all 4 gates and 5 color workflows
  • Raises blockers to humans when agent confidence is below threshold
  • Maintains the boarding extraction pack as a living document
  • Triggers builder agents when extraction outputs are ready
  • Reports status via structured progress updates

PURPLE Swarm -- Positioning (7 Agents)

Market Research Agent

TAM/SAM/SOM calculation, market sizing, growth trend analysis, entry barrier assessment.

Competitive Analysis Agent

Scrapes competitor websites, builds positioning matrix, identifies market gaps and differentiation opportunities.

Audience Profiling Agent

Extracts persona details from transcripts and source material. Builds 2-3 detailed buyer personas with pain points and buying journey maps.

Niche Validation Agent

Scores niche viability across predefined criteria. Market size validation, growth potential, go/no-go recommendation.

Keyword Research Agent

SEO keyword discovery and clustering. 30+ keywords researched, opportunity scored, content pillars defined.

Trend Analysis Agent

Macro trend identification, opportunity window assessment, competitive timing analysis.

Psychology Agent

Dreams/Fears/Suspicions/Enemies extraction. Hero's Journey narrative mapping. Sales objection playbook generation.

GREEN Agent -- Brand Extraction (1 Agent)

Brand System Extractor

Takes logo, colors, fonts from expert and generates a complete design system. Uses the design-system-extractor skill pattern already built.

  • CSS variables and custom properties
  • Tailwind configuration
  • Component library (buttons, cards, badges)
  • Brand guidelines document
  • LLM usage instructions for consistent generation

BLUE Swarm -- Content Extraction (3 Agents)

Transcript Analyzer

Processes call transcripts. Extracts key decisions, frameworks, direct quotes, personality traits, communication style markers, and audience signals.

Document Extractor

Reads PDFs, DOCX files, knowledge files, slide decks. Converts to structured data with tagged sections, framework diagrams, and methodology maps.

Content Auditor

Inventories YouTube videos, articles, social posts, books. Builds a complete content library with metadata, categories, and relevance scores.

YELLOW Agent -- Network Mapping (1 Agent)

Network Intelligence Agent

LinkedIn API integration for connection analysis. Partnership opportunity scoring, warm intro path identification, conflict checking, and network intelligence brief generation.

RED Agent -- Domain Generation (1 Agent)

Domain Generator

Uses positioning brief + psychology output to generate 50+ domain name candidates. Checks availability via GoDaddy MCP/API, scores and ranks options by memorability, brandability, SEO value, and availability.

Builder Agents -- Post-Extraction (5 Agents)

Lead Magnet Builder

Takes the expert's primary framework and builds an interactive assessment page. Single-file HTML with scoring logic, brand-consistent design, and lead capture.

Offer Page Builder

Takes positioning brief + pricing tiers to build a complete offer page. Hero section, roadmap, testimonials layout, FAQ, and CTA with tier-specific messaging.

Onboarding Flow Builder

Takes boarding questions and builds a card-based interactive onboarding flow. Progressive disclosure, validation logic, and data capture.

System Prompt Builder

Takes frameworks + personality traits + governance rules and generates a complete system prompt. Voice calibration, tool instructions, boundary rules, and escalation logic.

Knowledge File Builder

Takes tool stacks, methodology docs, and extracted content to generate chunked knowledge files optimized for RAG retrieval.

2B: Data Flow Architecture

Visual Diagram ▼
Raw Inputs Transcripts, Documents, Brand Assets, URLs
▼
Chief of Staff Agent Decompose + Assign + Track
▼
PURPLE
7 Agents -- Positioning
GREEN
1 Agent -- Brand System
BLUE
3 Agents -- Content
YELLOW
1 Agent -- Network
RED
1 Agent -- Domain
▼
Positioning
Brief
Design System
+ Guidelines
Content
Library
Network Intel
Brief
Domain
Selection
│ │ │ │ │
└──────┴──────┴──────┴──────┘
▼
BOARDING EXTRACTION PACK Living Document -- Auto-Updated by Chief of Staff
▼
Lead Magnet
Offer Page
Onboarding Flow
System Prompt
Knowledge Files
▼
Human Review Gates (4) Brand Assets | Extraction | Permissions | Offer Approval
▼
LAUNCH
Parallel Execution Model

