Tooling
15 slash commands, 9 specialized agents, template library and quality rules
15 slash commands
quick and focused commands for documentation and development tasks - click to see examples
9 specialized agents
autonomous agents that orchestrate the 7-phase workflow
Analyst
phase 1, 3Discovery and specification
Architect
phase 2Architecture design (HIGH complexity)
Orchestrator
phase 3.5Task decomposition (CRITICAL)
Developer
phase 4Code implementation
Gatekeeper
phase 4Quality gates validation
Reviewer
phase 5Code review
Tester
phase 5Test validation
Documenter
phase 6Documentation generation
Guardian
phase 7Pre-commit validation
Complete 7-phase workflow
from user request to production-ready code
Phase 1: Discovery
analystUnderstands requirements, evaluates complexity (LOW/MEDIUM/HIGH) and creates changes/[id]/proposal.md
Phase 2: Architecture
architect(HIGH complexity only)Architecture design with C4 diagrams and creates changes/[id]/design.md
Phase 3: Specification
analystWrites Arc42 documentation + BDD scenarios + ADRs and creates changes/[id]/spec.md
Phase 3.5: Task Decomposition
orchestratorCRΓTICOBreaks specs into atomic tasks <100 LOC to prevent hallucinations and creates changes/[id]/tasks.md
Phase 4: Implementation
developer + gatekeeperImplements task by task with tests, validated by gatekeeper
Phase 5: Review
reviewer + testerCode review + test validation (β₯80% coverage required)
Phase 6: Documentation
documenterUpdates docs and README to match implementation
Phase 7: Validation & Release
guardian(optional)Final validation: constitution compliance, tests, linters, updated specs
Why Task Decomposition prevents AI hallucinations
The problem: Large specs cause quadratic attention degradation O(nΒ²)
Experimental data (Chronicle 10):
π Without specification: 60.4% hallucination rate
π Complete Arc42 spec (3,570 lines): 20.0% hallucination
β Large context = 625M attention operations
β 60% loss in middle of context ("Lost in the Middle")
π Spec + Task Decomposition (450 lines/task): 7.5% hallucination
β Small context = 6.25M operations per task
β <5% context loss
β 88% reduction in hallucination rate
The solution: Orchestrator breaks specs into ~100 LOC tasks, keeping scope bounded and context focused
Based on research: Liu et al. (2023) "Lost in the Middle", Vaswani et al. (2017) "Attention Is All You Need"
Templates & rules
Template library
Ready-to-use templates in Arc42 (12 chapters), C4 diagrams, BDD scenarios, and ADRs in Markdown, JSON and Obsidian formats
39 quality rules
Object Calisthenics (9 rules) + Extensions (30 rules) covering functional programming, package design and much more
Categories:
- β’ Object Calisthenics (001-009): Fundamental OOP principles
- β’ Functional Programming (010-019): FP patterns and immutability
- β’ Package Design (020-029): Module organization and dependencies
- β’ Testing (030-039): Quality patterns and test coverage
.claude/ directory structure
.claude/
βββ constitution.md # Governance principles
βββ commands/ # 15 slash commands
β βββ vision.md
β βββ feature.md
β βββ ...
βββ skills/ # 9 specialized agents
β βββ analyst/
β βββ orchestrator/
β βββ ...
βββ templates/ # Arc42, C4, BDD, ADR
β βββ arc42/
β βββ c4/
β βββ ...
βββ rules/ # 39 quality rules
βββ 001_nivel-unico-indentacao.md
βββ ...