AI4Dev Lab — Small model. Better context. Better output.
Stack: Phaser 3 + TypeScript + Vite
Model constraint: Small models only — gpt-5-mini, gpt-5.4-mini, claude-haiku-4-5, gemini-3-flash
CLI: OpenCode or GitHub Copilot
In GitHub Copilot:
gpt-5-minicosts 0× (free),gpt-5.4-mini/claude-haiku-4-5/gemini-3-flashcost 0.33×.
Presented by @bertrandgressier & @ThomasRumasLM
Which files to fill in?
Tool Agent stubs Prompt/Command stubs OpenCode .opencode/agent/*.md.opencode/command/*.md(already wired)GitHub Copilot .github/agents/*.md.github/prompts/*.prompt.md
git clone https://github.com/bertrandgressier/devsummit2026-ai-lab.git
cd devsummit2026-ai-lab
npm install
npm run dev # → http://localhost:3000https://opencode.ai/docs/#install
npm install -g opencode-ai
opencode providers login select github-copilot → opens browser → login to GitHub → authorize device
Running opencode in your terminal
opencodeThen select a small model: In opencode settings (/models), select a small model only:
github-copilot/gpt-5-mini(0×)github-copilot/gpt-5.4-mini(0.33×)github-copilot/claude-haiku-4-5(0.33×)github-copilot/gemini-3-flash(0.33×)
Forbidden: claude-sonnet, claude-opus, gpt-4o, gemini-pro
Open the project in VS Code with the GitHub Copilot extension.
Use Copilot Chat in agent mode — select a small model in Copilot Chat settings.
Agents: invoke via @<agent-name> in Copilot Chat.
Prompts: run via the Copilot Chat prompt runner or / prefix.
| OpenCode command | Copilot prompt | What it does |
|---|---|---|
/plan <game description> |
plan.prompt.md |
Generate an implementation plan (architect) |
/design <game + mood> |
design.prompt.md |
Generate visual specs (designer) |
/code [instruction] |
code.prompt.md |
Implement plan files in order, or act on an explicit instruction |
/qa <game name> |
qa.prompt.md |
Review all game files, save report to .agents/qa-report.md |
/archive |
archive.prompt.md |
Archive the current spec before the next game |
OpenCode: commands live in .opencode/command/
Copilot: prompts live in .github/prompts/
Four roles — fill in the stub that matches your tool:
| Agent | Role | OpenCode | Copilot |
|---|---|---|---|
architect |
Reads game description + ARCHITECTURE.md → produces ordered plan saved to .agents/spec.md |
.opencode/agent/architect.md |
.github/agents/architect.md |
designer |
Produces DESIGN.md with exact colors, layout, animations | .opencode/agent/designer.md |
.github/agents/designer.md |
coder |
Implements all spec files in order (skips already-done), or acts on an explicit instruction | .opencode/agent/coder.md |
.github/agents/coder.md |
qa |
Reviews code against ARCHITECTURE.md rules, saves report to .agents/qa-report.md, never modifies files |
.opencode/agent/qa.md |
.github/agents/qa.md |
One skill in .agents/skills/:
| Skill | What it provides |
|---|---|
phaser-patterns |
Phaser 3 code patterns (interactive objects, tweens, transitions...) |
OpenCode: skills are declared in agent frontmatter (skills: [phaser-patterns]).
Copilot: instruct the agent to read .agents/skills/phaser-patterns/SKILL.md at startup.
.agents/
skills/ ← reusable knowledge injected into agents
spec.md ← generated plan (gitignored)
qa-report.md ← latest QA report (gitignored)
.opencode/
agent/ ← OpenCode role definitions
command/ ← OpenCode slash commands
.github/
agents/ ← Copilot agent definitions
prompts/ ← Copilot reusable prompts
checkpoints/ ← reference implementations (facilitator unlocks)
src/ ← Phaser project (implement inside games/)
AGENTS.md ← auto-loaded by every agent, maps key project files
ARCHITECTURE.md ← folder structure, mandatory rules, scene types
LAB.md ← step-by-step lab guide
| Branch | What's in it |
|---|---|
checkpoint-1 |
All agents pre-filled, Tic Tac Toe working — start here for Step 9 |
bonus-orchestrator |
Full orchestrator setup with all subagents configured |
git fetch origin && git checkout checkpoint-1npm run build # must exit 0