All five color workflows run simultaneously once the Chief of Staff has decomposed inputs. PURPLE, GREEN, and BLUE swarms can begin immediately from raw inputs. YELLOW requires LinkedIn access (Gate 3). RED requires PURPLE output (positioning brief) to generate semantically relevant domain names. Builder agents activate only after the boarding pack has sufficient data (typically after PURPLE + BLUE complete).

3

Recursive Extraction Protocol

▼
Core Innovation

This multi-pass extraction protocol is the key differentiator. Rather than a single extraction pass, the system iterates through four progressive passes -- each one building on the output of the previous. By the fourth pass, the system has not only extracted all available information, but has also identified gaps, generated quality rubrics, and produced targeted follow-up questions.

1

First Pass -- Raw Extraction

Parallel Ingest

Agent reads ALL source material simultaneously (transcripts, docs, PDFs, URLs). Outputs structured data with confidence scores.

Extracts
  • Facts -- biographical data, business history, product details, pricing, team structure
  • Frameworks -- named methodologies, step-by-step processes, proprietary models
  • Quotes -- memorable phrases, repeated language, taglines, brand voice samples
  • Personality Traits -- communication style markers, energy level, formality, humor use
  • Audience Signals -- who the expert talks to, pain points mentioned, success stories
  • Pricing Data -- current offers, price points, tier structures, revenue models
  • Partnership Details -- existing collaborations, affiliate relationships, joint ventures
Output Format
JSON Schema 
{
  "extraction_pass": 1,
  "source_count": 6,
  "data": {
    "facts": [
      {
        "content": "Founded Align360 in 2019...",
        "source": "transcript_01.txt",
        "confidence": 0.95,
        "category": "business_history"
      }
    ],
    "frameworks": [
      {
        "name": "5-Phase Alignment Model",
        "steps": ["Design", "Career", "Integrate", "Figures", "Legacy"],
        "source": "governance_doc.docx",
        "confidence": 0.98
      }
    ],
    "quotes": [],
    "personality": {},
    "audience": {},
    "pricing": {},
    "partnerships": {}
  },
  "metadata": {
    "processing_time_ms": 45000,
    "total_tokens_analyzed": 125000,
    "timestamp": "2026-03-09T10:00:00Z"
  }
}
2

Second Pass -- Pattern Recognition

Cross-Reference

Takes the raw extraction output and cross-references data across ALL sources to find patterns, consistency, and contradictions.

Identifies
  • Repeated Themes -- concepts that appear across 2+ sources (high signal)
  • Consistent Language Patterns -- phrases, metaphors, and framings the expert uses consistently
  • Framework Structures -- how the expert organizes knowledge, common numbering (3-step, 5-phase, etc.)
  • Contradictions -- conflicting data between sources (flags for human review)
  • Voice Fingerprint -- sentence length distribution, vocabulary complexity, emotional tone markers
Builds
  • Framework Taxonomy -- hierarchical map of all frameworks, sub-frameworks, and tools
  • Voice/Tone Profile -- quantified profile used to calibrate all generated content
  • Terminology Glossary -- expert-specific terms with definitions and usage context
3

Third Pass -- Rubric Building

Quality Assurance

From the patterns identified in Pass 2, generates scoring rubrics that become the quality assurance layer for all generated artifacts.

Rubrics Generated
  • Brand Voice Consistency -- does generated content sound like the expert? Scores tone, vocabulary, sentence structure, energy level against the voice fingerprint.
  • Framework Fidelity -- do tools and assessments accurately implement the expert's methodology? Checks step names, sequence, terminology, and logic against the framework taxonomy.
  • Audience Alignment -- does positioning resonate with identified personas? Scores pain point relevance, aspiration mapping, and language match.
Rubric as QA Layer

These rubrics serve a dual purpose: (1) automated quality scoring of every generated artifact before human review, and (2) specific, actionable feedback when artifacts fail quality thresholds. Instead of vague "make it better" feedback, the rubric produces targeted revision instructions like "Framework step 3 uses 'Integrate' but expert source consistently says 'Alignment Integration' -- update terminology."

