AI Project Showcase: Career Coach

Document type: AI Project Showcase

Project: Career Coach

Status: Draft

Last updated by Claude Code: April 12, 2026

Populated from: CLAUDE.md, CareerCoach/project.yml, CareerCoach/CareerCoach/App/CareerCoachApp.swift, knowledgebase/subagents/SUBAGENTS-INDEX.md, knowledgebase/subagents/SUBAGENT-TEMPLATE.md, knowledgebase/subagents/meta-worldview.md, knowledgebase/frameworks/career-development-frameworks.md, knowledgebase/user-profiles/peter-westerman-career-profile.md, Career Coach Prompts/Career Coach, Career Coach Prompts/What Color is Your Parachute, documentation/# UPDATE THE CAREER-COACH PROJECT.md, CareerCoach/CareerCoach/Services/LLM/LLMRouter.swift, CareerCoach/CareerCoach/Services/RAG/PromptAssembler.swift, CareerCoach/CareerCoach/Services/RAG/SubagentLoader.swift

Section 1 — Product Overview

1.1 Product name and tagline

Name: Career Coach
Tagline: An AI-powered career guidance app with 68 thought-leader advisor personas (filesystem count as of Apr 12, 2026; SUBAGENTS-INDEX.md still cites 65), multi-LLM synthesis, and personalized RAG — helping professionals navigate career transitions with wisdom from the world’s leading thinkers.
Current status: Beta (Phase 6 complete — authentication, distribution builds, codebase audit)
First commit / project start: Early 2026 (Phase 1 knowledgebase build); native app development through Phase 6

1.2 What it is

Career Coach is a native iOS/macOS SwiftUI application that provides AI-powered career guidance through a curated panel of 68 thought-leader “advisor” personas (the SUBAGENTS-INDEX.md master index still reads “65 built” but three additional persona files have been added to the subdirectories — index is stale). Users select which advisors inform their coaching sessions — from Richard Bolles (What Color Is Your Parachute) to Pema Chödrön (When Things Fall Apart) to Jensen Huang (AI infrastructure) — and receive guidance shaped by those specific thinkers’ frameworks, voices, and perspectives. The app supports multiple LLMs (Claude, Gemini, ChatGPT) with an “All” synthesis mode, budget-controlled RAG with priority-ranked context assembly, Tavily web search, voice I/O, document upload, a 50-question onboarding assessment, and per-user career profiles.

1.3 What makes it meaningfully different

Career Coach does not simply give AI career advice — it channels the intellectual frameworks of 68 real thinkers, organized into 8 groups (career frameworks [10], business strategy [15], negotiation/communication [4], technology/innovation [8], mindfulness/wellness [6], AI thought leaders [5], media/politics [12], arts/architecture [8]). A meta-worldview subagent synthesizes these perspectives into 7 convergent patterns, 7 genuine tensions, and 12 unified principles (synthesis written against the original 65-advisor baseline). The RAG system uses budget-controlled prompt assembly (P0–P6+ priority tiers) that adapts to each LLM’s context window, ensuring the most relevant advisor knowledge and user profile data always fits within token limits.

1.4 Platform and deployment context

Platform: Native iOS 17+ / macOS 14+ app (SwiftUI + SwiftData)
Deployment: Local-first with optional CloudKit sync; macOS DMG distribution and TestFlight/App Store Connect configuration
Primary interface: Multi-tab SwiftUI app — Chat (with advisor picker, LLM selector, voice input), Library (conversation history), Advisors (browser/detail), Companies (612 AI company profiles), Profile (document library, settings)


Section 2 — User Needs and Problem Statement

2.1 Target user

Primary user: Senior professionals and executives navigating career transitions — layoffs, pivots, retirement planning, fractional/consulting transitions
Secondary users: Mid-career professionals seeking development guidance; career coaches using the tool with clients
User environment: Personal iOS/macOS devices; private career coaching sessions; document upload for resume/LinkedIn analysis

2.2 The problem being solved

Senior professionals facing career transitions need guidance that integrates multiple perspectives — strategic, psychological, practical, philosophical — but traditional career coaching is expensive, single-perspective, and not available on-demand. Generic AI chatbots give shallow career advice that lacks the intellectual depth, framework diversity, and personalization that complex career decisions require.

2.3 Unmet needs this addresses

Need How the product addresses it Source of evidence
Multi-perspective career guidance beyond generic AI advice 68 thought-leader advisor personas with distinct frameworks, voices, and coaching approaches; users select which advisors inform each session CLAUDE.md — subagent system; SUBAGENTS-INDEX.md (index says 65; filesystem has 68)
Personalized guidance grounded in the user’s actual career context Per-user career profiles loaded into RAG; 50-question onboarding assessment; document upload (resume, LinkedIn, appraisals) CLAUDE.md — Phase 5 features; PromptAssembler.swift
Synthesis across conflicting expert perspectives Meta-worldview subagent that identifies 7 convergent patterns, 7 tensions with reconciliations, and 12 unified principles from all 65 advisors meta-worldview.md
Framework-based decision support for specific career situations 4 chunked career framework documents (12 frameworks, transition decision model, layoff recovery, exit strategy) knowledgebase/frameworks/

2.4 What users were doing before this existed

Professionals facing career transitions hired executive coaches ($200–500/hour), read multiple career books (Bolles, Duckworth, Newport), used generic AI chatbots for surface-level advice, and consulted their professional networks. No tool combined the depth of book-level career frameworks with AI’s ability to personalize and synthesize across multiple thinkers simultaneously.


