Architect Skill

Bootstrap and manage a unified AI context system using the three-layer model: Product (stable reference), Project (current build cycle), and Task (subagent capabilities).

Commands

/architect init

Bootstrap the Architect framework in the current project.

Collects via AskUserQuestion:

• PROJECT_NAME - Human-readable name (e.g., "My Awesome App")
• PROJECT_DESCRIPTION - 2-3 sentence domain description
• TECH_STACK - Primary language/framework (TypeScript, Python, Go, etc.)
• DEPLOYMENT - Target platform (Replit, Vercel, Local, AWS, etc.)

Creates:

• docs/architect/ directory with three-layer structure
• docs/architect/Master_Context.md - Hub file, single source of truth
• docs/architect/product/ - Stable reference (architecture, design, domain)
• docs/architect/project/ - Current build cycle (active, decisions, state)
• docs/plans/ - Task documentation and backlog (at docs/ level)
• docs/architect/subagents/ - Specialized agent capabilities
• docs/architect/hooks/ - Automation scripts
• docs/architect/sessions/ - Exported conversation history
• .claude/settings.json - Hook configuration
• .claude/agents/architect.md - Project-specific orchestrator
• IDE symlinks: CLAUDE.md, replit.md, .cursorrules, .github/copilot-instructions.md

Implementation Steps:

1. Use AskUserQuestion to collect project details
2. Create all directories under docs/architect/ and docs/plans/
3. Process templates from the plugin's templates/ directory (use ${CLAUDE_PLUGIN_ROOT}/templates/)
4. Substitute {{VARIABLE}} placeholders with collected values
5. Copy hook scripts (executable)
6. Create symlinks for IDE integration
7. Report success with next steps

/architect status

Check context health and report any issues.

Checks:

• Required files exist (Master_Context.md, project/active.md, etc.)
• Symlinks are valid
• Hooks are executable
• project/state.md is recent (warn if > 7 days stale)

Reports:

• Missing files
• Broken symlinks
• Permission issues
• Stale context warnings

/architect archive

Archive current project context and prepare for next cycle.

Actions:

1. Create docs/architect/archive/{feature-name}/ directory
2. Move project/active.md, project/decisions.md, project/state.md to archive
3. Move relevant plans/*.md files to archive
4. Create fresh project/ files from templates
5. Update Master_Context.md to reflect cleared state
6. Report what was archived

Usage:

/architect archive auth-system

/architect setup-context

Initialize the pgvector context system in the project's database.

Prerequisites:

• PostgreSQL database with pgvector extension available
• DATABASE_URL environment variable set
• OPENAI_API_KEY environment variable set (for embeddings)
• ANTHROPIC_API_KEY environment variable set (for classification)

Actions:

1. Enable pgvector extension in database
2. Create session_embeddings table with vector column
3. Create indexes for efficient similarity search
4. Copy context system files to docs/architect/context/
5. Report success

Usage:

/architect setup-context

/architect log

Manually ingest the current or specified session into the context database.

Actions:

1. Export current session to sessions/ (if not already exported)
2. Chunk session by conversation turns
3. Classify each chunk by layer/sublayer using Claude
4. Generate embeddings using OpenAI
5. Store in pgvector database

Usage:

/architect log                    # Log current session
/architect log sessions/my-session.md  # Log specific session file

/architect context

Query the context database for relevant information.

Modes:

• project - Query based on current project/active.md (default)
• layer <layer> [sublayer] - Query specific layer
• search <query> - Free-text semantic search

Usage:

/architect context                      # Auto-query based on current work
/architect context layer user patterns  # Query user patterns
/architect context search "auth flow"   # Semantic search

/architect session

Manage named sessions for better context organization.

Actions:

• start <name> - Start a named session (writes to .current-session)
• end - End current named session
• (no args) - Show current session name

Usage:

/architect session start auth-refactor  # Start named session
/architect session                      # Show current session
/architect session end                  # End session

When a named session is active, all /compact exports append to the same session file instead of creating new timestamped files.

Context Layers

The Architect system organizes knowledge into five retrieval layers, each with sublayers:

Five-Layer Model (for Context Retrieval)

Layer	Sublayers	What it captures
USER	patterns, preferences, learnings, mistakes	How you work, recurring approaches, lessons learned
PRODUCT	architecture, design, domain	System knowledge that persists across features
PROJECT	active, decisions, state	Current feature context and choices
PLAN	strategy, steps, backlog	Implementation approaches and future work
TASK	implementation, debugging, verification	How specific work was executed

Three-Layer File Model (for Organization)

PRODUCT (Architecture Agent)     - What is this? How does it work?
        |                          Stable reference, rarely changes
        v
PROJECT (Build Agent)            - What are we building right now?
        |                          Current cycle, cleared when done
        v
TASK    (Subagents)              - How do specific things work?
                                   Capabilities for specialized work

Layer	Folder	Purpose	Changes
PRODUCT	docs/architect/product/	Stable reference (architecture, design, domain knowledge)	Rarely
PROJECT	docs/architect/project/	Current build cycle (active work, decisions, state)	Each feature
TASK	docs/architect/subagents/	Specialized capabilities (domain experts)	As needed

How They Connect

• Files (product/, project/, plans/) are the explicit, curated context you maintain
• Sessions are ingested into pgvector and classified by the five-layer model
• Queries combine file context (active.md) with semantic search of session history

Lifecycle

1. New feature starts - Create/update project/active.md
2. During build - Update state.md, log decisions
3. Feature ships - Run /architect archive {feature}
4. Learnings graduate - Move stable patterns to product/
5. Clear project/ - Ready for next cycle

Auto-Detection (SessionStart Hook)

The skill includes an optional SessionStart hook that detects uninitialized projects.

What it does:

• Runs when you start a new Claude Code session
• Checks if docs/architect/ folder exists in the project
• If missing, suggests running /architect init

Installation:

cd $CLAUDE_PLUGIN_ROOT
./install-hook.sh

This installs a user-level hook to ~/.claude/settings.json that runs for all projects.

Hooks

The framework includes automation hooks for session management and context retrieval.

Ingestion Hooks (write to context DB)

Hook	Trigger	Action
PreCompact	Before /compact	Export session, then ingest into pgvector
ExitPlanMode	Plan approved	Commit, export session, ingest with plan metadata
post-commit	After git commit	Ingest recent session with commit hash

Query Hooks (read from context DB)

Hook	Trigger	Action
SessionStart	Opening project	Query relevant context, inject into system prompt

Activity Tracking

Hook	Trigger	Action
post-commit	After git commit	Update project/state.md and Master_Context.md Recent Activity

Hook Details

PreCompact (export-session.sh)

• Exports conversation to sessions/{timestamp}-session.md or sessions/{session-name}.md
• If context system configured, ingests session into pgvector

ExitPlanMode (plan-approved.sh)

• Commits staged changes
• Exports session
• Creates plan doc in plans/
• Ingests session with plan_approved trigger and commit hash

SessionStart (session-start-context.sh)

• Queries pgvector for context relevant to current project/active.md
• Outputs context wrapped in <architect-context> tags
• Silently skips if context system not configured

IDE Symlinks

All AI tools read the same context via symlinks to Master_Context.md:

File	Tool
CLAUDE.md	Claude Code
replit.md	Replit AI
.cursorrules	Cursor
.github/copilot-instructions.md	GitHub Copilot
.windsurfrules	Windsurf

One file to edit, all tools updated.

Template Variables

Templates use {{VARIABLE}} syntax:

Variable	Example	Description
{{PROJECT_NAME}}	"Brain Pinecone Manager"	Human-readable name
{{PROJECT_SLUG}}	"brain-pinecone-manager"	URL-safe lowercase
{{PROJECT_DESCRIPTION}}	"A vector database management app..."	Brief description
{{TECH_STACK}}	"TypeScript"	Primary language/framework
{{DEPLOYMENT}}	"Replit"	Target platform
{{DATE}}	"2026-01-17"	Current date

Subagent Catalog

Optional specialized agents available in catalog/:

Agent	Focus
ui-agent.md	Frontend components, styling, UX
api-agent.md	Backend endpoints, services, data flow
db-agent.md	Database schema, queries, migrations

Domain-specific agents can be added based on project needs.