Scoring Format
JSON 
{
  "rubric": "brand_voice_consistency",
  "version": "1.0",
  "criteria": [
    {
      "dimension": "tone_match",
      "weight": 0.3,
      "scoring": {
        "5": "Indistinguishable from expert's natural voice",
        "3": "Generally matches but occasional mismatches",
        "1": "Clearly generic / does not sound like expert"
      }
    },
    {
      "dimension": "terminology_accuracy",
      "weight": 0.25,
      "scoring": {
        "5": "All expert-specific terms used correctly",
        "3": "Most terms correct, 1-2 errors",
        "1": "Generic language, no expert terminology"
      }
    },
    {
      "dimension": "framework_reference",
      "weight": 0.25,
      "scoring": {
        "5": "Naturally references expert's frameworks",
        "3": "Mentions frameworks but unnaturally",
        "1": "No framework references"
      }
    },
    {
      "dimension": "energy_calibration",
      "weight": 0.2,
      "scoring": {
        "5": "Matches expert's energy level perfectly",
        "3": "Somewhat close but off in intensity",
        "1": "Completely wrong energy"
      }
    }
  ],
  "minimum_pass_score": 3.5,
  "target_score": 4.2
}
4

Fourth Pass -- Gap Analysis

Completeness

Compares extracted data against the standard boarding template and identifies what is missing, what blocks launch, and what can be deferred.

Process
  • Compare extraction output against the complete boarding template (all required fields)
  • Identify MISSING information with severity levels (blocker, important, nice-to-have)
  • Generate specific, contextual follow-up questions for the expert
  • Prioritize: what blocks launch vs. what can be deferred or inferred
Output: Follow-up Question Generation
Markdown 
## Follow-up Questions for [Expert Name]

### BLOCKERS (Required Before Launch)
1. **Pricing Structure**: We extracted references to "3 tiers" but
   could not determine specific price points. What are the
   exact prices for each tier?
   _Source gap: mentioned in transcript_02 at 14:32 but
   not specified_

2. **Governance Rules**: What topics should the AI assistant
   refuse to discuss? What boundaries exist?
   _Source gap: no governance document provided_

### IMPORTANT (Recommended Before Launch)
3. **Case Studies**: Can you share 2-3 specific client
   success stories with metrics?
   _Source gap: referenced in transcript_01 but no
   details provided_

### NICE-TO-HAVE (Can Be Added Post-Launch)
4. **Video Testimonials**: Do you have video testimonials
   we can embed on the offer page?
4

Input Specifications

▼

Minimum Viable Inputs

Gets Through Gate 2 ▼

The absolute minimum to begin the boarding process and generate initial artifacts.

  • 1+ transcripts -- any format: .txt, .docx, .pdf, or audio files that will be auto-transcribed
  • Brand assets -- logo file + color codes, OR a website URL to extract from automatically
  • 1 document describing their methodology/framework -- book, course outline, slide deck, or written description
Boarding coverage with minimum inputs ~40%

Optimal Inputs

Gets Through All 4 Gates ▼

The full input set that enables the swarm to reach maximum coverage with minimal follow-up questions.

  • 3+ transcripts covering: methodology discussion, audience/market discussion, pricing/business model discussion
  • Complete brand kit -- logo (PNG/SVG), primary/secondary/accent colors (hex), headline and body fonts, tagline
  • Framework documentation -- books, courses, slide decks, whitepapers, or any structured IP
  • Existing website URL -- for brand extraction and competitive positioning context
  • YouTube channel / content library access -- for content audit and RAG training material
  • LinkedIn profile URL -- for network mapping and partnership intelligence
  • List of existing products/pricing -- for offer page generation and tier structuring
Boarding coverage with optimal inputs ~95%

Sparse Input Handling

Graceful Degradation ▼

The system must handle incomplete inputs gracefully -- extracting maximum value from what is available while clearly flagging gaps.

