diff --git a/.serena/.gitignore b/.serena/.gitignore
deleted file mode 100644
index 14d86ad..0000000
--- a/.serena/.gitignore
+++ /dev/null
@@ -1 +0,0 @@
-/cache
diff --git a/CHANGELOG.md b/CHANGELOG.md
deleted file mode 100644
index 0abb295..0000000
--- a/CHANGELOG.md
+++ /dev/null
@@ -1,136 +0,0 @@
-# CHANGELOG
-
-<!-- version list -->
-
-## v0.1.0 (2025-11-29)
-
-- Initial Release
-
-## v0.4.0 (2025-09-10)
-
-### Features
-
-- Add production logging system + project structure improvements
-  ([#6](https://github.com/ai-enhanced-engineer/ai-base-template/pull/6),
-  [`ed20bb5`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/ed20bb5a4b9e6c1f5a50379bf7e32c9406c40f22))
-
-### Refactoring
-
-- Modernize and streamline logging system with functionality-driven tests
-  ([#6](https://github.com/ai-enhanced-engineer/ai-base-template/pull/6),
-  [`ed20bb5`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/ed20bb5a4b9e6c1f5a50379bf7e32c9406c40f22))
-
-- Modernize logging module and simplify tests
-  ([#6](https://github.com/ai-enhanced-engineer/ai-base-template/pull/6),
-  [`ed20bb5`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/ed20bb5a4b9e6c1f5a50379bf7e32c9406c40f22))
-
-- Rename ai_base_template/ to src/ and integrate formatting
-  ([#6](https://github.com/ai-enhanced-engineer/ai-base-template/pull/6),
-  [`ed20bb5`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/ed20bb5a4b9e6c1f5a50379bf7e32c9406c40f22))
-
-
-## v0.3.0 (2025-09-02)
-
-### Documentation
-
-- Showcase production logging system in README
-  ([#5](https://github.com/ai-enhanced-engineer/ai-base-template/pull/5),
-  [`75ab7ad`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/75ab7ad3fa027447c2b360fece807fac2d737141))
-
-- Transform README to production-first pedagogical approach
-  ([#5](https://github.com/ai-enhanced-engineer/ai-base-template/pull/5),
-  [`75ab7ad`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/75ab7ad3fa027447c2b360fece807fac2d737141))
-
-- Transform README to production-first pedagogical approach
-  ([#4](https://github.com/ai-enhanced-engineer/ai-base-template/pull/4),
-  [`7f109e7`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/7f109e7ba6a968d7b67caed39f58d0a6c3fdf83e))
-
-### Features
-
-- Add structured logging and integrate formatting into CI/CD
-  ([#5](https://github.com/ai-enhanced-engineer/ai-base-template/pull/5),
-  [`75ab7ad`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/75ab7ad3fa027447c2b360fece807fac2d737141))
-
-### Refactoring
-
-- Improve import organization and remove redundant docstrings
-  ([#5](https://github.com/ai-enhanced-engineer/ai-base-template/pull/5),
-  [`75ab7ad`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/75ab7ad3fa027447c2b360fece807fac2d737141))
-
-
-## v0.2.1 (2025-08-21)
-
-### Bug Fixes
-
-- Add sync target and update CI workflows
-  ([#3](https://github.com/ai-enhanced-engineer/ai-base-template/pull/3),
-  [`95cb140`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/95cb140b7bc7f2a619bbe0731b96129ccf42a122))
-
-### Chores
-
-- Improve Makefile interface and consistency
-  ([#3](https://github.com/ai-enhanced-engineer/ai-base-template/pull/3),
-  [`95cb140`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/95cb140b7bc7f2a619bbe0731b96129ccf42a122))
-
-### Refactoring
-
-- Improve Makefile with better naming and consistency
-  ([#3](https://github.com/ai-enhanced-engineer/ai-base-template/pull/3),
-  [`95cb140`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/95cb140b7bc7f2a619bbe0731b96129ccf42a122))
-
-
-## v0.2.0 (2025-08-06)
-
-### Chores
-
-- Add pre-commit configuration with Make commands
-  ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-### Documentation
-
-- Add Architecture Decision Record (ADR)
-  ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-- Add Architecture Decision Record (ADR) for project structure
-  ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-### Features
-
-- Add Apache 2.0 license ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-### Refactoring
-
-- Use Make commands in CI workflow for consistency
-  ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-### Testing
-
-- Verify pre-commit hooks are working
-  ([#2](https://github.com/ai-enhanced-engineer/ai-base-template/pull/2),
-  [`38a8fcc`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/38a8fccd8e882747ec8879ed1c0afd5db9f77ded))
-
-
-## v0.1.0 (2025-08-06)
-
-### Features
-
-- Add Apache 2.0 license ([#1](https://github.com/ai-enhanced-engineer/ai-base-template/pull/1),
-  [`28bc302`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/28bc3025d31029824b72586ce6b41fa6c128844c))
-
-
-## v0.0.2 (2025-07-31)
-
-### Bug Fixes
-
-- Comment out ML dependencies to speed up CI builds
-  ([`ab0c011`](https://github.com/ai-enhanced-engineer/ai-base-template/commit/ab0c0114d70f745460f532603a4cd90a52d9f00d))
-
-
-## v0.0.1 (2025-07-31)
-
-- Initial Release
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..f055a9a
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,34 @@
+# Python Agentic Template
+
+Autonomous Python template. Describe what you want; agents build it.
+
+## Workflow
+
+Run `workflows/PROJECT_INIT_WORKFLOW.md` for the complete specification.
+
+Phases: **Research → Architecture → MVP → Enhancement** with user approval gates.
+
+## Development Standards
+
+### Validation (Required Before Commits)
+```bash
+make validate-branch   # Runs format, lint, type-check, test
+```
+
+## Artifact Locations
+
+| Output | Location |
+|--------|----------|
+| User seeds | `context/PRODUCT.md`, `context/ENGINEERING.md` |
+| PRD | `context/PRD.md` |
+| Architecture | `ADR.md` |
+| Project plan | `context/PROJECT_PLAN.md` |
+| Code | `src/*.py` |
+| Tests | `tests/test_*.py` |
+
+## Key Rules
+
+1. Read all context files before acting
+2. Use templates in `workflows/templates/` for output formats
+3. Stop at approval gates—don't jump phases
+4. Max 3 fix cycles per deliverable, then escalate
diff --git a/README.md b/README.md
index 1aaf87a..ca5e1b5 100644
--- a/README.md
+++ b/README.md
@@ -2,13 +2,15 @@
 
 > Describe what you want to build. Let agents build it.
 
-Autonomous multi-agent Python project template. *Part of [Bot Brewers](https://github.com/bot-brewers).*
+An autonomous Python project template that bootstraps itself into production-ready code through a multi-agent workflow.
 
 ## Why This Template?
 
-**The Problem:** Starting AI/ML projects requires extensive setup—architecture decisions, project structure, testing patterns, CI/CD, logging, and more. Most developers copy-paste from old projects or spend days configuring from scratch.
+Starting an AI/ML project means three days of ceremony before writing real code. Project structure. Logging. Tests. CI/CD. Type hints. Pre-commit hooks.
 
-**The Solution:** This template **bootstraps itself** into a complete, production-ready project through a multi-agent workflow. You describe your project in plain language; agents research, plan, and build it.
+Most developers copy-paste from old repos or configure from scratch. Both approaches carry technical debt from day one.
+
+This template bootstraps itself. Describe the project in two seed files; agents research, plan, and implement it with production patterns built in.
 
 ## How It Works
 
@@ -54,14 +56,14 @@ Autonomous multi-agent Python project template. *Part of [Bot Brewers](https://g
 └─────────────────────────────────────────────────────────────────┘
 ```
 
-**Human-in-the-loop**: You approve each phase before continuing. No runaway automation.
+You approve each phase before continuing.
 
 ## Quick Start
 
 1. **Create your repository**: Click "Use this template" on GitHub
 2. **Set up environment**: `make init`
 3. **Fill out your seeds**: Edit `context/PRODUCT.md` and `context/ENGINEERING.md`
-4. **Start brewing**: In Claude Code, say `"Run the project initialization workflow"`
+4. **Start building**: In Claude Code, say `"Run the project initialization workflow"`
 
 ### Filling Out Seeds
 
@@ -79,32 +81,26 @@ See `workflows/PROJECT_INIT_WORKFLOW.md` for the complete workflow specification
 
 ## What You Get
 
-Beyond the autonomous workflow, this template provides a **production-ready foundation**:
+The autonomous workflow is the main attraction, but the template also provides a production-ready foundation:
 
 ### Modern Python Tooling
-- Python 3.12+, FastAPI, Pydantic
-- Type hints throughout
+- Python 3.12+ with type hints throughout
 - uv for fast dependency management
+- Ready for any framework you choose
 
 ### Production Logging
-- Structured JSON logging with structlog
-- Correlation ID tracking across requests
-- Dual-mode: human-readable (dev) / JSON (prod)
+Structured JSON logging with structlog. Correlation ID tracking. Dual-mode output: human-readable for development, JSON for production.
 
 ### Development Automation
-- Pre-configured linting (Ruff), formatting (Black), type checking (mypy)
-- Pre-commit hooks for quality gates
-- `make validate-branch` runs all checks
+- Ruff (linting), Black (formatting), mypy (types)
+- Pre-commit hooks as quality gates
+- `make validate-branch` runs everything
 
-### Testing Patterns
-- Unit, functional, and integration test structure
-- pytest with markers for test organization
-- 21+ logging system tests included as examples
+### Testing Structure
+Unit, functional, and integration test directories with pytest markers. The logging system alone has 21+ tests as examples.
