Node.js Project Architecture for AI-Friendly Development

Architecture standards that keep files small enough for AI agents to read/edit without blowing the context window.

Core Rules

• Single file max 400 lines, index.html max 200 lines, server.js entry max 100 lines
• All tunable values in config.json, loaded at runtime, editable via admin dashboard
• Backend: routes/ by domain, services/ for shared logic, db.js for database
• Frontend: HTML skeleton only, JS/CSS in separate files
• Every project gets admin.html + routes/admin.js for config hot-reload

Project Type Selection

Determine project type, then read the corresponding reference:

Quick Start (All Types)

1. Identify project type from table above
2. Read the corresponding reference file
3. Create directory structure per the reference
4. Extract hardcoded values → config.json
5. Split large files by function (each <400 lines)
6. Add routes/admin.js + admin.html
7. Frontend: config.js fetches /api/config at startup, code reads GAME_CONFIG.xxx or APP_CONFIG.xxx
8. Test locally → backup → deploy

config.json Pattern (Universal)

// Server: load and serve config
const config = JSON.parse(fs.readFileSync('./config.json', 'utf8'));
app.get('/api/config', (req, res) => {
  const safe = { ...config };
  delete safe.admin; // strip secrets
  res.json(safe);
});

// Admin: hot-reload
app.post('/admin/config', requireAdmin, (req, res) => {
  fs.writeFileSync('./config.json.bak', fs.readFileSync('./config.json'));
  fs.writeFileSync('./config.json', JSON.stringify(req.body, null, 2));
  Object.assign(config, req.body);
  res.json({ ok: true });
});

Admin Dashboard Pattern (Universal)

admin.html auto-generates form from config structure:

• Password login (x-admin-password header)
• Visual config editor with save + hot-reload
• Stats overview (users/data/uptime)
• Config backup history + restore

Why This Matters

Large single files consume massive context tokens when AI reads them:

• 3000-line file → ~40K tokens per read (20% of 200K window)
• 200-line module → ~2.7K tokens per read (1.3% of window)
• Result: 10-15 productive rounds vs 3-5 before context compression