If Only 1 Transcript
  • Extract everything possible from the single source
  • Flag all data points with confidence levels (lower thresholds when single-source)
  • Generate targeted follow-up questions to fill gaps
  • Mark positioning brief as "preliminary" -- requires additional input before Gate 4
If No Brand Assets
  • If website URL provided: GREEN agent extracts design system from live site
  • If no website: generate brand recommendations based on positioning brief + industry standards
  • Flag design system as "recommended" vs "extracted" -- requires expert approval
If No Framework Documentation
  • Extract frameworks from transcripts (look for numbered steps, repeated structures, named processes)
  • Scrape public content (YouTube descriptions, blog posts, social media) for framework references
  • Build "inferred framework" taxonomy -- clearly labeled as inferred, not confirmed
  • Generate specific questions: "We identified a possible 5-step process you mentioned at 14:32. Can you confirm: [steps listed]?"
Confidence Degradation Rules

Single-source data points receive a maximum confidence of 0.7. Data points confirmed across 2+ sources receive up to 0.95. Inferred data (not directly stated but derived from patterns) receives a maximum confidence of 0.5. Anything below 0.5 confidence is flagged for human review rather than being included in generated artifacts.

5

Output Specifications

▼

Every artifact the system produces is defined here with its format, generating agent, and human review requirements.

Artifact Format Generated By Human Review
Positioning Brief HTML PURPLE Swarm Required
Design System HTML + CSS vars GREEN Agent Light review
Content Library Structured JSON BLUE Swarm Spot check
Network Intel Brief PDF YELLOW Agent Required
Domain Shortlist HTML RED Agent Selection only
Boarding Pack HTML (living doc) Chief of Staff Required
Lead Magnet Page HTML Builder Agent Required
Offer Page HTML Builder Agent Required
Onboarding Flow HTML Builder Agent Required
System Prompt Markdown Builder Agent Required
Knowledge Files Markdown Builder Agent Required
Follow-up Questions Markdown Gap Analysis N/A (sent to expert)
QA Rubrics JSON Rubric Builder Light review
Living Document Pattern

The Boarding Extraction Pack is not a static document. It is a living HTML file maintained by the Chief of Staff agent that updates in real-time as each swarm completes its work. Sections are progressively filled in, confidence scores update as new sources are processed, and status indicators show which sections are complete, in-progress, or blocked. This pack serves as the single source of truth for the entire boarding process.

6

Implementation Phases

▼
Phase 1 -- NOW (Samuel / Align360)

Manual + Individual Agents

Humans drive the process. Individual agents handle specific tasks as tools. This is the current operating mode.

  • Humans initiate each step and provide context to agents
  • Design System Extractor skill handles GREEN workflow
  • Claude Code handles transcript analysis and content extraction
  • Manual positioning research with agent assistance for specific subtasks
  • Artifacts built one at a time with human orchestration
Learning Objectives
  • What works well with agent assistance vs. what requires human judgment
  • Where agents excel: extraction, pattern matching, artifact generation
  • Where agents fail: subjective quality assessment, expert voice calibration, strategic decisions
  • Timing data: how long does each step actually take with agent assistance
Automation Level ~30%

Semi-Automated Swarm

Chief of Staff agent orchestrates the process. Swarms run automatically from input. Humans stationed at review gates only.

  • Chief of Staff agent receives inputs and orchestrates all workflows
  • Color swarms execute in parallel without human initiation
  • Boarding pack auto-updates as swarms complete
  • Humans review at each gate, provide approval or revision notes
  • Rubric-based QA catches quality issues before human review
Data Capture
  • Timing data for every agent task (wall clock and token usage)
  • Quality scores via rubrics (automated) vs. human scores (manual)
  • Human override patterns -- where and why did humans change agent output
  • Calibration: how accurate are the rubrics vs. human judgment
Automation Level ~70%
Phase 3 -- Partner 4+

Full Swarm Autonomy

Expert uploads inputs. Swarm runs end-to-end. Humans review final artifacts only.

  • Expert uploads inputs via self-service onboarding dashboard
  • Full swarm executes: extraction, research, artifact generation, quality scoring
  • Automated quality scoring via rubrics refined from Phase 1-2 learning
  • Humans review only the final boarding pack and generated artifacts
  • Target: less than 10 hours human time per partner, end-to-end
Automation Level ~95%
7

Technology Stack

▼
Layer Technology Role
Orchestration Claude Code CLI Chief of Staff runs as a CLAUDE.md-configured agent with access to all skills and tools
Agent Framework Claude Code Task tool or Anthropic Agent SDK Specialist subagents spawned via Task tool for parallel execution. Agent SDK for production deployment.
Skill Library ~/.claude/skills/ Reusable skill definitions: design-system-extractor (built), positioning-extractor (to build), content-extractor (to build), lead-magnet-builder (to build)
Data Store Supabase Boarding pack persistence, extraction data, progress tracking, artifact versioning. Already in use for NowPage.
Publishing NowPage API / folio-saas Artifact publishing to live URLs. Design systems, boarding packs, lead magnets all published via NowPage.
Quality LLM-as-Judge Rubric-based evaluation. A separate Claude instance scores artifacts against generated rubrics.
MCP Servers Hugging Face, Vercel, Context7, GoDaddy External integrations: model search, deployment, documentation, domain availability checking

Chief of Staff CLAUDE.md Configuration

Markdown (CLAUDE.md) 
# Chief of Staff Agent -- Boarding Orchestrator

## Role
You are the Chief of Staff for Athio's expert boarding process.
You orchestrate the entire journey from raw inputs to launched
Expert OS. You decompose work, assign to specialist agents,
track progress, and maintain the boarding pack.

## Available Skills
- design-system-extractor: Extract brand system from assets/URLs
- positioning-extractor: Run PURPLE swarm for positioning brief
- content-extractor: Run BLUE swarm for content library
- lead-magnet-builder: Generate assessment pages from frameworks

## Workflow
1. Receive raw inputs (transcripts, docs, brand assets, URLs)
2. Classify each input by type and relevant color workflow
3. Spawn parallel Task agents for each color workflow
4. Monitor progress, collect outputs
5. Assemble boarding extraction pack
6. Run gap analysis (Pass 4)
7. Generate follow-up questions if needed
8. Trigger builder agents when pack is sufficient
9. Run rubric-based QA on all artifacts
10. Present to human for gate review

## Gates
- Gate 1: Brand Assets uploaded and validated
- Gate 2: Extraction interview processed
- Gate 3: Content permissions granted
- Gate 4: Offer page approved by expert

## Rules
- Never publish artifacts without human gate approval
- Flag any confidence score below 0.5 for human review
- Always maintain the boarding pack as living document
- Track timing data for every agent task

Skill Architecture Pattern

File Structure 
~/.claude/skills/
  design-system-extractor/     # BUILT -- GREEN workflow
    SKILL.md
    references/
      architecture.md
      color-science.md

  positioning-extractor/       # TO BUILD -- PURPLE workflow
    SKILL.md                   # 7-agent swarm orchestration
    prompts/
      market-research.md
      competitive-analysis.md
      audience-profiling.md
      niche-validation.md
      keyword-research.md
      trend-analysis.md
      psychology.md

  content-extractor/           # TO BUILD -- BLUE workflow
    SKILL.md                   # 3-agent content pipeline
    prompts/
      transcript-analyzer.md
      document-extractor.md
      content-auditor.md

  lead-magnet-builder/         # TO BUILD -- Builder agent
    SKILL.md
    templates/
      assessment.html
      quiz.html
      scorecard.html

  boarding-orchestrator/       # TO BUILD -- Chief of Staff
    SKILL.md                   # Master orchestration logic
    templates/
      boarding-pack.html
      gap-analysis.md
      follow-up-questions.md
8

Integration with Existing Tools

▼
Claude Code CLI

Primary interface for all agent operations. Chief of Staff runs here. Task tool spawns subagents. All file I/O happens through CLI tools.

Claude Code Desktop / Co-Work

Used for live demos with expert partners. Real-time boarding process visualization. Expert can watch agents work in real-time during extraction interviews.

NowPage Publishing (folio-saas)

All HTML artifacts (design systems, boarding packs, lead magnets, offer pages) publish to live URLs via the NowPage API. Single POST endpoint for instant publishing.

Design System Extractor Skill

Already built at ~/.claude/skills/design-system-extractor/. Handles the entire GREEN workflow. Takes brand assets or URLs and produces a complete design system HTML file.

Forge (CLAUDE.md Enhancer)

Used to generate and maintain CLAUDE.md files for each expert's workspace. Ensures consistent project configuration and agent instructions across all partner builds.

MCP Server Ecosystem

Hugging Face (model/paper search), Vercel (deployment), Context7 (documentation), GoDaddy (domain availability), Mermaid (diagram generation). All available to agents via MCP protocol.

Integration Data Flow

Integration Map 
Expert Inputs
    |
    v
Claude Code CLI (Chief of Staff)
    |
    +---> Task Tool ---> PURPLE Swarm Agents
    |                       |---> WebSearch (market data)
    |                       |---> WebFetch (competitor sites)
    |                       |---> Context7 (methodology docs)
    |
    +---> Task Tool ---> GREEN Agent
    |                       |---> Design System Extractor Skill
    |                       |---> WebFetch (website extraction)
    |
    +---> Task Tool ---> BLUE Swarm Agents
    |                       |---> Read tool (transcripts, docs)
    |                       |---> WebFetch (YouTube, articles)
    |
    +---> Task Tool ---> YELLOW Agent
    |                       |---> LinkedIn API (network data)
    |
    +---> Task Tool ---> RED Agent
    |                       |---> GoDaddy MCP (domain check)
    |
    +---> Builder Agents
    |       |---> Write tool (HTML artifacts)
    |       |---> NowPage API (publish)
    |
    +---> Supabase (persist boarding pack data)
    |
    +---> Vercel (deploy if needed)
    |
    v
Published Expert OS
9

Success Criteria

▼
<2h
Expert to First Artifacts
<24h
Boarding Pack Complete
<2wk
Target MVP Launch
<10h
Human Hours / Partner
85+
Quality Score / 100
8+/10
Expert Satisfaction
Metric Current (Phase 1) Target (Phase 2) Target (Phase 3)
Expert to first artifacts ~48 hours <4 hours <2 hours
Expert to boarding pack complete ~2 weeks <3 days <24 hours
Expert to MVP launch 8 weeks 4 weeks <2 weeks
Human review hours per partner ~200 hours ~40 hours <10 hours
Artifact quality score (rubric) Baseline TBD >75/100 >85/100
Expert satisfaction Baseline TBD 7+/10 8+/10
Automation coverage ~30% ~70% >90%
10

Risks & Mitigations

▼
Risk Severity Likelihood Mitigation
Agent hallucination in extraction
Agents fabricate facts, frameworks, or quotes that the expert never said		 High Medium Rubric-based validation at Pass 3 catches inconsistencies. Human review gates required before any artifact publishes. All extracted data includes source citations and confidence scores. Double-check system compares generated content against source material.
Sparse inputs from expert
Expert provides minimal material, insufficient for quality extraction		 Medium High Graceful degradation protocol (Section 4). System generates follow-up questions targeting specific gaps. Clear confidence scoring so humans know which data to trust. Gate system prevents advancement with insufficient data.
Over-engineering the swarm
Building too much automation before validating the manual process		 Medium Medium Ship Phase 1 (manual + individual agents) NOW. Learn from Align360. Only build Phase 2 orchestration after Phase 1 validates the workflow. Iterative approach: automate the highest-value steps first.
Framework misrepresentation
Generated artifacts incorrectly represent the expert's methodology		 High Medium Expert review gate required before launch (Gate 4). Framework fidelity rubric scores all methodology references. Terminology glossary ensures correct naming. Expert approves system prompt and knowledge files before deployment.
Agent coordination failures
Swarm agents produce conflicting outputs or duplicate work		 Medium Medium Chief of Staff agent maintains single source of truth (boarding pack). Clear input/output contracts for each agent. Dependency graph ensures correct execution order. Conflict resolution protocol: when outputs disagree, flag for human review.
Cost escalation (token usage)
Multi-agent swarm consumes excessive tokens, making per-partner cost untenable		 Medium Medium Track token usage per agent task from Phase 1. Set token budgets per workflow. Use smaller/faster models for simple extraction tasks. Reserve large models for synthesis and creative generation. Target: total cost per partner under $50 in API usage.
11