 
 ### CI/CD Ready
-- GitHub Actions workflows
-- Semantic versioning
-- Docker-ready structure
+GitHub Actions workflows included. Semantic versioning. Docker-ready structure.
 
 ## Project Structure
 
@@ -166,14 +162,9 @@ The autonomous workflow ensures these patterns are built in from the start, not
 
 ## Who Should Use This
 
-### Teams Starting AI/ML Projects
-Stop reinventing infrastructure. Describe your project and let agents build a production-ready foundation.
-
-### Senior Engineers New to AI
-Get the safety rails you're accustomed to in production systems while learning AI concepts.
+Starting an AI/ML project and don't want to spend days on infrastructure? This is for you. Senior engineer who expects production-grade tooling from day one? This is for you. Leading a team and want a consistent starting point that embodies engineering discipline? This is for you.
 
-### Technical Leaders
-Give your team a consistent, production-ready starting point that embodies engineering best practices.
+The common thread: reliability over speed, describing what you want over configuring it manually.
 
 ## Learn More
 
@@ -187,18 +178,14 @@ Give your team a consistent, production-ready starting point that embodies engin
 - [Hidden Technical Debt in ML Systems](https://papers.nips.cc/paper/5656-hidden-technical-debt-in-machine-learning-systems.pdf)
 
 ### Technologies
-- [FastAPI](https://fastapi.tiangolo.com/) - Modern Python web framework
-- [Pydantic](https://docs.pydantic.dev/) - Data validation
 - [structlog](https://www.structlog.org/) - Structured logging
 - [uv](https://docs.astral.sh/uv/) - Fast Python package management
+- [pytest](https://docs.pytest.org/) - Testing framework
+- [Ruff](https://docs.astral.sh/ruff/) - Fast Python linter
 
 ## Contributing
 
-When contributing, prioritize:
-1. **Reliability over features**
-2. **Simplicity over cleverness**
-3. **Documentation over assumptions**
-4. **Tests over trust**
+Prioritize reliability over features, simplicity over cleverness, documentation over assumptions, tests over trust.
 
 ## License
 
diff --git a/research/EDA.ipynb b/research/EDA.ipynb
deleted file mode 100644
index 91e9588..0000000
--- a/research/EDA.ipynb
+++ /dev/null
@@ -1,55 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# EDA - Exploratory Data Analysis\n",
-    "\n",
-    "This notebook will be continued soon..."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": "# Import production code\nimport sys\nfrom pathlib import Path\n\nsys.path.append(str(Path().parent.absolute()))\n\nimport src\nfrom src.main import get_version, hello_world"
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": "# Explore the package\nprint(f\"Version: {src.__version__}\")\nprint(f\"Hello: {hello_world()}\")\nprint(f\"Get version: {get_version()}\")"
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.7"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}
\ No newline at end of file
diff --git a/workflows/PROJECT_INIT_WORKFLOW.md b/workflows/PROJECT_INIT_WORKFLOW.md
index c556ceb..50d0e69 100644
--- a/workflows/PROJECT_INIT_WORKFLOW.md
+++ b/workflows/PROJECT_INIT_WORKFLOW.md
@@ -24,6 +24,8 @@ This workflow uses **abstract roles** instead of specific agent names for portab
 | `architecture` | ADRs, system design, MVP scoping | 2 | project-architect, ml-systems-architect |
 | `implementation` | Code writing, testing, domain expertise | 3-4 | ai-engineer, backend-engineer, data-engineer, agentic-engineer |
 | `review` | Test enforcement, code quality auditing | 3-4 | python-code-quality-auditor |
+| `domain-expert` | Business alignment, PRD compliance, user value validation | 3-4 feedback | Inferred from PRODUCT.md domain; product-manager as fallback |
+| `architecture-reviewer` | Over-engineering detection, ADR compliance, pattern validation | 3-4 feedback | project-architect, ml-systems-architect |
 
 **Role Matching Keywords**:
 
@@ -33,6 +35,8 @@ This workflow uses **abstract roles** instead of specific agent names for portab
 | architecture | architect, design, planning, ADR, system, infrastructure |
 | implementation | engineer, developer, implement, code, build |
 | review | review, audit, quality, test, enforce, validate, check, gate |
+| domain-expert | domain, business, product, user, requirements, stakeholder |
+| architecture-reviewer | architect, design, patterns, system, simplicity, complexity |
 
 **Note**: The `implementation` role supports **multiple specialists** - different agents can be assigned to different features based on domain expertise.
 
@@ -48,47 +52,383 @@ This workflow uses **abstract roles** instead of specific agent names for portab
 | architecture | [pending discovery] | - | ⏳ |
 | implementation | [pending discovery] | - | ⏳ |
 | review | [pending discovery] | - | ⏳ |
+| domain-expert | [pending discovery] | - | ⏳ |
+| architecture-reviewer | [pending discovery] | - | ⏳ |
 
 ---
 
-## Phase 0: Agent Discovery
+## Workflow State
 
-Before starting the workflow, Claude Code discovers available specialists and maps them to roles.
+*Updated automatically during workflow execution. Clear this section after workflow completion or when starting fresh.*
+
+| Field | Value |
+|-------|-------|
+| Current Phase | - |
+| Current Step | - |
+| Last Updated | - |
+
+### Created Agents (pending validation)
+
+*List agents created in Phase 0.3, pending restart and validation.*
+
+| Agent File | Role | Status |
+|------------|------|--------|
+| - | - | - |
+
+### Linked Skills
+
+*Skills linked to created agents (if any).*
+
+| Agent | Skills |
+|-------|--------|
+| - | - |
+
+---
+
+## Phase 0: Agent Discovery & Setup
+
+Before starting the workflow, Claude Code discovers available specialists, identifies gaps, and optionally creates missing specialists tailored to the project.
+
+```
+Phase 0: Agent Discovery & Setup
+├── 0.1: Scan & Match
+├── 0.2: Gap Analysis & User Choice
+│   ├── [Create] → 0.3
+│   ├── [Generalist] → Phase 1
+│   └── [Hybrid] → 0.3 for selected roles
+└── 0.3: Specialist Creation (conditional)
+```
+
+---
+
+### Phase 0.1: Scan & Match
 
 **Actions**:
 1. Scan for agents at:
    - Project level: `.claude/agents/*.md`
    - User level: `~/.claude/agents/*.md`
-2. Read agent descriptions and match to roles using keywords
-3. Present candidates to user for approval
-4. Update "Discovered Agents" section above with selections
+2. Read agent descriptions and match to roles using keywords (see Role Matching Keywords table above)
+3. Present candidates to user
 
-**User Approval**:
+**Output**:
 
-Claude Code presents discovered agents:
 ```
-Found specialists for this workflow:
+Phase 0.1: Agent Discovery
+──────────────────────────
 
-research:       context-engineer (user) - "Web research and evidence synthesis"
-architecture:   project-architect (user) - "ADRs and system design"
-implementation: ai-engineer (user) - "RAG and LLM integration"
-                backend-engineer (user) - "APIs and databases"
-review:         python-code-quality-auditor (user) - "Code quality and test enforcement"
+Scanning for specialists...
 
-Approve these selections? [Y/n/modify]
+✅ Found:
+   implementation: backend-engineer (user) - "APIs and databases"
+   review:         python-code-quality-auditor (user) - "Quality gate"
+
+❌ Missing:
+   research:       No agent found
+   architecture:   No agent found
 ```
 
-**If No Matches Found**:
+If all roles have matches → Present for approval, update "Discovered Agents" table, proceed to Prerequisites.
+
+If any roles missing → Proceed to Phase 0.2.
+
+#### Domain Expert Selection
+
+The `domain-expert` role requires special handling - it should match the project's domain, not just generic keywords.
+
+**Selection Logic**:
+
+1. Analyze `context/PRODUCT.md` for domain keywords
+2. Map to appropriate specialist agent
+3. If no clear match, ask user
 
 ```
-⚠️  WARNING: No specialist found for role [research].
-    Workflow will proceed with generalist Claude Code.
-    Consider creating a specialist agent for better results.
+Analyzing PRODUCT.md for domain...
+
+Domain keywords found:
+- "machine learning", "model training" → ML domain
+- "e-commerce", "checkout", "cart" → Commerce domain
+- "healthcare", "patient", "clinical" → Healthcare domain
+- "API", "backend", "database" → Backend/infrastructure domain
+
+Suggested domain expert:
+┌─────────────────────────────────────────────────────────┐
+│ Domain: ML/Data Science                                 │
+│ Recommended: data-science-modeler (user)                │
+│ Reason: Keywords "model training", "pipeline" in PRD    │
+└─────────────────────────────────────────────────────────┘
+
+Accept this selection? [Y/n/specify different]
 ```
 
-**After Discovery**:
+**Domain-to-Agent Mapping**:
+
+| Domain Keywords | Suggested Agent | Fallback |
+|-----------------|-----------------|----------|
+| ML, model, training, prediction | data-science-modeler, ml-research-expert | product-manager |
+| API, backend, database, service | backend-engineer | product-manager |
+| UI, frontend, component, user interface | frontend-engineer | product-manager |
+| Data, ETL, pipeline, warehouse | data-engineer | product-manager |
+| Agent, autonomous, LLM, RAG | ai-engineer, agentic-engineer | product-manager |
+| (No clear match) | Ask user | product-manager |
+
+**If user must specify**:
+
+```
+No clear domain match found in PRODUCT.md.
 
-Update the "Discovered Agents" table above, then proceed to Prerequisites.
+What domain expertise is most relevant for validating this project?