Section 3 — Market Context and Competitive Landscape

3.1 Market category

Primary category: AI career coaching / professional development
Market maturity: Emerging (AI career tools exist but none offer multi-advisor persona architecture)
Key dynamics: The executive coaching market is large (~$20B globally ⚡) but expensive and supply-constrained. AI career tools are commoditizing basic advice while creating opportunity for depth-differentiated products. ⚡

3.2 Competitive landscape

Product / Company Approach Strengths Key gap this project addresses Source
ChatGPT / Claude / Gemini (general) ⚡ Generic career advice via prompting Broad knowledge, accessible No persistent advisor personas, no personalized career profile, no framework depth ⚡ Public products
LinkedIn AI features ⚡ Resume/profile optimization, job matching Massive professional data Optimization-focused, not coaching-focused; no multi-perspective guidance ⚡ Public product
BetterUp / CoachHub ⚡ Human coaching platforms Deep human relationships Expensive ($200+/session), limited availability, single-coach perspective ⚡ Public products
Ama (AI coach) ⚡ AI coaching with personality assessment Structured coaching framework Single AI persona, no thought-leader diversity, no RAG knowledge base ⚡ Public product

3.3 Market positioning

Career Coach positions as a “panel of the world’s leading thinkers in your pocket” — the intellectual depth of reading 65 career-relevant books, combined with the personalization of a private coach who knows your career history, and the synthesis capabilities of multi-LLM AI. It occupies the space between generic AI chat and expensive human coaching.

3.4 Defensibility assessment

The product’s primary defensibility is its 65-advisor knowledge base — each persona file represents deep research into a thinker’s biography, frameworks, key works, coaching applications, signature questions, and response style. The meta-worldview synthesis (7 patterns, 7 tensions, 12 principles) is a unique intellectual artifact. The budget-controlled RAG architecture (P0–P6+ priority tiers) and multi-LLM synthesis mode add technical differentiation.


Section 4 — Requirements Framing

4.1 How requirements were approached

Requirements evolved across 6 development phases, documented through iterative Cursor AI sessions and a phased design document (documentation/# UPDATE THE CAREER-COACH PROJECT.md). Each phase built on the previous: Phase 1–2 (knowledgebase), Phase 3 (native app), Phase 4 (multi-LLM + RAG), Phase 5 (shared knowledge, frameworks, assessment), Phase 6 (auth, audit, distribution).

4.2 Core requirements (what it must do)

  1. Provide AI career coaching conversations informed by user-selected thought-leader advisors (68 available; master index stale at 65)
  2. Support multiple LLMs (Claude, Gemini, ChatGPT) with an “All” synthesis mode that queries all three and synthesizes responses
  3. Assemble prompts with budget-controlled RAG that adapts to each LLM’s context window (P0–P6+ priority tiers)
  4. Maintain per-user career profiles with private knowledge (loaded into RAG when authenticated)
  5. Administer a 50-question onboarding assessment across 8 sections, routing responses to the RAG pipeline
  6. Support voice I/O (speech-to-text input, text-to-speech output) and document upload (PDF, DOCX, image OCR)
  7. Persist conversations via SwiftData with optional CloudKit sync across devices
  8. Provide 612 searchable AI company profiles with 5-segment classification
  9. Include O’Reilly-style help system (9 articles, 7 chapters)

4.3 Constraints and non-goals

Hard constraints:

  • iOS 17.0+ / macOS 14.0+, Swift 5.9
  • Requires API keys for Claude, OpenAI, Gemini, and optionally Tavily (user-provided via settings)
  • User profiles are private and per-user (not shared between users)

Explicit non-goals:

  • Not a job board or job-matching service
  • Not a resume writer (analyzes uploaded resumes for coaching context, does not generate them)
  • Not a replacement for licensed therapy or counseling

4.4 Key design decisions and their rationale

Decision Alternatives considered Rationale Evidence source
68 individual thought-leader persona files (not a single “career coach” prompt) Single comprehensive career coaching prompt Multi-advisor architecture lets users choose which perspectives inform their session; creates intellectual depth and diversity CLAUDE.md — subagent system design
Budget-controlled RAG with P0–P6+ priority tiers Fixed-size context window allocation Different LLMs have vastly different context windows (200K Claude vs. 1M Gemini vs. 128K GPT-4); priority tiers ensure the most important context always fits PromptAssembler.swift — priority sections
Multi-LLM with “All” synthesis mode Single LLM (Claude only) Cross-model synthesis surfaces different perspectives and reduces single-model bias; users can compare CLAUDE.md — multi-LLM features; SynthesisService.swift
Native SwiftUI app (not WordPress or web) WordPress plugin, web app Career coaching is private and personal; native app provides better UX, offline capability, device integration (voice, camera for OCR), and data privacy CLAUDE.md — Phase 3 decision
Meta-worldview as synthesized philosophy (not just a collection) Individual advisors only, no synthesis The meta-worldview identifies genuine convergences and tensions across 65 thinkers, creating unique coaching value beyond any individual advisor meta-worldview.md

