This is a submission for the GitHub Copilot CLI Challenge
What I Built
Say "Write me a novel" — and 13 AI agents will deliver a publication-quality story.
Novel Studio for Copilot CLI is a multi-agent novel creation system built entirely on GitHub Copilot CLI's native features. No external frameworks, no LangChain, no CrewAI — just AGENTS.md, .github/agents/, and hooks.
The idea started from a simple question: Is Copilot CLI just a coding tool? I discovered that its agent architecture — AGENTS.md auto-loading, @agent invocation, hooks lifecycle, and Agent Mode — forms a complete orchestration platform capable of creative workflows far beyond code generation.
The 13-Agent Team
| Phase | Agents |
|---|---|
| Planning (6) |
@main-writer · @story-writer · @character-writer · @setting-writer · @genre-specialist · @pacing-manager
|
| Writing (4) |
@dialogue-writer · @action-writer · @emotion-writer · @prose-writer
|
| Quality (3) |
@editorial-team · @feedback-agent · @research-agent
|
How it works
These 13 agents collaborate in a 3-phase pipeline:
Phase 1 — Planning: @main-writer coordinates while @story-writer, @character-writer, @setting-writer, and @genre-specialist work in parallel to build plot structure, character profiles, world-building, and genre analysis. @pacing-manager then designs the overall tempo.
Phase 2 — Writing: Each chapter is broken into scenes, and scene-type specialists take over: @dialogue-writer for conversations, @action-writer for combat/tension, @emotion-writer for introspection, and @prose-writer for general narrative.
Phase 3 — Quality Control: @editorial-team proofreads for consistency, @feedback-agent evaluates from 5 reader perspectives, and @research-agent fact-checks.
The user only intervenes at key checkpoints (5-7 times in Review mode). Everything else is automated.
User Experience
- 30 seconds to first novel — Clone, open Copilot, type "Write me a novel." That's it.
- 3 work modes — Auto (hands-off), Review (recommended, 5-7 checkpoints), Manual (full control). Beginners and power users both covered.
- Multilingual — English and Korean supported out of the box, with dedicated agent variants for each language.
-
Direct agent access — Call any specialist directly for advice, e.g.,
@genre-specialist What are typical cyberpunk tropes?
Why this matters
Multi-agent AI systems typically require heavyweight frameworks. This project demonstrates that Copilot CLI's native features are sufficient for full agent orchestration — routing, delegation, lifecycle hooks, and quality automation — with zero additional dependencies.
Demo
GitHub Repository: tiny-flowlab/novel-studio-copilot-cli
Sample Outputs (included in repository)
The repository contains two complete novels generated by the agent system:
- 📖 English:
projects/love_story_heroes/— 3-chapter romance novel → Read the full novel - 📖 Korean:
projects/ko/first_love_001— 3-chapter romance novel (demonstrates multilingual support)
Each project folder contains the full pipeline output — from planning documents to editorial reports to final manuscripts — so you can trace exactly how the 13 agents collaborated.
Quick Start
git clone https://github.com/tiny-flowlab/novel-studio-copilot-cli.git
cd novel-studio-copilot-cli
# Open in VS Code with Copilot, or run: gh copilot
# Then type: "Write me a novel"
My Experience with GitHub Copilot CLI
Everything in this project was built with GitHub Copilot CLI. The AI models used were Claude Opus 4.6 and Sonnet 4.5.
Before this challenge, I had only used GitHub Copilot through VS Code's Copilot Chat. This challenge gave me the opportunity to experience Copilot CLI in depth for the first time. Compared to other CLI tools like Claude Code, Codex, and Gemini CLI, Copilot CLI proved to be remarkably powerful in terms of agent control. Features like AGENTS.md auto-loading, the .github/agents/ directory convention, @agent invocation, and the hooks lifecycle enabled multi-agent orchestration without any external framework.
Through building this project, I came to understand what "agents" really mean in practice — role definition, delegation, collaboration, and quality gates.
Specific Copilot CLI Features Used
1. AGENTS.md Auto-Loading — The Entry Point
Copilot CLI automatically injects AGENTS.md into the system prompt when you open a project. I designed this file as the Sales Representative agent — the front door to the entire system. The moment a user enters the project directory and opens Copilot, the novel studio activates automatically. No setup commands, no configuration — just start talking.
2. .github/agents/ + @agent Invocation — The Backbone
Each of the 13 agents lives as an .agent.md file in .github/agents/. Copilot CLI recognizes this convention and makes every agent callable via @agent-name. This enabled a hierarchical delegation pattern: @main-writer coordinates the entire project, calling @story-writer for plot, @dialogue-writer for conversation scenes, @editorial-team for proofreading — all through simple @name references.
This is the same orchestration pattern that frameworks like LangChain implement with hundreds of lines of Python — achieved here with just markdown files and Copilot's native agent system.
3. Hooks — Automated Quality Gates
The hooks/ directory provided lifecycle automation that was critical for novel quality:
-
preToolUse.sh— Checks character/setting consistency before every file edit -
postToolUse.sh— Runs spell check, readability analysis, and auto-backup after every save -
sessionStart.sh— Restores previous project context when resuming work
These hooks transformed Copilot CLI from a writing tool into a quality-assured writing pipeline.
4. Agent Mode — True Autonomy
Agent Mode allowed Copilot to execute multi-step workflows autonomously: creating project directories, writing files, running Python quality scripts, and delegating to sub-agents — all without manual intervention between steps. This was essential for the 3-phase pipeline to work as a cohesive system.
5. Model Selection
Being able to select different AI models (ChatGPT, Claude, Gemini) directly in Copilot settings was valuable for optimizing each phase — faster models for planning iterations, higher-quality models for final prose writing.
Key Takeaway
GitHub Copilot CLI is not just a coding assistant — it's a general-purpose agent orchestration platform. The combination of AGENTS.md auto-loading, @agent invocation, lifecycle hooks, and Agent Mode provides everything needed to build sophisticated multi-agent systems for creative (and non-creative) workflows, without any external framework.
Top comments (0)