+Examples: "ML/data science", "backend services", "e-commerce", "healthcare"
+
+Domain: _
+```
+
+---
+
+### Phase 0.2: Gap Analysis & User Choice
+
+When roles are missing, present options to the user:
+
+```
+Phase 0.2: Gap Resolution
+─────────────────────────
+
+For missing roles, choose an option:
+
+[1] Create specialists (recommended)
+    → I'll design agents based on your PRODUCT.md and ENGINEERING.md
+    → Follows Anthropic best practices
+    → Requires Claude Code restart after creation
+
+[2] Proceed with generalists
+    → Claude Code handles these roles directly
+    → Works but less specialized
+    → No restart needed
+
+[3] Hybrid approach
+    → Select which roles get specialists vs. generalists
+
+Your choice: _
+```
+
+**If user chooses [1] Create** → Proceed to Phase 0.3 for all missing roles.
+
+**If user chooses [2] Generalist** → Update "Discovered Agents" table with "generalist" entries, proceed to Prerequisites.
+
+**If user chooses [3] Hybrid** → Ask which roles to create specialists for:
+
+```
+Which roles should get specialists? (comma-separated)
+Missing: research, architecture
+
+Create specialists for: _
+```
+
+Then proceed to Phase 0.3 for selected roles, mark others as "generalist".
+
+---
+
+### Phase 0.3: Specialist Creation
+
+**Prerequisite**: User must have filled out `context/PRODUCT.md` and `context/ENGINEERING.md` seed documents.
+
+#### 0.3.1: Analyze Seeds for Domain Context
+
+Read seed documents and extract:
+
+| From PRODUCT.md | Extract |
+|-----------------|---------|
+| Problem statement | Domain keywords |
+| Target users | Interaction patterns |
+| Core requirements | Tool/skill needs |
+| Success criteria | Validation focus |
+
+| From ENGINEERING.md | Extract |
+|---------------------|---------|
+| Tech stack | Specific technology expertise |
+| Constraints | What agent should avoid/prefer |
+| Architecture patterns | DDD, event-driven, etc. |
+| Non-functional requirements | Performance, security focus |
+
+#### 0.3.2: Fetch Best Practices
+
+Use `claude-code-guide` subagent to fetch Anthropic documentation on:
+- Agent file structure and frontmatter
+- Effective agent prompts
+- Skill integration patterns
+
+#### 0.3.3: Draft Agent Files
+
+Generate agent markdown files following this structure:
+
+```markdown
+---
+name: [role]-specialist
+description: [One-line domain-specific description]
+---
+
+# [Role] Specialist for [Project Name]
+
+You are a [role] specialist for [domain] projects.
+
+## Expertise
+[Derived from seeds - specific technologies, patterns, domain knowledge]
+
+## Responsibilities
+[Role-specific tasks from workflow]
+
+## Constraints
+[From ENGINEERING.md - what to avoid, preferences]
+
+## Skills
+[Optional: linked skills if user requested]
+```
+
+#### 0.3.4: Present Drafts for User Approval
+
+```
+Phase 0.3.4: Agent Draft Review
+───────────────────────────────
+
+I've drafted the following specialists based on your seeds:
+
+┌─────────────────────────────────────────────────────────┐
+│ research-specialist.md                                   │
+├─────────────────────────────────────────────────────────┤
+│ Description: Research ML pipeline technologies and       │
+│ synthesize evidence on [domain] patterns                 │
+│                                                         │
+│ Expertise: [tech stack from seeds]                      │
+│ Skills: [suggested skills]                              │
+└─────────────────────────────────────────────────────────┘
+
+[View full draft? Y/n]
+
+Options:
+[1] Approve all drafts
+[2] Modify a draft (opens for editing)
+[3] Regenerate with different focus
+[4] Cancel specialist creation
+```
+
+#### 0.3.5: Skills Configuration
+
+Before creating agent files, ask about skills:
+
+```
+Skills Configuration
+────────────────────
+
+Would you like to link specific skills to these agents?
+
+Options:
+[1] Auto-suggest skills based on project domain
+    → I'll recommend skills matching your seeds
+[2] Manually specify skills for each agent
+[3] Skip - let Claude Code auto-load skills as needed
+[4] Create new project-specific skills
+    → I'll help design skills tailored to this project (sub-workflow)
+
+Your choice: _
+```
+
+If user chooses [4], trigger skill creation sub-workflow before continuing.
+
+#### 0.3.6: Create Agent Files
+
+Create approved agents in `.claude/agents/`:
+
+```
+Creating agents...
+  ✅ .claude/agents/research-specialist.md
+  ✅ .claude/agents/architecture-specialist.md
+
+Updating workflow state...
+```
+
+Update the "Discovered Agents" table with created agents (status: "created, pending validation").
+
+#### 0.3.7: Update Workflow State
+
+Update the "Workflow State" section (below) with:
+- Current phase: 0.3
+- Current step: pending_restart
+- Created agents list
+- Linked skills (if any)
+
+#### 0.3.8: Prompt User to Restart
+
+```
+───────────────────────────────────────────────────────────
+Agent files created successfully!
+
+Claude Code needs to restart to load the new agents.
+
+Please:
+1. Exit this session (Ctrl+C or type 'exit')
+2. Restart Claude Code in this directory
+3. Say: "Continue project initialization workflow"
+
+Progress saved to this file (Workflow State section below).
+───────────────────────────────────────────────────────────
+```
+
+#### 0.3.9: Post-Restart Validation
+
+After user restarts and continues, validate each created agent:
+
+```
+Phase 0.3.9: Agent Validation
+─────────────────────────────
+
+Validating created agents...
+
+research-specialist:
+  Invoking: "Introduce yourself in one sentence."
+  Response: "I research ML pipeline technologies, synthesize
+            evidence on XGBoost and FastAPI patterns, and prepare
+            knowledge packages for architecture decisions."
+  ✅ Agent responds correctly.
+
+architecture-specialist:
+  Invoking: "Introduce yourself in one sentence."
+  Response: "I design system architectures for data pipelines,
+            create ADRs, and define MVP scope with MoSCoW prioritization."
+  ✅ Agent responds correctly.
+```
+
+**If validation fails**:
+
+```
+❌ Agent 'research-specialist' failed to load.
+   Error: [error message]
+
+   Options:
+   [1] Review and fix the agent file manually
+   [2] Delete and recreate
+   [3] Skip - use generalist for this role
+```
+
+#### 0.3.10: Finalize Discovery
+
+After validation:
+1. Update "Discovered Agents" table with final status (✅)
+2. Clear "Workflow State" section (workflow complete)
+3. Proceed to Prerequisites
+
+---
+
+### Generalist Fallback Behavior
+
+When proceeding with generalists (by choice or after failed creation):
+
+- Claude Code handles the role directly without a specialist agent
+- Quality is acceptable but less domain-specific
+- User can create specialists later and re-run Phase 0
+
+Update "Discovered Agents" table:
+```markdown
+| research | generalist (Claude Code) | - | ✅ |
+```
 
 ---
 
@@ -364,17 +704,159 @@ If a deliverable fails review 3 times:
 - `make validate-branch` passes
 - Coverage >= 80% for new code
 
+---
+
+### 3.4 Domain & Architecture Feedback
+
+**Roles**: `domain-expert` + `architecture-reviewer` (read-only, sequential)
+**Triggered**: After all deliverables pass code review, before phase gate approval.
+
+This feedback gate ensures implementations are aligned with business outcomes and not over-engineered.
+
+#### 3.4.1 Domain Expert Review (First)
+
+**Role**: `domain-expert`
+
+**Input**:
+- Implemented code from Phase 3
+- `context/PRD.md` - Business requirements
+- `context/PROJECT_PLAN.md` - MVP scope
+
+**Checks**:
+1. Does implementation serve stated user needs from PRD?
+2. Is it aligned with success criteria?
+3. Does it address the core problem statement?
+4. Are there unnecessary features not in scope?
+
+**Report Format**:
+
+```
+## Domain Alignment Review
+**Phase**: 3 (MVP)
+**Status**: ALIGNED | MISALIGNED | OVER-SCOPED
+
+### Alignment with PRD
+- [ ] Addresses problem statement
+- [ ] Serves target users
+- [ ] Meets success criteria
+- [ ] Within defined scope
+
+### Concerns
+[List any business/domain misalignments]
+
+### Recommendations
+[Suggested adjustments]
+```
+
+**If MISALIGNED or OVER-SCOPED** → Blocking. Implementation must fix issues before proceeding.
+
+#### 3.4.2 Architecture Review (Second)
+
+**Role**: `architecture-reviewer`
+
+**Input**:
+- Implemented code from Phase 3
+- `ADR.md` - Architecture decisions
+- Domain review findings (from 3.4.1)
+
+**Checks**:
+1. Does implementation follow patterns in ADR.md?
+2. Is complexity justified by requirements?
+3. Are there simpler alternatives for current phase?
+4. Is it over-engineered for MVP scope?
+
+**Over-Engineering Signals**:
+- Abstractions without current use cases
+- Premature optimization
+- Features beyond Must-Have scope
+- Configuration/flexibility not requested
+- Layers/indirection without clear benefit
+
+**Report Format**:
+
+```
+## Architecture Review
+**Phase**: 3 (MVP)
+**Status**: APPROPRIATE | OVER-ENGINEERED | UNDER-DESIGNED
+
+### ADR Compliance
+- [ ] Follows ADR-001: [decision]
+- [ ] Follows ADR-002: [decision]
+...
+
+### Complexity Assessment
+- Complexity level: [Simple | Moderate | Complex]
+- Justified by requirements: [Yes | No]
+
+### Over-Engineering Detected
+[List any unnecessary complexity]
+
+### Simplification Recommendations
+[Concrete suggestions to reduce complexity]
+```
+
+**If OVER-ENGINEERED** → Blocking. Implementation must simplify before proceeding.
+
+#### 3.4.3 Combined Feedback Summary
+
+```
+Phase 3.4: Domain & Architecture Feedback
+─────────────────────────────────────────
+
+┌─────────────────────────────────────────────────────────┐
+│ DOMAIN REVIEW                                           │
+│ Status: ALIGNED                                         │
+│ ✓ Addresses problem statement                           │
+│ ✓ Serves target users                                   │
+│ ✓ Within MVP scope                                      │
+└─────────────────────────────────────────────────────────┘
+
+┌─────────────────────────────────────────────────────────┐
+│ ARCHITECTURE REVIEW                                     │
+│ Status: OVER-ENGINEERED                                 │
+│ ✗ Abstract factory pattern not needed for 2 variants   │
+│ ✗ Configuration system beyond requirements             │
+│                                                         │
+│ Recommendations:                                        │
+│ 1. Replace factory with simple if/else                  │
+│ 2. Hardcode config values, add flexibility in Phase 4   │
+└─────────────────────────────────────────────────────────┘
+
+Blocking Issues: 2
+Action Required: Simplify implementation before approval gate
+```
+
+#### 3.4.4 Fix Feedback Issues
+
+**Role**: `implementation`
+
+If domain or architecture review identifies blocking issues:
+
+1. Address each blocking issue from feedback
+2. Apply simplification recommendations
+3. Re-run `make test` to ensure changes don't break tests
+4. Return to 3.4.1 for re-review (max 2 feedback cycles)
+
+**Loop Control**:
+- Maximum **2 feedback cycles** per phase
+- If still failing after 2 cycles → **escalate to user**
+
+---
+
 ### [APPROVAL GATE 3]
 
 User validates:
 - MVP functionality (end-to-end)
-- All deliverables passed review loop
+- All deliverables passed code review loop
+- Domain alignment review passed (ALIGNED)
+- Architecture review passed (APPROPRIATE)
 - Test coverage meets requirements
 
 **Options**:
 - ✅ **Approve** - Proceed to Phase 4
 - 🐛 **Request fixes** - Specify additional issues to address
 - ✏️ **Adjust scope** - Modify MVP boundaries
+- 🔄 **Re-run feedback** - Request another domain/architecture review
 
 ---
 
@@ -415,18 +897,60 @@ Phase 4 follows the same **Implement → Review → Fix** loop as Phase 3 for ea
 **Actions**:
 1. Run same validation as Phase 3.2 (tests, naming, coverage, quality)
 2. Generate review report
-3. If PASS → proceed to user approval
+3. If PASS → proceed to domain/architecture feedback
 4. If FAIL → implementation agent fixes (max 3 cycles)
 
 ---
 
-### 4.3 Feature Approval
+### 4.3 Domain & Architecture Feedback
+
+**Roles**: `domain-expert` + `architecture-reviewer` (read-only, sequential)
+**Triggered**: After feature passes code review, before feature approval.
+
+Same structure as Phase 3.4, applied per feature:
+
+#### 4.3.1 Domain Expert Review
+
+**Checks**:
+1. Does this feature serve stated user needs from PRD?
+2. Is it aligned with roadmap priorities?
+3. Does it integrate well with existing MVP?
+4. Is it within the Should-Have/Could-Have scope for this phase?
+
+**Report Format**: Same as Phase 3.4.1
+
+**If MISALIGNED** → Blocking. Must fix before approval.
+
+#### 4.3.2 Architecture Review
+
+**Checks**:
+1. Does feature follow established patterns from Phase 3?
+2. Is complexity appropriate for a Phase 4 enhancement?
+3. Does it introduce unnecessary technical debt?
+4. Is it over-engineered for the stated requirement?
+
+**Report Format**: Same as Phase 3.4.2
+
+**If OVER-ENGINEERED** → Blocking. Must simplify before approval.
+
+#### 4.3.3 Fix Feedback Issues
+
+If blocking issues found:
+1. Address domain/architecture concerns
+2. Re-run tests
+3. Return to 4.3.1 for re-review (max 2 cycles)
+
+---
+
+### 4.4 Feature Approval
 
 ### [APPROVAL GATE per feature]
 
-After review passes, user validates:
+After code review AND domain/architecture feedback pass, user validates:
 - Feature functionality
-- Review report shows PASS
+- Code review report shows PASS
+- Domain alignment review passed (ALIGNED)
+- Architecture review passed (APPROPRIATE)
 - Integration with existing MVP
 
 **Options**:
@@ -434,13 +958,15 @@ After review passes, user validates:
 - 🔄 **Request changes** - Modify current feature
 - ⏭️ **Skip to next** - Move to next roadmap item
 - 🛑 **Stop** - End enhancement phase
+- 🔄 **Re-run feedback** - Request another domain/architecture review
 
 ---
 
 **Output**: Enhanced system with additional features
 
 **Success Criteria**:
-- Each feature passes review loop
+- Each feature passes code review loop
+- Each feature passes domain/architecture feedback
 - `make validate-branch` passes
 - Coverage maintained >= 80%
 
@@ -597,6 +1123,93 @@ Templates in `workflows/templates/` define output formats. Agents use these to p
 - Ask agent to re-read specific context file
 - Update context if requirements changed
 
+### "Created agent doesn't load after restart"
+
+**Cause**: Syntax error in agent file or invalid frontmatter
+
+**Fix**:
+- Check agent file has valid YAML frontmatter (between `---` markers)
+- Verify `name` and `description` fields are present
+- Look for special characters that need escaping
+- Try recreating the agent using Phase 0.3
+
+### "Agent validation fails with error"
+
+**Cause**: Agent file malformed or Claude Code can't invoke it
+
+**Fix**:
+- Review error message for specifics
+- Compare agent file structure to working examples in `~/.claude/agents/`
+- Delete and recreate the agent
+- Fall back to generalist for this role
+
+### "Workflow state shows pending_restart but I already restarted"
+
+**Cause**: Claude Code didn't detect the continuation prompt
+
+**Fix**:
+- Say exactly: "Continue project initialization workflow"
+- Check the "Workflow State" section shows the correct phase
+- Manually update "Current Step" to "validation" and re-run
+
+### "Seeds are empty but I want to create specialists"
+
+**Cause**: Phase 0.3 requires PRODUCT.md and ENGINEERING.md to design domain-specific agents
+
+**Fix**:
+- Fill out `context/PRODUCT.md` with at least: problem statement, target users, tech stack
+- Fill out `context/ENGINEERING.md` with at least: tech preferences, constraints
+- Re-run Phase 0.3
+
+### "Domain review says MISALIGNED but implementation matches PRD"
+
+**Cause**: PRD may have evolved since implementation, or domain expert misunderstood context
+
+**Fix**:
+- Review PRD - has it changed since Phase 2?
+- Ask domain expert to re-read specific PRD section
+- If PRD is outdated, update it and re-run domain review
+- Override with user approval if alignment is subjective
+
+### "Architecture review flags everything as over-engineered"
+
+**Cause**: Reviewer may be too aggressive, or project genuinely needs simplification
+
+**Fix**:
+- Review specific recommendations - are they actionable?
+- Check if ADR.md justifies the complexity (documented decisions)
+- If patterns are intentional, add rationale to ADR and re-run review
+- Escalate to user if disagreement persists
+
+### "Implementation keeps failing feedback after simplification"
+
+**Cause**: Feedback criteria may be too strict, or fundamental approach needs rethinking
+
+**Fix**:
+- After 2 feedback cycles, workflow escalates to user
+- User can override with documented risk
+- Consider splitting deliverable into smaller pieces
+- May need to revisit architecture decisions in ADR.md
+
+### "No domain expert found for my project type"
+
+**Cause**: Project domain doesn't match available agents
+
+**Fix**:
+- Use `product-manager` as fallback (general business alignment)
+- Create a domain-specific agent using Phase 0.3
+- Specify domain manually when prompted during Phase 0.1
+
+### "Domain and architecture reviews contradict each other"
+
+**Cause**: Different perspectives on what's appropriate
+
+**Fix**:
+- Domain review runs first, architecture considers its findings
+- If conflict persists, escalate to user
+- User decides which feedback takes priority
+- Document decision in ADR.md for future reference
+
 ---
 
 ## Agent Best Practices
@@ -626,3 +1239,9 @@ After each workflow execution, update this file to:
 | 2025-11-28 | Reorganized: context/ for seeds, workflows/ for support | Clearer separation of user inputs vs. system files |
 | 2025-11-28 | Added Phase 0 agent discovery, abstract roles | Workflow portability - users may have different specialists |
 | 2025-11-28 | Added `review` role with Implement → Review → Fix loop | Enforce test creation and code quality before progression |
+| 2025-11-29 | Enhanced Phase 0 with sub-phases 0.1-0.3 for agent creation | Handle missing specialists by offering creation vs. generalist fallback |
+| 2025-11-29 | Added Workflow State section for restart recovery | Claude Code requires restart to load new agents; state persistence enables continuation |
+| 2025-11-29 | Added skills configuration step in Phase 0.3.5 | Allow users to link existing skills or create project-specific ones |
+| 2025-11-29 | Added `domain-expert` and `architecture-reviewer` roles | Validate business alignment and prevent over-engineering |
+| 2025-11-29 | Added Phase 3.4 and 4.3 Domain & Architecture Feedback gates | Blocking feedback loop after code review, before approval |
+| 2025-11-29 | Added domain expert selection logic to Phase 0.1 | Infer domain from PRODUCT.md, ask user if unclear |