Section 5 — Knowledge System Architecture

5.1 Knowledge system overview

KB type: File-based subagent persona library + career frameworks + user profiles, loaded via SubagentLoader and assembled by PromptAssembler with budget-controlled priority tiers
Location in repo: knowledgebase/ (subagents, frameworks, user-profiles, embeddings, guardrails, disambiguations)
Estimated size: 68 subagent persona files (filesystem) + meta-worldview + SUBAGENTS-INDEX + SUBAGENT-TEMPLATE + 4 framework documents + 2 user profiles + 612 company profiles; collectively thousands of pages of curated knowledge

5.2 Knowledge system structure


knowledgebase/
├── subagents/
│   ├── SUBAGENTS-INDEX.md            # Master index with status tracking
│   ├── SUBAGENT-TEMPLATE.md          # Template for new subagents
│   ├── meta-worldview.md             # Synthesized philosophy (7 patterns, 7 tensions, 12 principles)
│   ├── career-frameworks/            # 10 subagents
│   ├── business-strategy/            # 15 subagents
│   ├── negotiation-communication/    # 4 subagents
│   ├── technology-innovation/        # 8 subagents
│   ├── mindfulness-wellness/         # 6 subagents
│   ├── ai-thought-leaders/           # 5 subagents
│   ├── media-politics/               # 12 subagents
│   └── arts-architecture/            # 8 subagents
├── frameworks/
│   ├── career-development-frameworks.md         # 12 frameworks (Flower Exercise, GROW, Ikigai, Habit Loop, etc.)
│   ├── career-transition-decision-framework.md  # 3-path decision model (Continue/Fractional/Retire)
│   ├── senior-executive-layoff-recovery-frame.md # Executive layoff recovery
│   └── senior-executive-exit-strategy-framework.md # Proactive exit planning (EXIT framework)
├── company-profiles/               # CSV data for 612 AI companies
├── embeddings/                     # (empty — for future vector search)
├── guardrails/                     # (empty — for future guardrails)
└── disambiguations/                # (empty — for future disambiguation)
CareerCoach/CareerCoach/Resources/
├── KnowledgeBase/                  # Bundled copy of subagent groups for app
└── CompanyProfiles/                # Bundled company profile data

5.3 Knowledge categories

Category Files / format Purpose Update frequency
Thought-leader subagents 68 .md files following SUBAGENT-TEMPLATE.md Advisor personas with biography, frameworks, key works, coaching application, system prompt Built across 8 sessions + gap-filling + 3 post-index additions; master index still reads “65 built”
Meta-worldview meta-worldview.md Synthesized philosophy from all 65 subagents (7 patterns, 7 tensions, 12 principles) Updated when subagents are added
Career frameworks 4 .md files (chunked for RAG) Shared decision frameworks for career transitions, layoff recovery, exit strategy Stable; new frameworks added as needed
User profiles 2 .md files (private per-user) Personalized career context loaded into RAG when user is authenticated Updated by users / curated from career data
Company profiles CSV / bundled data (612 companies) AI company intelligence with 5-segment classification Periodically updated
Assessment responses SwiftData (runtime) 50-question onboarding assessment routed to PromptAssembler Per-user, one-time with option to retake

5.4 How the knowledge system was built

The subagent library was built across 8 dedicated Cursor AI sessions (documented in CLAUDE.md), each focused on a thematic group. For each thinker, a deep-research persona file was created containing biography, intellectual context, 3–5 core frameworks, key works with career coaching relevance, coaching application guidance, signature questions, response style descriptors, and an injectable system prompt. The meta-worldview was synthesized by analyzing all 65 subagents (original cohort) for convergent patterns and genuine tensions. A gap-filling session (Phase 2) added 12 additional subagents. Three further subagents were added after SUBAGENTS-INDEX.md was last refreshed, bringing the filesystem total to 68 while the index still reads “65 built”. Career frameworks were derived from the subagent content and organized as standalone RAG-injectable documents.

5.5 System prompt and agent configuration

System prompt approach: Dynamic prompt assembly via PromptAssembler.swift with P0–P6+ priority tiers. P0 = system identity and constraints; P1 = selected advisor system prompts; P2 = user career profile; P3 = assessment data; P4 = company profiles; P5 = framework documents; P6 = Tavily web search results. Each tier is included or trimmed based on remaining token budget for the active LLM.
Key behavioural guardrails: Not a therapist or licensed counselor; advises seeking professional help for mental health concerns; does not guarantee career outcomes; transparently identifies which advisor perspectives are informing the response.
Persona / tone configuration: Each of the 68 subagent files defines a distinct voice, tone, vocabulary, and metaphor style. The meta-worldview uses a synthesized “collective wisdom” voice. Users choose which advisor(s) shape the conversation.
Tool use / function calling: Tavily web search for real-time information; no LLM function calling — context is assembled into prompts server-side (client-side in Swift).


Section 6 — Build Methodology

6.1 Development approach

Developed across 6 phases using Cursor AI-assisted sessions, with the knowledge system (65 subagents + meta-worldview) built before the application code. The native SwiftUI app was built on XcodeGen for multi-platform targeting (iOS + macOS). Each phase had a clear scope: knowledge (1–2), app (3), multi-LLM + RAG (4), personalization + frameworks (5), auth + distribution (6).

6.2 Build phases

Phase Approximate timeframe What was built Key commits or milestones
Phase 1 Early 2026 53 thought-leader subagent personas across 8 thematic sessions Subagent library foundation
Phase 2 Early 2026 12 gap-filling subagents + meta-worldview synthesis 65-advisor panel with unified philosophy (later grew to 68 post-index)
Phase 3 Early 2026 Native iOS/macOS SwiftUI app with SwiftData, multi-tab UI, conversation management Working app with chat interface
Phase 4 Early 2026 Multi-LLM support (Claude, Gemini, ChatGPT, All-synthesis), LLMRouter, budget-controlled PromptAssembler Cross-model synthesis capability
Phase 5 March 2026 Shared vs. isolated knowledge architecture, 4 career framework documents, per-user profiles, 50-question assessment, Tavily web search Personalization and framework depth
Phase 6 March 2026 Gmail-based authentication, codebase alignment audit, TestFlight/App Store Connect config, universal build script, macOS DMG distribution Distribution-ready

6.3 Claude Code / AI-assisted development patterns

The CLAUDE.md (360 lines) provides comprehensive project context including directory structure, development status by phase, complete subagent inventory, user profiles, app features, and n8n workflow integration details. Each subagent was created through dedicated Cursor AI research sessions. The app architecture follows patterns from the ITI Shared Library (RAG architecture, prompt assembly, budget control). Development conversations are preserved in cursor-chats/.

6.4 Key technical challenges and how they were resolved

Challenge How resolved Evidence
Context window limits across LLMs with vastly different sizes (128K–1M tokens) Budget-controlled PromptAssembler with P0–P6+ priority tiers that adapts per LLM; uses maxContextTokens / 2 for prompt budget PromptAssembler.swift, LLMProvider.swift
Loading 65 subagent files efficiently at runtime SubagentLoader reads from app bundle; caches parsed subagent data; lazy loading per user selection SubagentLoader.swift
Synthesizing responses from 3 different LLMs in “All” mode SynthesisService queries Claude, Gemini, and ChatGPT in parallel, then uses a follow-up prompt to synthesize a unified response SynthesisService.swift
Per-user private knowledge that must not leak between users Authentication gate (RootView); user profiles loaded into RAG only when the authenticated user matches; SwiftData model isolation CLAUDE.md — Phase 5/6 features

Section 7 — AI Tools and Techniques

7.1 AI models and APIs used

Model / API Provider Role in product Integration method
Claude Sonnet 4.6 (claude-sonnet-4-6) Anthropic Primary career coaching LLM; default model Direct API (https://api.anthropic.com/v1/messages) via ClaudeService.swift
GPT-4o OpenAI Alternative LLM for coaching conversations Direct API (https://api.openai.com/v1/chat/completions) via OpenAIService.swift
Gemini 2.0 Flash Google Alternative LLM with 1M token context window Direct API via GeminiService.swift
Tavily Search Tavily Real-time web search for current career market data Direct API (https://api.tavily.com/search) via TavilyService.swift
n8n (optional) Self-hosted Workflow routing for Claude non-streaming requests HTTP webhook (localhost:5678/webhook/career-coach) via N8nWorkflowClient.swift

7.2 AI orchestration and tooling

Tool Category Purpose
PromptAssembler RAG / prompt engineering Budget-controlled prompt assembly with P0–P6+ priority tiers
SubagentLoader Knowledge management Loads and parses subagent persona files from app bundle
LLMRouter Model routing Routes requests to appropriate LLM service based on user selection
SynthesisService Multi-model orchestration Queries multiple LLMs in parallel and synthesizes responses
KnowledgeService RAG Manages shared and per-user knowledge loading
N8nWorkflowClient Workflow automation Optional n8n webhook routing for Claude requests

7.3 Prompting techniques used

  • [x] System prompt persona/role setting (65 distinct advisor personas + meta-worldview)
  • [x] RAG context injection (subagent knowledge, user profiles, assessment data, frameworks, company profiles, web search)
  • [x] Multi-turn conversation management (SwiftData conversation persistence)
  • [x] Structured / JSON output prompting (API response parsing)
  • [x] Chain-of-thought reasoning [CLAUDE NOTE: inferred from advisor framework application guidance]
  • [x] Output guardrails / content filtering (not-a-therapist disclaimers, professional referral guidance)
  • [x] Fallback / error recovery prompting (graceful degradation when APIs unavailable)
  • [ ] Few-shot examples in prompts
  • [x] Tool use / function calling (Tavily web search integration)

