Skip to content

wilsonkichoi/agentic_development_workflow

Repository files navigation

Agentic Development Workflow

A structured, 5-phase framework for building software with AI coding agents. Gives LLMs clear specifications, atomic tasks, and role-matched expertise — so you get predictable, high-quality output instead of hoping the AI figures it out.

Note: I created this project to better understand agentic software development methodology. There may be a bunch of stupid stuff here. For real projects, check out Superpowers, GSD, Spec Kit, EPCC, oh-my-claudecode, Compound Engineering, Agent Skills and incorporate whatever works for you.

The Problem

Most people use AI coding tools by throwing a vague prompt at an LLM and hoping for the best. This works for small tasks but falls apart for anything real:

  • The AI makes assumptions instead of asking
  • Specifications live in your head, not in documents the AI can read
  • Tasks are too big, so context degrades and quality drops
  • No verification — you merge code you haven't meaningfully reviewed
  • Knowledge is lost between sessions

How This Helps

This workflow breaks AI-assisted development into 5 phases with human gates between each:

Phase 1        Phase 2          Phase 3           Phase 4         Phase 5
Research  -->  Specification  -->  Task Breakdown  -->  Execution  -->  Verification
          [GATE]             [GATE]               [GATE]          [GATE]

Phase 1: Research — Human gathers materials, AI performs deep research to fill gaps and cross-reference findings.

Phase 2: Specification — AI generates architecture specs (SPEC.md) with milestones, schemas, API contracts, and diagrams. No code. Optional multi-role review (security, domain, product).

Phase 3: Task Breakdown — Spec is decomposed into atomic tasks grouped into parallel waves, with role assignments and test separation.

Phase 4: Execution — One task per session, role-matched agents, automated verification, human review via task review files, PR on approval.

Phase 5: Verification — Per-milestone end-to-end testing and retrospective.

Each phase produces persistent documents (SPEC.md, PLAN.md, PROGRESS.md, decision records, task reviews) that survive across sessions and prevent context rot.

What's Included

Component Count Description
Skills 8 init-project, research, spec, plan, execute, review, verify, auto
Agents 13 Software Architect, Backend Engineer, QA Engineer, Security Reviewer, and more

Installation

Claude Code Plugin (recommended)

# Add the marketplace
/plugin marketplace add wilsonkichoi/agentic_development_workflow

# Install the plugin — select "Install for you, in this repo only (local scope)"
/plugin install agentic-dev@wilsonkichoi-agentic-dev

When prompted for install scope, choose "Install for you, in this repo only" (local scope). This keeps the plugin scoped to your project rather than applying globally, which is recommended while the workflow is still evolving.

After installation:

  • Skills are available as /agentic-dev:research, /agentic-dev:spec, etc.
  • Agents appear in /agents (e.g., agentic-dev:backend-engineer)
  • Auto-updates when the repo is updated

Copilot CLI

copilot plugin install wilsonkichoi/agentic_development_workflow

After installation:

  • Skills are available as /init-project, /research, /spec, /plan, /execute, /review, /verify, /auto
  • Agents appear in /agent (e.g., backend-engineer, qa-engineer)
  • Plugins install globally to ~/.copilot/state/installed-plugins/

Note: Copilot CLI does not support per-project plugin scoping. Installed plugins are always global — there is no --scope flag or disable/enable per workspace (unlike Gemini CLI). If you need project-level control, use Claude Code's local scope install or Gemini CLI's workspace scope instead.

Gemini CLI

To avoid polluting the global context, install the extension globally but keep it disabled by default. Enable it strictly on a per-workspace basis:

# 1. Install the extension from GitHub
gemini extensions install https://github.com/wilsonkichoi/agentic_development_workflow

# 2. Disable the extension globally (user scope) so it doesn't pollute all your projects
gemini extensions disable agentic-development-workflow --scope user

# 3. Navigate into specific projects and enable it locally (workspace scope)
cd /path/to/your-project
gemini extensions enable agentic-development-workflow --scope workspace

After installation:

  • Skills are available directly via the CLI (e.g., gemini plan).
  • Subagents are available via the @ mention or interactively (e.g., gemini --agent=backend-engineer).

Local Plugin Testing

claude --plugin-dir /path/to/agentic_development_workflow

Updating

Claude Code Plugin

# Update marketplace listings
/plugin marketplace update wilsonkichoi-agentic-dev

# Or enable auto-updates in /plugin → Marketplaces → Enable auto-update

Important: After updating, start a new session (/clear or new terminal). Skill file paths are cached per session and /reload-plugins may not fully refresh them.

Copilot CLI

# Update the plugin
copilot plugin update agentic-dev

# View installed plugins
copilot plugin list

# Uninstall
copilot plugin uninstall agentic-dev

Gemini CLI

# Update the extension
gemini extensions update agentic-development-workflow

# View installed extensions
gemini extensions list

# Uninstall
gemini extensions uninstall agentic-development-workflow

Usage

Starting a New Project

# From parent directory
/agentic-dev:init-project ./my-project

# Already inside the project directory
/agentic-dev:init-project .

This creates:

my-project/
├── CLAUDE.md                 # Project instructions (edit first)
├── workflow/
│   ├── research/manual/      # Place human research here
│   ├── research/final/       # AI research output
│   ├── spec/                 # SPEC.md, HANDOFF.md
│   ├── plan/                 # PLAN.md, PROGRESS.md
│   │   └── reviews/          # Per-task review files
│   ├── decisions/            # Architecture Decision Records
│   └── retro/                # Retrospectives
├── src/
└── tests/

Running Phases

Phase Command What It Does
1. Research /agentic-dev:research AI deep research on materials in workflow/research/manual/
2. Specification /agentic-dev:spec Generate SPEC.md with milestones, schemas, contracts
3. Task Breakdown /agentic-dev:plan Decompose into atomic tasks with wave grouping
4. Execution /agentic-dev:execute Implement one task with role-matched agent
4b. Review /agentic-dev:review Code review, fix plan validation, fix verification
5. Verification /agentic-dev:verify End-to-end testing + retrospective
Auto /agentic-dev:auto Full pipeline (execute+review+fix+verify) per wave or milestone

Using Agents (Phase 4)

During execution, tasks are matched to specialized agents:

Task Type Agent
API endpoints, services agentic-dev:backend-engineer
UI components agentic-dev:frontend-developer
Database migrations agentic-dev:data-engineer
Test implementation agentic-dev:qa-engineer
Auth, encryption agentic-dev:security-engineer
Infrastructure, CI/CD agentic-dev:devops-engineer

Test tasks run in separate sessions from implementation tasks — the QA agent sees only the spec and public interface, not the implementation. This prevents the AI from writing tests that match its code rather than testing the contract.

Key Concepts

Human gates — Every phase produces reviewable artifacts. Nothing moves forward without human approval.

One task per session — Fresh context for each task. No context rot.

Decision records — Significant decisions are captured in workflow/decisions/DR-NNN-*.md with options, rationale, and threaded human/AI discussion.

Task review files — Every task gets workflow/plan/reviews/task-X.Y.md with a work summary and space for *FEEDBACK:* / *AI:* discussion before PR creation.

Review loop — After execution, /agentic-dev:review produces durable review files (wave-mM-N.md or appends to task-X.Y.md). Any AI session in any tool can read the review file and continue the loop: review issues, validate fix plans, verify fixes. No copy-paste between sessions — the review file is the connective tissue. Review and fixes operate on the wave's feature branch; the feature branch merges to main only after all issues are resolved.

Parallel waves — Independent tasks run concurrently when they belong to different components, have frozen API contracts, and share no mutable state.

Test separation — Tests are written by a QA agent in a separate session from the implementation agent, testing the spec contract rather than the implementation.

Contributing

See CONTRIBUTING.md for development setup, testing checklist, and PR guidelines.

Documentation

  • WORKFLOW.md — Full workflow reference with all phases, principles, cost optimization, and tool-specific guides
  • CI_CD.md — CI/CD pipeline design notes

References

  • Superpowers — composable skills, TDD-first, subagent-driven execution
  • GSD (Get Shit Done) — meta-prompting, wave-based parallel execution, context rot mitigation
  • Spec Kit — spec-driven development, multi-agent support, extension ecosystem
  • EPCC Workflow — explore-plan-code-commit, session continuity, feature tracking
  • oh-my-claudecode — reference for skill implementation and architecture
  • Anthropic Skills — official skills reference
  • Agency Agents — role-based agent definitions (adapted into agents/)
  • PUA — debugging persistence framework (adapted into Phase 4)
  • Compound Engineering — Agent-Driven Development using the power of LLM.
  • Agent Skills — Production-grade engineering skills for AI coding agents.

License

MIT

About

No description, website, or topics provided.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors