Skip to content

warren830/super-aidlc

Repository files navigation

Super-AIDLC

Stop vibe coding. Start engineering.

Super-AIDLC is a structured development skill for AI coding agents (Claude Code, Kiro). It routes tasks by complexity, designs before coding, builds with TDD in parallel, reviews with specialist agents, and compounds knowledge across sessions -- with security hardening enabled by default.

Chinese / 中文 | Blog | Benchmarks

Why Another AI Workflow?

We benchmarked 4 approaches on identical tasks (full results):

Approach Speed Tests Security Vulns Design Docs
Raw (no methodology) 4 min 33 Shell injection, path traversal, memory leak None
Superpowers 14 min 69 Shell injection, path traversal, memory leak None
AIDLC-workflows 9 min 49 Shell injection, path traversal, memory leak 13 files (audit)
Super-AIDLC 16 min 85 None 2 files (design + build log)

Super-AIDLC is the only approach that produces code with zero known security vulnerabilities.

Cross-session benchmark (full results): After 3 sessions on DevLake, Track A (Super-AIDLC) was 13% faster than Track B (Superpowers) in Session 3, with zero issues -- because 2 sessions of accumulated knowledge provided a clear implementation path.


Quick Start

Install

git clone https://github.com/warren830/super-aidlc.git ~/super-aidlc

# Global install (all projects)
~/super-aidlc/adapters/claude-code/install.sh --global

# Or single project
~/super-aidlc/adapters/claude-code/install.sh /path/to/your/project

# Kiro
~/super-aidlc/adapters/kiro/install.sh /path/to/your/project

# Verify
~/super-aidlc/adapters/claude-code/install.sh --verify --global

Symlink-based -- cd ~/super-aidlc && git pull updates all installed projects instantly.

Your First Session

cd /your/project
claude

# Just describe what you want:
/super-aidlc Build a user authentication system with JWT

That's it. Super-AIDLC will auto-detect complexity, ask structured questions, produce a design doc, build with TDD, review, verify, and offer to ship.


Usage Guide

Daily Workflow (3 commands you'll use most)

# 1. Build anything
/super-aidlc implement rate limiting for the API

# 2. After a hard debugging session or solving a tricky problem:
/super-aidlc:compound

# 3. Weekly knowledge hygiene:
/super-aidlc:janitor --days=7

All 11 Commands

Core Pipeline

Command When to Use Example
/super-aidlc [task] Any development task /super-aidlc add WebSocket support

This is the main entry point. It auto-detects complexity and routes:

Complexity Auto-detected When What Happens
Light Bug fix, config change, ≤2 files TDD build → review → verify
Medium New feature, moderate scope Questions → design doc → parallel build → review → verify
Heavy New system, multi-component Brainstorm → research → full design → parallel worktree builds → specialist review → verify

Flags: --light --medium --heavy (force complexity), --dry-run (preview), --lang=zh (Chinese docs), --skip-review (escape hatch), --auto (fully automated after design approval)

Before Building

Command When to Use Example
/super-aidlc:brainstorm [idea] Vague idea, need to explore /super-aidlc:brainstorm I want some kind of notification system
/super-aidlc:design [task] Want design doc without coding /super-aidlc:design payment processing module

Brainstorm asks WHO/WHAT/WHY, then runs 3 parallel lenses (Product: "right problem?", Engineering: "hard risks?", Design: "UX issues?") on each approach. Defines scope, outputs requirements.md that feeds into the main pipeline.

Design runs full inception (research → questions → design doc → review) but stops before writing code. Use when you want to review the architecture before committing.

During / After Building

Command When to Use Example
/super-aidlc:review [scope] Review current changes /super-aidlc:review backend/api/
/super-aidlc:debug [bug] Systematic bug investigation /super-aidlc:debug login returns 401 for valid tokens
/super-aidlc:qa [url] Test a running app /super-aidlc:qa http://localhost:3000
/super-aidlc:ship [branch] Verify + commit + push + PR /super-aidlc:ship

Review runs two stages: spec compliance (did you build what was asked?) then up to 8 parallel specialist reviewers (correctness, quality, maintainability always-on + security, performance, reliability, API contract, adversarial conditionally selected).

Debug follows the Iron Law: investigate → analyze → hypothesize → implement. No shotgun fixes. Always produces a regression test.