Appendix: Align360 as Case Study

▼
First Partner -- Manual Baseline

Samuel Ekeh / Align360 is Athio's first JV expert partner. Every step of the boarding process was performed manually with ad-hoc agent assistance. This section documents what was done, how it would map to the agentic system, and what we learned.

Step (Manual) What Humans Did What Agents Would Do Est. Time Saved
Transcript Analysis Team manually read 3 call transcripts (Will/Derek/Jason discussions about Samuel's boarding, Align360 touchpoint meeting, Will/Jason status call). Extracted key decisions, action items, and framework details by hand. BLUE Swarm Transcript Analyzer processes all 3 simultaneously. Extracts structured data with confidence scores in minutes. Cross-references across all sources in Pass 2. ~15 hours
Source Document Review Manually read Align360 Background & Tools Overview, Governance Document, Knowledge Files (Part 1 & 2), System Prompt v6.1. Synthesized into understanding of Samuel's methodology. BLUE Document Extractor reads all files in parallel. Builds framework taxonomy, terminology glossary, and governance rule set automatically. ~20 hours
Design System Extraction Analyzed A360 logo, branding document (colors: dark #2e3c45, teal #7aa49c, orange #e09b67), Quicksand font. Manually built the design system HTML file with all variables, components, and guidelines. GREEN Agent uses the design-system-extractor skill. Takes logo + branding doc as input, auto-generates the complete design system HTML. Already proven to work -- this skill is built. ~25 hours
Positioning Research Manually researched alignment coaching market, identified competitors, profiled target audience (executives, teams, career changers). Built positioning brief through multiple team discussions. PURPLE Swarm's 7 agents run in parallel: market sizing, competitor scraping, persona building, niche validation, keyword research, trend analysis, and psychology mapping. Produces comprehensive brief in hours. ~60 hours
Boarding Pack Assembly Manually compiled all extraction data into a structured boarding pack. Updated as new information emerged. Tracked completion across multiple documents. Chief of Staff maintains the boarding pack as a living document, auto-updating as each swarm completes. Single HTML file with progress indicators and confidence scores. ~15 hours
Lead Magnet Build Took Samuel's 5-Phase Alignment Model (Design, Career, Integrate, Figures, Legacy) and manually coded an interactive assessment page with scoring logic, brand-consistent design, and lead capture. Lead Magnet Builder agent takes the primary framework from the extraction output and the design system, auto-generates the assessment page. Human reviews and approves. ~30 hours
Launch Roadmap Manually built the interactive HTML roadmap showing all tasks, ownership, and timeline. Created command center for project tracking. Chief of Staff generates the roadmap from the GTM template, customized with partner-specific data from the boarding pack. Published via NowPage. ~20 hours
~185h
Total Manual Hours (Align360)
~8h
Estimated Agent Hours (if automated)
23x
Efficiency Multiplier
7
Artifacts Produced

Key Learnings from Align360 Boarding

  • Design system extraction is a solved problem. The design-system-extractor skill produced a comprehensive design system from brand assets reliably. This validates the GREEN workflow.
  • Transcript analysis is high-value for agents. Reading and extracting structured data from transcripts is exactly what LLMs excel at. This should be fully automated first.
  • Positioning research is the biggest time sink. The 7-track PURPLE workflow consumed the most manual hours. This is where the swarm will deliver the most value.
  • Expert voice calibration requires human judgment. Ensuring generated content truly sounds like the expert (not generic AI) needs human review. Rubrics help, but cannot replace expert approval.
  • The boarding pack pattern works. Having a single living document that all workflows feed into proved essential for coordination. This becomes the Chief of Staff's primary output.
  • 4-gate system provides natural checkpoints. The gate structure (Brand Assets, Extraction Interview, Content Permissions, Offer Approval) maps cleanly to agent workflow outputs. Each gate is a natural human review point.

Align360 Source Material Inventory

Files Processed 
Transcripts (3):
  - 26.3.4 will, derek, jason about brian and samuel boarding...
  - Align360 Touchpoint - March 05.docx
  - 26.3.6 will and jason and align360 and status...

Source Documents (5):
  - Align360_Background_Tools_Overview_v2.md.pdf
  - Align360 Governance Document.docx
  - Align360_System_Prompt_v6.1.md
  - Align360_Knowledge_File_Part1.md
  - Align360_Knowledge_File_Part2.md

Brand Assets (2):
  - A360logo.jpg
  - Align360 Branding & Colors.docx

Generated Artifacts (4):
  - align360-design-system.html (GREEN output)
  - align360-launch-roadmap.html (roadmap)
  - command-center.html (project tracking)
  - lead-magnet (assessment page -- in progress)
LLM

LLM Context Block

Context for AI Agents 
## ATHIO AGENTIC BOARDING SYSTEM -- LLM CONTEXT

### What This Document Is
Product Requirements Document (PRD) for Athio's agentic expert
boarding system. Describes how to automate the process of
onboarding JV expert partners from raw IP to launched Expert OS.

### Key Architecture
- Chief of Staff Agent: Orchestrator that decomposes, assigns, tracks
- PURPLE Swarm (7 agents): Positioning research
- GREEN Agent (1): Brand/design system extraction
- BLUE Swarm (3 agents): Content extraction from transcripts/docs
- YELLOW Agent (1): Network mapping via LinkedIn
- RED Agent (1): Domain name generation + availability
- Builder Agents (5): Lead magnet, offer page, onboarding flow,
  system prompt, knowledge files

### Core Innovation: Recursive Extraction Protocol
4-pass system:
  Pass 1: Raw extraction with confidence scores
  Pass 2: Cross-source pattern recognition
  Pass 3: Rubric generation for QA
  Pass 4: Gap analysis + follow-up question generation

### Gates (Human Review Points)
  Gate 1: Brand Assets uploaded and validated
  Gate 2: Extraction interview processed
  Gate 3: Content permissions granted
  Gate 4: Offer page approved by expert

### Implementation Phases
  Phase 1 (NOW): Manual + individual agents (Align360)
  Phase 2 (Next 2): Semi-automated swarm, humans at gates
  Phase 3 (Partner 4+): Full swarm autonomy, <10h human time

### Success Targets
  - Time-to-first-artifact: <2 hours
  - Boarding pack complete: <24 hours
  - MVP launch: <2 weeks (down from 8)
  - Human hours: <10 (down from ~200)
  - Quality score: >85/100

### Existing Tools
  - Design System Extractor skill (built, at ~/.claude/skills/)
  - NowPage publishing API (operational)
  - Supabase data store (operational)
  - Claude Code CLI (primary interface)
  - MCP servers: Hugging Face, Vercel, Context7, GoDaddy

### When Building for This System
  - Use Athio branding: #1a1a2e navy, #e94560 red accent,
    #533483 purple, #2ecc71 green, Inter font
  - All artifacts are single-file HTML (self-contained)
  - Include LLM instruction blocks in every generated artifact
  - Use confidence scores for all extracted data
  - Maintain the boarding pack as a living document
  - Respect the 4-gate human review system
  - Track timing and token usage for Phase 2 calibration