7.4 AI development tools used to build this

Tool How used in build
Cursor Primary development IDE; 8+ dedicated subagent research sessions; app architecture and Swift code generation; iterative debugging and codebase audit
Claude (via Cursor) Deep research for 65 subagent personas; meta-worldview synthesis; framework document creation; Swift code generation
Antigravity Autonomous test execution, browser QA, visual regression testing — used per global CLAUDE.md tool lane
XcodeGen Project generation from project.yml for iOS + macOS targets

Section 8 — Version History and Evolution

8.1 Version timeline

Version / Phase Date Summary of changes Significance
Phase 1 Early 2026 53 thought-leader subagent personas (8 thematic sessions) Knowledge base foundation
Phase 2 Early 2026 12 gap-filling subagents + meta-worldview synthesis 65-advisor panel (later 68 with post-index additions)
Phase 3 Early 2026 Native SwiftUI app — iOS 17+/macOS 14+, SwiftData, multi-tab UI Working application
Phase 4 Early 2026 Multi-LLM (Claude, Gemini, ChatGPT, All-synthesis), LLMRouter, PromptAssembler Multi-model architecture
Phase 5 March 2026 Shared/isolated knowledge, 4 framework docs, user profiles, 50-question assessment, Tavily Personalization depth
Phase 6 March 2026 Gmail auth, codebase audit, TestFlight config, build scripts, macOS DMG Distribution-ready

8.2 Notable pivots or scope changes

The project began as a WordPress plugin concept (evidenced by plugin-installs/ directory and WordPress-style project structure) but pivoted to a native iOS/macOS SwiftUI application. This was a significant architectural pivot driven by the need for better UX, offline capability, voice integration, and data privacy for personal career coaching. [CLAUDE NOTE: inferred from directory structure containing both WordPress and native app paths]

8.3 What has been cut or deferred

  • CloudKit sync (configured but set to .none in current app init)
  • Embeddings / vector search (directory exists but is empty — for future RAG enhancement)
  • Content guardrails (directory exists but is empty)
  • Disambiguation rules (directory exists but is empty)
  • WordPress plugin version (deferred in favor of native app)
  • n8n workflow integration is optional / localhost-only (not deployed to production)

Section 9 — Product Artifacts

9.1 Design and UX artifacts

Artifact Path Type What it shows
SwiftUI Views CareerCoach/CareerCoach/Views/ (Chat, Library, Advisors, Companies, Profile, Assessment, Auth, Settings, Mac, Help) SwiftUI source Complete multi-tab app UI
O’Reilly-style help CareerCoach/CareerCoach/Help/ SwiftUI + content 9 articles across 7 chapters
XcodeGen config CareerCoach/project.yml Project config iOS + macOS target definitions
Build scripts CareerCoach/scripts/build-all.sh Shell script Universal build (iOS archive, macOS archive, DMG, notarization)

9.2 Documentation artifacts

Document Path Type Status
CLAUDE.md CLAUDE.md (360 lines) Project context Complete (Phase 6)
Design requirements documentation/# UPDATE THE CAREER-COACH PROJECT.md (321 lines) Phased requirements Complete
Subagent index knowledgebase/subagents/SUBAGENTS-INDEX.md Master index Complete (65 subagents)
Subagent template knowledgebase/subagents/SUBAGENT-TEMPLATE.md Creation template Complete
Meta-worldview knowledgebase/subagents/meta-worldview.md Synthesized philosophy Complete
Career frameworks knowledgebase/frameworks/*.md (4 documents) Framework references Complete

9.3 Data and output artifacts

Artifact Path Description
68 subagent persona files knowledgebase/subagents/*/ Deep-researched thought-leader personas (biography, frameworks, coaching application, system prompts); SUBAGENTS-INDEX.md still lists 65
User career profiles knowledgebase/user-profiles/ Private per-user career context (2 profiles)
612 AI company profiles knowledgebase/company-profiles/ + bundled resources 5-segment classified AI company database
Standalone prompt files Career Coach Prompts/ Career Discovery Navigator, WCYP persona, Portfolio Strategy prompts
Career Coach Exports Career Coach Exports/ Generated portfolio positioning documents

Section 10 — Product Ideation Story

10.1 Origin of the idea

Career Coach was born from the insight that effective career guidance requires not one perspective but many — and that AI can synthesize diverse expert viewpoints in ways that no single human coach can. The project was inspired by the real career transitions of its initial users (senior B2B media executives) who needed guidance integrating strategic, psychological, and practical perspectives. Rather than building a generic AI career chatbot, the approach was to create a “panel of advisors” architecture where each advisor represents the authentic intellectual framework of a real thinker. [CLAUDE NOTE: inferred from user profiles and product architecture]

10.2 How the market was assessed

Research approach used: Analysis of existing AI career tools, executive coaching pricing models, and career book landscape; identification of the gap between expensive human coaching and shallow AI advice.
Key market observations:

  1. Executive coaching costs $200–500/hour and is single-perspective
  2. Career books contain deep frameworks but cannot personalize or synthesize across authors
  3. Generic AI chatbots give career advice without framework depth or persistent user context
  4. No product existed that combined multi-author frameworks with AI personalization

What existing products got wrong:
Existing AI career tools either replicate a single coaching methodology (limiting) or use generic prompts (shallow). They treat career advice as a Q&A problem rather than a multi-perspective synthesis challenge.