QA auto-detects mode (browser/API/CLI) and tests user flows with evidence (screenshots, response bodies).

Ship verifies all tests/build/lint pass, creates meaningful commits, pushes, and opens a PR with summary.

Knowledge Management

Command When to Use Example
/super-aidlc:compound [context] After solving a hard problem /super-aidlc:compound
/super-aidlc:compound-refresh [scope] After refactors or migrations /super-aidlc:compound-refresh performance-issues
/super-aidlc:janitor [--days=N] Periodic knowledge scan /super-aidlc:janitor --days=7
/super-aidlc:metrics [--days=N] See how you're improving /super-aidlc:metrics --days=30

Compound extracts structured solutions from the current session (Problem → What Didn't Work → Solution → Prevention) into aidlc-docs/solutions/. Cross-project generic solutions auto-promote to ~/.aidlc/global-solutions/.

Compound-refresh reviews existing solutions against current code. Five actions: Keep, Update, Consolidate, Replace, Delete. Use after refactors that may have invalidated prior knowledge.

Janitor scans build logs and auto-scores sessions (debugger invoked? multiple fix attempts? new patterns?). High-value sessions get auto-compounded. Low-value sessions are skipped. Run weekly.

Metrics generates trend reports: are you getting faster? fewer bugs? more tests? Shows which strategies (INLINE/SERIAL/PARALLEL) work best for your projects.

Knowledge System (How It Gets Smarter)

Session 1:  aidlc-docs/ empty → starts from scratch
Session 2:  Researcher reads patterns.md + build logs → avoids Session 1's mistakes
Session 3:  Researcher reads solutions/ + patterns.md + logs → leverages all prior knowledge
Session N:  Four-layer search → project conventions → project solutions → global solutions → build history

Your knowledge accumulates in:

aidlc-docs/                              # Per-project (auto-created)
  patterns.md                            # Conventions (≤50 lines, always read first)
  solutions/                             # Structured knowledge base
    runtime-issues/                      # Bug fixes
    patterns/                            # Best practices
    config-issues/                       # Configuration gotchas
    testing-issues/                      # Testing insights
    ...
  2026-04-05-feature-name/
    design.md                            # Architecture + error map + decisions
    build-log.md                         # Session history + structured metrics

~/.aidlc/                                # Cross-project (auto-created)
  global-solutions/                      # Knowledge shared across all projects

Typical Workflows

New feature (Medium):

/super-aidlc add user profile editing with avatar upload
→ asks ~5 groups of questions
→ produces design doc with architecture diagram
→ parallel builds 2-3 units with TDD
→ spec review → parallel quality review
→ auto-verify (test + build + lint)
→ offers to ship
→ auto-evaluates compound score

Bug fix (Light):

/super-aidlc:debug the email queue is dropping messages under load
→ investigate → reproduce → root cause
→ regression test → fix → verify
→ auto-compound if score >= 3

Exploring a new idea:

/super-aidlc:brainstorm I want real-time collaboration like Google Docs
→ WHO/WHAT/WHY questions
→ 2-3 approaches (WebSocket vs SSE vs polling)
→ scope definition
→ outputs requirements.md

/super-aidlc:design real-time collaboration
→ reads requirements.md automatically
→ produces design doc

/super-aidlc build real-time collaboration
→ reads design doc automatically
→ builds it

Weekly maintenance:

/super-aidlc:janitor --days=7              # compound any unprocessed valuable sessions
/super-aidlc:compound-refresh              # check for stale knowledge
/super-aidlc:metrics --days=30             # are we improving?

Fully automated (--auto):

/super-aidlc --auto add rate limiting with Redis backend
→ inception: asks questions, produces design doc (interactive)
→ user approves design
→ [AUTO] construction runs as fresh subagent (isolated context)
→ [AUTO] gate check: all units DONE? reviews PASS? tests green?
→ [AUTO] operations runs as fresh subagent (verification + commit)
→ asks user about push/PR (irreversible actions always need confirmation)

Auto mode solves context degradation in long sessions. Each stage runs in a fresh subagent so no agent accumulates stale context. Green = auto-proceed, red = pause for human.

Updating

cd ~/super-aidlc && git pull
# All installed projects update instantly (symlinks)

Uninstalling

~/super-aidlc/adapters/claude-code/uninstall.sh /path/to/your/project
# Or for global: remove ~/.claude/skills/super-aidlc* directories

Note: aidlc-docs/ in your project is NOT removed -- it contains your design docs and accumulated knowledge.


Five Iron Laws

  1. No code without a failing test first. Violations get deleted.
  2. No fixes without root-cause investigation. No shotgun debugging.
  3. No completion claims without evidence. "Should work" is not evidence.
  4. No shipping without all-green verification. Auto-fix up to 3 times.
  5. No unsanitized input to shell/filesystem/templates. Security is default-on.

Anti-Laziness System (v4.2)

Three layers prevent AI agents from cutting corners:

  1. Richer Unit Specs -- each unit in the design doc has GIVEN/WHEN/THEN acceptance criteria, required test scenarios (happy/error/edge), and explicit "Done means" definition.
  2. Builder Self-Discipline -- builders must produce an Acceptance Criteria Coverage table mapping each AC to its test (file:line), meet minimum test depth requirements, and pass an anti-skeleton scan.
  3. Reviewer Depth Enforcement -- spec reviewers verify each AC with file:line evidence. Quality reviewers must cite code for key security/correctness checks. Zero-finding PASS triggers a mandatory Confidence Check proving the reviewer actually read the code.

Architecture

Brainstorm:    WHO → WHAT → WHY → Approaches → Scope (optional)
                  ↓
Inception:     Parallel Research (4 agents) → Questions → Design Doc → Approve
                  ↓
Construction:  [U1] [U2] [U3]  ← INLINE / SERIAL / PARALLEL (auto-selected)
                  ↓    ↓    ↓
               Spec Review → Parallel Quality Reviews → Merge
                  ↓
Verify:        Test → Build → Lint → (fail? → debugger → retry x3) → All green
                  ↓
Ship:          Commit → Push → PR
                  ↓
Compound:      Score session → auto-extract if valuable → aidlc-docs/solutions/

18 Agents

Agent Role
Researcher Codebase patterns + four-layer knowledge search
Learnings Researcher Search solutions knowledge base
Git History Analyzer Code evolution + hotspot detection
Best Practices Researcher External patterns + framework docs
Architect Design doc producer (no code)
Builder TDD builder + self-check protocol
Design Reviewer Independent design review (Heavy)
Spec Reviewer Stage 1: built what was asked?
Quality Reviewer Stage 2: overall quality gate
Correctness Reviewer Logic errors, edge cases, state bugs
Security Reviewer Vulnerabilities, exploits, OWASP
Performance Reviewer N+1, memory, scalability
Maintainability Reviewer Coupling, complexity, naming, dead code
Reliability Reviewer Error handling, retries, timeouts, failure modes
API Contract Reviewer Breaking changes, response shapes, versioning
Adversarial Reviewer Failure scenarios, attack vectors
QA Agent Browser/API/CLI testing with real Chromium
Debugger Root-cause investigation (4-phase)

Project Structure

super-aidlc/
  VERSION                           # 4.1.0
  SKILL.md                          # Entry point
  phases/                           # brainstorm, inception, construction, operations
  agents/                           # 18 specialized agents
  skills/                           # 10 slash commands
  guards/                           # careful, freeze, verification
  rules/                            # tdd, review-protocol, anti-patterns, overconfidence
  extensions/                       # security-baseline
  adapters/                         # claude-code, kiro install scripts
  docs/                             # blogs, benchmarks

Credits

Built on ideas from:

  • AIDLC-workflows (AWS Labs) -- adaptive workflow steering rules for AI agents, three-phase lifecycle (Inception→Construction→Operations), design-first with "Bolts" rapid iteration
  • Superpowers (Jesse Vincent) -- composable skills for coding agents, non-negotiable TDD + verification-before-completion, subagent-driven development with two-stage review
  • gstack (Garry Tan) -- virtual engineering team as slash commands (15 specialists + 6 power tools), Think→Plan→Build→Review→Test→Ship→Reflect sprint cycle, real-browser QA with Playwright
  • Compound Engineering (Every Inc) -- "each unit of engineering work should make subsequent units easier", knowledge compounding via documented solutions, 80% planning+review / 20% execution

License

MIT

About

Structured AI development skill combining AIDLC-workflows + Superpowers + gstack. Supports Kiro and Claude Code.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors