Know Tool - Specification Graph Mastery

What is the Specification Graph?

The specification graph (.ai/know/spec-graph.json) is a directed acyclic graph (DAG) representing software systems as interconnected nodes with explicit dependencies. Everything is a node, relationships are explicit, nothing is implied.

Three node types:

1. Entities - Structural nodes that participate in dependencies (user, feature, component, etc.)
2. References - Terminal nodes with implementation details (business_logic, data-models, etc.)
3. Meta - Project metadata (phases, assumptions, decisions, qa_sessions)

Key principle: The graph IS the source of truth. All relationships are explicit.

Phases in meta.phases

The meta.phases section tracks feature lifecycle and scheduling:

Phase Types:

• I, II, III - Scheduling phases (immediate, next, future)
• pending - Not yet scheduled
• done - Completed and archived

Phase Status:

• incomplete - Feature added but not started
• in-progress - Active development
• review-ready - Implementation complete, awaiting review
• complete - Finished (in done phase)

Phase Lifecycle:

/know:add    → pending phase, status: incomplete
/know:build  → status: in-progress → review-ready
/know:done   → done phase, status: complete

Core Mental Model

Single Product Chain

Project → User → Objective → Feature → Action → Component → Operation

Flows from who uses it, through what they want, to how the system delivers it. requirement and interface are reference types, not entities.

Dependency Rules

Dependencies are strict and unidirectional:

• Only entities participate in dependencies
• References are terminal nodes (no dependencies)
• Graph must remain a DAG (no cycles)

Command Groups

Know CLI uses a flat structure with auto-detection:

Command	Purpose
know get <type:key>	Get entity or reference (auto-detects)
know list [--type TYPE]	List entities or references (auto-detects)
know search <pattern>	Search all text content (supports regex)
know add <type> <key> [key2 ...]	Add one or more entities/references (auto-detects)
know link <from> <to> [to2 ...]	Add one or more dependencies
know unlink <from> <to> [to2 ...]	Remove one or more dependencies
know nodes	Node operations: deprecate, merge, rename, delete, cut, clone, update
know meta	Get, set, and delete meta sections (project, phases, decisions)
know graph	Traverse, uses, used-by, connect, clean, suggest, diff, migrate, coverage, cross connect, cross coverage
know graph coverage	Show % of spec entities reachable from root users
know graph cross connect [feature]	Auto-connect spec features/components to code via token matching
know graph cross coverage	Show spec↔code link coverage (% with code-link refs)
know check	Validate, health, stats, gaps, orphans, cycles, completeness
know gen	Specs, feature-specs, docs, traces, rules, codemap, code-graph, sitemap
know feature	Lifecycle: status, connect, review, done, impact, validate, contracts, coverage
know req	Requirements: add, list, status, block, complete
know op	Op-level progress: start, done, next, reset, status
know phases	Phase management: list, add, move, status, remove
know init	Initialize know workflow (installs graph protection hooks)

Using know gen rules Commands

These commands expose the dependency structure for LLMs:

# Understand any type
know gen rules describe feature
know gen rules describe business_logic
know gen rules describe phases

# See dependency rules
know gen rules before component    # What can depend on component?
know gen rules after feature        # What can feature depend on?

# Visualize the structure
know gen rules graph                # See the full dependency map

Always start with know gen rules commands before manipulating the graph.

Essential Commands

Discovery & Exploration

know list                           # List all entities
know list --type feature            # List entities of type (auto-detects entity vs reference)
know list --type business_logic     # List references of type (auto-detects)
know get feature:real-time-telemetry   # Get entity (auto-detects)
know get business_logic:login          # Get reference (auto-detects)

# Search
know search "authentication"                        # Plain text search (case-insensitive)
know search "auth.*login" --regex                   # Regex search
know search "API" --section references              # Search only references
know search "user" --field description              # Search specific field
know search "Feature.*" -rc                         # Regex, case-sensitive

# Dependencies
know graph uses feature:real-time-telemetry          # What does this entity use? (dependencies)
know graph used-by component:websocket-manager       # What uses this entity? (dependents)
know graph up feature:x                              # Alias for 'uses' (go up dependency chain)
know graph down component:y                          # Alias for 'used-by' (go down chain)

# Cross-Graph Navigation (spec ↔ code)
know graph traverse feature:auth --direction impl    # Show code implementations
know graph traverse module:auth --direction spec     # Show spec feature
know graph traverse feature:profile                  # Auto-detects direction

# Statistics
know graph check stats                    # Graph statistics (entity counts, dependencies)
know graph check completeness feature:x   # Completeness score for an entity

Modification

know add feature new-feature '{"name":"...", "description":"..."}'       # Add single entity
know add feature feat-a feat-b feat-c -f data.json                      # Add multiple (shared data from file)
know add documentation new-doc '{"title":"...", "url":"..."}'            # Add reference (auto-detects)
know meta set project key '{"value":"..."}'                              # Set meta value
know meta get project                                                    # Get meta section
know meta delete phases I                                                # Delete meta key (prompts)
know meta delete requirements auth-login -y                              # Delete meta key (skip prompt)
know link feature:auth action:login                                      # Add single dependency
know link feature:auth action:login action:logout component:session      # Add multiple at once
know unlink feature:auth action:login action:logout                      # Remove multiple at once

Validation

know graph check validate                 # Must run after changes (includes fix commands in errors)
know graph check health                   # Comprehensive check
know graph check cycles                   # Find circular dependencies

Requirements (know req)

know req list feature:auth              # List requirements with status
know req add feature:auth req-name '{"description":"..."}' # Add requirement
know req status feature:auth req-name in-progress          # Update status
know req block feature:auth req-name --by "reason"         # Mark blocked
know req complete feature:auth req-name                    # Mark complete

Status values: pending, in-progress, blocked, complete, verified

Op-Level Progress (know op)

know op status feature:auth             # Show op status for a feature
know op next feature:auth               # Print the next op number
know op start feature:auth              # Mark current op as in-progress
know op done feature:auth               # Mark current op as complete with commits
know op reset feature:auth              # Reset an op to pending

Feature Lifecycle (know feature)

know graph cross connect feature:auth            # Create bidirectional spec ↔ code linkage
know feature connect auth module:x module:y      # Link feature to multiple code entities at once
know feature review feature:auth        # Review for completion: validate graph, check coverage
know feature done feature:auth          # Complete: tag commits, update phase, archive
know feature impact entity:x            # Show features depending on an entity or file
know feature validate feature:auth      # Check if codebase changes warrant revisiting
know feature contract feature:auth      # Display contract info
know feature validate-contracts         # Validate for drift between contracts
know feature coverage feature           # Aggregate coverage from feature level
know feature coverage feature --detail  # Per-component breakdown

Node Operations (know nodes)

# Deprecation
know nodes deprecate entity:id --reason "..." [--replacement entity:new]
know nodes undeprecate entity:id
know nodes deprecated             # List all deprecated entities
know nodes deprecated --overdue   # Only entities past removal date

# Modification
know nodes update entity:id '{"name":"New Name"}'  # Update properties
know nodes rename entity:id new-key                # Rename entity key (shows confirmation)
know nodes rename entity:id new-key -y             # Skip confirmation
know nodes clone entity:id new-key                 # Clone with all dependencies
know nodes clone entity:id new-key --no-upstream  # Clone without incoming deps

# Removal (works with entities AND references)
know nodes delete feature:old              # Remove entity, clean up dependencies
know nodes delete feature:old action:bar   # Remove multiple at once
know nodes delete data-model:old-schema    # Remove reference, clean up dependencies
know nodes delete component:temp -y        # Skip confirmation
know nodes cut entity:id                   # Remove node only, leave deps orphaned
know nodes cut reference:id -y             # Skip confirmation

# Merging
know nodes merge from:entity into:entity      # Merge entities, transfer deps (shows confirmation)
know nodes merge from:entity into:entity -y   # Skip confirmation
know nodes merge from:entity into:entity --keep  # Keep source after merge

# Graph Operations
know link feature:x action:y action:z    # Add one or more dependencies
know unlink feature:x action:y action:z  # Remove one or more (shows confirmation)
know unlink feature:x action:y -y        # Skip confirmation

Important: All destructive operations (delete, cut, rename, merge, unlink) now show detailed confirmation prompts by default. Use -y or --yes to skip confirmation in scripts.

Note: Validation errors include fix commands. For example:

✗ Invalid dependency: feature:x → component:y. feature can only depend on: action
  Fix: know unlink feature:x component:y

Analysis

know graph check gap-analysis feature:x  # Find missing dependencies
know graph check gap-missing             # List missing connections in chains
know graph check gap-summary             # Overall implementation status
know graph check orphans                 # Find unused references
know graph check usage                   # Reference usage statistics
know graph suggest                 # Suggest connections for orphaned references
know graph clean                   # Clean up unused references (dry run)
know graph clean --remove --execute # Actually remove unused references
know graph build-order             # Topological sort
know graph connect entity:x        # Suggest valid connections for an entity

Generation (know gen)

know gen spec entity:x              # Generate spec for a single entity
know gen feature-spec feature:x     # Generate detailed feature specification
know gen docs feature:x             # Generate .md files from graph for a feature
know gen trace entity:x             # Trace entity across product-code boundary
know gen trace-matrix               # Show requirement traceability matrix
know gen trace-matrix -t component  # Trace specific entity type
know gen sitemap                    # Generate sitemap of all interfaces
know gen codemap                    # Generate code structure map
know gen code-graph                 # Generate code-graph from codemap

Migration (know graph migrate)

know graph migrate                  # Check graph structure against current rules
know graph migrate --format json    # Structured output for LLM consumption
know graph migrate-rules /path/to/new-rules.json  # Analyze impact of rules change
know graph migrate-rules /path/to/new-rules.json --format json --verbose

Advanced

know graph diff graph1.json graph2.json  # Compare two graph files
know init                                # Initialize know workflow in a project

Initialization & Protection

know init sets up the complete know workflow:

know init                    # Run in project root
know init --project-dir /path/to/project

What it installs:

1. Slash commands → .claude/commands/know/
2. know-tool skill → .claude/skills/know-tool/
3. Agents → .claude/agents/
4. Directory structure → .ai/know/
5. Project template → .ai/know/project.md
6. Graph protection hook → .claude/hooks/protect-graph-files.sh

Graph Protection Hook:

The hook automatically blocks direct Read/Edit/Write access to *-graph.json files, enforcing use of the know CLI:

❌ Direct Edit access to graph files is not allowed

Graph file: .ai/know/spec-graph.json

⚠️  Use the know CLI instead:
   • Read: know get <entity-id>
   • List: know list
   • Edit: know add <type> <key> <data>
   • Link: know link <from> <to>
   • Validate: know graph check validate

Why this matters: Direct file editing can corrupt the graph structure. The hook ensures all modifications go through validated CLI commands.

Configuration: The hook is installed in .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Read|Edit|Write",
      "hooks": [{
        "type": "command",
        "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/protect-graph-files.sh"
      }]
    }]
  }
}

Reference Files

For detailed information, read these reference files:

• creating.md - Guide to creating entities and references
• generating.md - Specification generation patterns
• validating.md - Graph validation and health checks
• qa-steps.md - QA testing patterns

Quick Workflow Pattern

When adding a new feature:

# 1. Understand the type
know gen rules describe feature

# 2. Add the entity
know add feature new-feature '{"name":"...", "description":"..."}'

# 3. Check what it can depend on
know gen rules after feature

# 4. Connect dependencies
know link feature:new-feature action:trigger-action

# 5. Validate
know graph check validate
know graph uses feature:new-feature --recursive

Phase Management

Phase = Roman numerals (I, II, III) - WHEN to do this feature (planning waves) Status = in-progress, complete, planned - current state of the work

Phase is the plan, status is the territory. A feature can be phase: III (planned for wave 3) but status: in-progress (started early).

know phases                          # Alias for phases list
know phases list                     # Show all features organized by phase
know phases add <phase> <entity>     # Add feature to phase (e.g., know phases add I feature:auth)
know phases move <entity> <phase>    # Move feature to different phase
know phases status <entity> <status> # Update status (planned, in-progress, complete)
know phases remove <entity>          # Remove entity from all phases

Output includes:

• Phase metadata (shortname, name, description) from meta.phases_metadata
• Features within each phase
• Requirement completion counts (from meta.requirements)
• Status icons (✅ complete, 🔄 in-progress, 📋 planned)
• Summary totals

Example output:

Phase I (Foundation)
  🔄 feature:auth (3/5) - Authentication system

Phase II (Features)
  📋 feature:api-gateway (0/4) - API routing

Phase III (Polish)
  📋 feature:dark-mode (--) - No requirements yet

Note: "--" indicates no requirements exist yet for that feature.

Feature Status Tracking

Virtual flags computed from graph state (not stored, derived):

know feature status feature:auth        # Show all three status flags
know phases list                        # Shows status icons inline

Status Flags

1. 📋 Planned - Feature exists in meta.phases (any phase)

   • Set by: /know:add or /know:plan adding to phases
   • Computed: Check if feature_id in any phase

2. ✅ Implemented - Code-graph links exist for this feature

   • Set by: /know:build creating bidirectional spec↔code links
   • Computed: Check for code-link references pointing to this feature
   • Auto-detected via graph traversal
   • Deprecated reference types: graph-link, implementation, product-component are deprecated. Use code-link instead.

3. ✅ Reviewed - Git commit with [feature:id] merged to main

   • Set by: Merging to main with feature tag in commit message
   • Computed: git log --grep "\[feature:auth\]" main
   • Pattern: feat: implement auth [feature:auth] in commit message

Workflow Integration

/know:add     → meta.phases[pending][feature:x]   → 📋 planned
/know:build   → creates code-links                → ✅ implemented
git merge     → [feature:x] in commit msg         → ✅ reviewed
/know:done    → removes from phases, archives     → done

Example Output

$ know feature status feature:auth

Feature Status: feature:auth

✅ Planned: Yes
   Phase: I
   Status: in-progress

✅ Implemented: Yes
   Modules: module:auth-handler, module:session-store

✅ Reviewed: Yes
   Commit: abc123f
   Date: 2026-02-13

✓ Feature is fully complete!

Important: Always include [feature:name] in merge commit messages to enable automatic reviewed status detection.

Requirements vs Todo.md

Requirements replace todo.md for progress tracking:

• Requirements are stored in meta.requirements
• Each feature links to requirement entities via depends_on
• Manage with know req commands (add, list, status, block, complete)
• Query all: know req list feature:x or know meta get requirements

Implementation Patterns

Discover Reference Types On Demand

know gen rules describe references      # List all reference types with descriptions
know gen rules describe <type>          # Deep dive on any type (entity or reference)
know gen rules after <entity-type>      # What can this entity depend on?
know gen rules before <entity-type>     # What can depend on this entity?

Run these before adding references. The rules file is the canonical source for what types exist and what they mean.

Cross-Graph Reference Type: code-link

code-link is the current reference type for linking spec and code graphs:

Graph	Schema
spec-graph	{ modules, classes, packages, status }
code-graph	{ feature, component, status }

Usage: Cross-graph link between spec entities (feature/component) and code entities (module/class).

Deprecated: graph-link, implementation, and product-component are deprecated cross-graph reference types. Use code-link instead.

Permissions (Access Control)

The permission reference type links users to features for access control:

"references": {
  "permission": {
    "admin-full-access": "*",
    "editor": ["feature:user-management", "feature:content-editor"],
    "viewer": ["feature:dashboard", "feature:reports"],
    "trusted-user": ["*", "!feature:admin-panel", "!feature:billing"]
  }
}

Users depend on permissions to define what they can access:

"graph": {
  "user:admin": {"depends_on": ["permission:admin-full-access"]},
  "user:trusted": {"depends_on": ["permission:trusted-user"]}
}

Permission values:

• "*" - All features
• ["feature:x", "feature:y"] - Specific features only
• ["*", "!feature:x"] - All features except those negated with !

External Artifact IDs

When a reference has a rendering in an external tool (Figma, Pencil, Storybook), store the external ID on the reference. This creates spec-to-design traceability.

• figma_id — Figma node or frame ID
• pen_file — Pencil .pen file path
• storybook_id — Storybook story identifier
• external_url — Generic link to external artifact

Applies to any reference type with an external rendering, not just interfaces.

Core Patterns

1. Screen → interface reference with route, identifiers, and external design ID when a rendering exists
2. Data-bearing feature → data-model reference. Features describe behavior; data-models describe shape
3. Multi-screen journey → sequence reference. One per journey, not per screen
4. Spec change → verify design. Design change → verify spec. Never update one in isolation
5. Requirements describe what. Decompose how into typed references (data-model, business_logic, sequence, api_contract)
6. Verify connectivity after every addition. know graph uses + know graph used-by + know graph check orphans
7. Keep implementation details out of requirements. That detail belongs in design artifact references
8. Extract shared references. Do not duplicate field definitions — one reference, multiple links

Critical Rules for LLMs

1. NEVER directly read/edit graph files - Always use know CLI commands (enforced by hooks)
2. Always validate after modifications - Run know graph check validate
3. Respect entity vs reference distinction - Entities participate in dependencies, references don't
4. Follow dependency rules - Use know gen rules to check before adding dependencies
5. Maintain DAG properties - No cycles allowed, check with know graph check cycles
6. Use full paths - Always use type:key format (e.g., feature:real-time-telemetry)
7. Never add dependencies to entity objects - Only in the graph section
8. Check completeness - Use know graph check gap-analysis to ensure full dependency chains
9. Use search for discovery - know search <pattern> is faster than reading the entire graph
10. Confirm destructive operations - Use -y flag to skip confirmation in automated scripts
11. Run write operations sequentially - The graph file is not concurrency-safe. Running know unlink, know add, know link, or any other write command in parallel (e.g., with & in bash) causes race conditions where the last writer overwrites earlier changes. Always chain write commands with && or run them in a single sequential script.

Installation Note

If know command is not found, run python3 know/know.py from the project root. See project INSTALL.md for setup.

---

Remember: The graph is dependency-driven. Use know gen rules to understand structure before making changes. Always validate after modifications.