10.3 The core product bet

If professionals can access a panel of 65 thought-leader advisors — each with authentic frameworks, voice, and coaching approach — through a multi-LLM AI app that knows their personal career context, they will get career guidance that is both deeper and more personalized than either human coaching or generic AI chat alone.

10.4 How the idea evolved

The project started with the knowledge system (Phase 1–2): 65 subagent persona files built across 8 dedicated research sessions, followed by the meta-worldview synthesis. Only after the knowledge base was complete did app development begin (Phase 3). This “knowledge-first” approach ensured the product’s intellectual depth was established before technical implementation. The app evolved through multi-LLM support (Phase 4), personalization depth with frameworks and assessment (Phase 5), and authentication and distribution (Phase 6). The pivot from WordPress to native SwiftUI reflected a design decision that career coaching deserves a private, high-quality native experience.


Section 11 — Lessons and Next Steps

11.1 Current state assessment

What works well: 68 deeply-researched advisor personas with authentic frameworks and voices; meta-worldview synthesis creating unique philosophical artifact; multi-LLM with synthesis mode; budget-controlled RAG with priority tiers; per-user career profiles; 50-question onboarding assessment; native iOS/macOS app with voice I/O and document upload; O’Reilly-style help system.
Current limitations: CloudKit sync disabled (set to .none); embeddings/vector search not implemented (directories empty); guardrails and disambiguation directories empty; API keys have fallback to hardcoded values in APIKeys.swift (security concern — should be Keychain-only); n8n integration is localhost-only.
Estimated completeness: Phase 6 of 6+ phases — ~85% feature-complete; distribution-ready but pre-public-release.

11.2 Visible next steps

  1. Refresh SUBAGENTS-INDEX.md to reflect the current 68 subagent files (currently stale at 65)
  2. Remove hardcoded API key fallbacks from APIKeys.swift (enforce Keychain-only)
  3. Enable CloudKit sync for cross-device conversation persistence
  4. Implement vector/embedding search for semantic subagent and framework retrieval
  5. Build content guardrails and disambiguation rules
  6. Deploy n8n workflow routing to production (not just localhost)
  7. Add more user profiles and expand the career framework library
  8. App Store / TestFlight public beta release

11.3 Lessons learned

_Manual input required — this section cannot be populated automatically._


Section 12 — Claude Code Validation Checklist

  • [x] Every placeholder has been replaced or marked NOT FOUND
  • [x] All externally-sourced competitive data is marked with ⚡
  • [x] All inferences are marked with [CLAUDE NOTE]
  • [x] Version history is derived from CLAUDE.md phase documentation
  • [x] Knowledge system paths reflect real directory structure
  • [x] AI tools are confirmed from code/config, not guessed
  • [x] Section 11.3 is left blank for manual input
  • [x] Document header shows today’s date and files examined

Sources Examined

File / Path What it contributed
CLAUDE.md (360 lines) Sections 1–10 — comprehensive project context, directory structure, development phases, subagent inventory, user profiles, app features
CLAUDE.md.backup (39 lines) Section 8 — earlier version for comparison
CareerCoach/project.yml (107 lines) Section 1, 4 — version numbers, platform targets, Swift/iOS/macOS versions
CareerCoach/CareerCoach/App/CareerCoachApp.swift (143 lines) Section 6 — app initialization, SwiftData schema, CloudKit config
knowledgebase/subagents/SUBAGENTS-INDEX.md Section 5 — master subagent index with 65 entries
knowledgebase/subagents/SUBAGENT-TEMPLATE.md Section 5 — subagent creation template
knowledgebase/subagents/meta-worldview.md Sections 1, 5 — synthesized philosophy (7 patterns, 7 tensions, 12 principles)
knowledgebase/frameworks/career-development-frameworks.md Section 5 — 12 career frameworks
knowledgebase/frameworks/career-transition-decision-framework.md Section 5 — 3-path decision model
knowledgebase/user-profiles/peter-westerman-career-profile.md Section 5 — user profile example
Career Coach Prompts/Career Coach (109 lines) Section 5 — Career Discovery Navigator prompt
Career Coach Prompts/What Color is Your Parachute (91 lines) Section 5 — WCYP executive-coach persona
Career Coach Prompts/Update Portfolio Positioning Strategy.md (85 lines) Section 9 — portfolio strategy prompt
documentation/# UPDATE THE CAREER-COACH PROJECT.md (321 lines) Section 4, 6 — phased requirements and design decisions
CareerCoach/CareerCoach/Services/LLM/LLMRouter.swift Section 7 — LLM routing architecture
CareerCoach/CareerCoach/Services/LLM/ClaudeService.swift Section 7 — Claude API integration
CareerCoach/CareerCoach/Services/LLM/OpenAIService.swift Section 7 — OpenAI API integration
CareerCoach/CareerCoach/Services/LLM/GeminiService.swift Section 7 — Gemini API integration
CareerCoach/CareerCoach/Services/LLM/SynthesisService.swift Section 7 — multi-LLM synthesis
CareerCoach/CareerCoach/Services/LLM/LLMProvider.swift Section 7 — model definitions, context window sizes
CareerCoach/CareerCoach/Services/RAG/PromptAssembler.swift Section 5, 7 — budget-controlled prompt assembly
CareerCoach/CareerCoach/Services/RAG/SubagentLoader.swift Section 5, 7 — subagent file loading
CareerCoach/CareerCoach/Services/RAG/TavilyService.swift Section 7 — Tavily web search
CareerCoach/CareerCoach/Services/RAG/N8nWorkflowClient.swift Section 7 — n8n workflow routing

Addendum — April 2026 Competitive Landscape and Roadmap Update

1. Industry Context

The AI career coaching market reached $6.69B in 2026 with a 22.3% CAGR, and the flood of vibe-coded career tools reflects that growth. Any developer with Cursor or Bolt.new can now build a “career coaching chatbot” in an afternoon — and many have. The result is a market saturated with thin wrappers over ChatGPT that offer the same generic career advice. Reddit threads consistently surface the same complaint: “ChatGPT validates plans rather than providing honest, challenging advice.” The irony of LLM convergence in this space is that as models get better at general career reasoning, the baseline quality of generic career chatbots improves — but the ceiling of what they can offer without structured domain knowledge and persistent user context remains low.

Career Coach’s architecture was designed around a different bet: that career guidance is a multi-perspective synthesis problem, not a Q&A problem. The 68 thought-leader advisor personas, each with researched frameworks, coaching approaches, and authentic voice, create intellectual depth that cannot be replicated by prompting a frontier model to “act like a career coach.” The meta-worldview synthesis (7 convergent patterns, 7 genuine tensions, 12 unified principles derived from all advisors) is a unique intellectual artifact. No competitor has attempted this.

However, the citizen developer wave has produced competitors that are faster to market on practical features. While Career Coach invested deeply in knowledge architecture, products like TopCV.io shipped four specialized AI coaches with cross-session memory, Melzi shipped mock interviews with scoring in 15 languages, and Careerflow shipped skills gap analysis plus job tracking in a single platform. The next phase for Career Coach requires closing these practical feature gaps while preserving and extending the multi-advisor depth that remains unmatched.

2. Competitive Landscape Changes

Eight significant competitors have entered or expanded since Career Coach’s development began:

Competitor Category Key Feature Career Coach Lacks Pricing
BetterUp Grow Enterprise AI coaching Slack/Teams integration, 17M data points, blended AI + human coaching Enterprise
Valence Nadia Enterprise AI manager coaching Real-time flow-of-work coaching, 45+ enterprise deployments Enterprise
TopCV.io Consumer AI coaching 4 specialized coaches, cross-session memory, interview + salary + review coaching $29/mo
Cruit Consumer AI career platform Unified context, resume inline editing, Sankey job flow visualization Not listed
Career Compass AI (iOS) Native iOS coaching Growth plan builder, AI weekly coaching emails, career KPI tracking $9.99/mo
Melzi Job Coach (iOS) Native iOS job search Mock interviews with scoring, ATS resume optimization, 15 languages, on-device processing Free (limited)
ApplyArc Consumer AI career platform Daily briefings, mock interviews, rejection analysis, salary negotiation scripts £19/mo
Final Round AI Consumer AI job search Auto-apply, interview coaching Not listed

Features that are now table stakes (Career Coach is missing all of them):

Feature Competitors Who Have It
Skills gap analysis (current skills vs. target role) Careerflow, Cruit, LinkedIn Career Explorer, Teal, ApplyArc
Mock interview practice with scoring/feedback Melzi, TopCV.io, Final Round AI, ApplyArc, Careerflow
Salary negotiation coaching with market benchmarks TopCV.io, Careerflow, ApplyArc, Career Compass AI
Job application tracking (Kanban/dashboard) Huntr, Teal, Careerflow, Cruit, ApplyArc
Resume tailoring to specific job descriptions Jobscan, Teal, Careerflow, Cruit, Kickresume

Differentiators that remain unique: 68 curated thought-leader personas (no competitor offers this), meta-worldview synthesis, multi-advisor panel selection per session, budget-controlled RAG with P0–P6+ priority tiers adapting per LLM context window, multi-LLM “All” synthesis mode, and 612 AI company profiles with 5-segment classification.

3. Our Competitive Response: Product Roadmap

The Tier 1 priorities balance closing parity gaps with exploiting the multi-advisor advantage through a new interaction mode.

Tier 1 — Critical (next build cycle):

Skills Gap Analyzer (M effort) — compares user skills from the assessment, career profile, and uploaded resume against target role requirements. Surfaces specific gaps with advisor-informed recommendations (Newport on deep work skill-building, Drucker on feedback analysis). Uses Tavily for real-time job market data. This addresses the most-cited missing feature across competitor comparisons.

Advisor Debate Mode (M effort) — the highest-value new interaction pattern. Users pose a career question and select 2-3 advisors; each responds in character, then a synthesis identifies convergences, tensions, and recommended action. This leverages the existing multi-LLM synthesis architecture while creating a fundamentally different experience from any competitor’s single-perspective advice. Reddit users explicitly complain that “ChatGPT validates plans rather than providing honest, challenging advice” — Debate Mode directly addresses this.

Mock Interview Module (L effort) — behavioral (STAR framework) and situational questions with advisor-persona feedback (Carnegie on interpersonal style, Porter on strategic framing). Includes scoring rubric and conversation saved to library. This is table stakes: Melzi, TopCV.io, Final Round AI, ApplyArc, and Careerflow all offer it.

Three housekeeping items (all S effort): reconcile SUBAGENTS-INDEX.md from 65 to 68 entries, remove hardcoded API key fallbacks from APIKeys.swift (enforce Keychain-only), and enable CloudKit sync for cross-device conversation persistence.

Tier 2 — High Value: Salary Negotiation Coach (integrating the Fisher/Ury/Patton, Nierenberg, and Stiglitz subagents with market data), Job Application Tracker (SwiftData-backed Kanban), Career Decision Journal (framework-based structured entries), Resume Tailoring Assistant, Advisor Recommendation Engine, and vector/embedding search for semantic subagent retrieval.

Tier 3 — Strategic: WordPress Plugin MVP (bringing the advisor-panel experience to WordPress via ITI Shared Library), Cross-Advisor Synthesis Reports (downloadable PDFs), Proactive Coaching Nudges, Network Strategy Module, LinkedIn Profile Optimizer, and Content Guardrails System.

Tier 4 — Exploratory: Advisor Voice Personas, Career Scenario Simulator, Multi-User Expansion, Employer/Coach B2B License, Apple Watch Companion.

The dependency logic: Skills Gap Analyzer requires the existing assessment and profile infrastructure (already built). Debate Mode extends the multi-LLM synthesis architecture (already built). Mock Interview is self-contained. Tier 2 items build on Tier 1 foundations — Salary Negotiation Coach leverages the same advisor-persona injection pattern as Debate Mode, and the Job Application Tracker feeds data into the Resume Tailoring Assistant.

4. New Capabilities Added Since Last Build

Two new Skills from the April 2026 roadmap cycle directly support Career Coach development:

Skill What It Provides
interview-coaching-design Design patterns for AI-powered mock interview experiences: behavioral/technical question generation, real-time feedback loops, scoring rubrics, STAR framework integration, and difficulty progression. Directly enables the Tier 1 Mock Interview Module.
salary-negotiation-frameworks Structured negotiation frameworks (BATNA, principled negotiation, information asymmetry, anchoring) applied to compensation discussions with market data integration. Negotiation script generation, offer evaluation matrices. Directly enables the Tier 2 Salary Negotiation Coach by operationalizing the Fisher/Ury/Patton, Nierenberg, and Stiglitz subagents.

Additionally, the existing career-assessment Skill provides the foundation for the Skills Gap Analyzer, and the kanban-board-builder Skill supports the Job Application Tracker implementation.

5. Honest Assessment

Strengths: Career Coach’s 68-advisor knowledge base is a genuine intellectual asset that would take months to replicate — each persona file represents deep research into a thinker’s biography, frameworks, key works, coaching applications, and response style. The meta-worldview synthesis (7 patterns, 7 tensions, 12 principles) is a unique artifact no competitor has attempted. The budget-controlled RAG architecture (P0–P6+ priority tiers adapting per LLM’s context window) is technically sophisticated. The multi-LLM “All” synthesis mode creates cross-model perspective diversity. The 50-question onboarding assessment and persistent user profiles give Career Coach deeper user context than any competitor except BetterUp Grow (which serves enterprises, not individuals).

Gaps: The product has zero practical career action tools — no skills gap analysis, no mock interviews, no salary negotiation coaching, no job tracking, no resume tailoring. These are table stakes across 5+ competitors. The native iOS/macOS platform limits distribution compared to web-based competitors. CloudKit sync is disabled. API keys have hardcoded fallbacks (security concern). The SUBAGENTS-INDEX.md is stale (reads 65, filesystem has 68). The product has deep knowledge but shallow practical utility for users who need to act on career decisions, not just think about them.

What we’re watching: BetterUp Grow’s Slack/Teams integration puts coaching in the flow of work — a distribution model native apps cannot match. TopCV.io’s four specialized coaches with cross-session memory show that persona-based coaching has commercial traction, validating Career Coach’s direction at scale. The WordPress plugin surface (Tier 3) could open a B2B distribution channel (career centers, coaching practices, universities) that no iOS competitor can access.

Career Coach demonstrates ITI’s ability to build knowledge-intensive AI products where the value is in the curated domain expertise, not the AI model. The competitive landscape confirms the multi-advisor approach is differentiated; the roadmap addresses the practical utility gap that separates a compelling prototype from a product people use.