Winning the
Hackathon
Ship more than any team in the room, using GitHub Copilot.
Prepare your agents in the evening. Let them build all night. Merge over coffee tomorrow. ☕
Julien Dubois
Principal Manager, Developer Relations at GitHub / Microsoft
github.com/jdubois · julien-dubois.com
Who am I?
Julien Dubois
Principal Manager, Developer Relations at GitHub / Microsoft
- Creator of JHipster · 22,000+ ⭐
- 200+ international talks (Devoxx, SpringOne, Microsoft Build…)
- Today: shipping real projects by managing AI agents
📊 My month with Copilot
One developer · ~30 days · real GitHub activity + Copilot usage data
759
commits authored
406
PRs merged (419 opened)
~443k
lines changed +347k / −96k
~14B
AI tokens / month *
Tokens by model · ~14B / month
Tokens: 95% served from cache
- 13.3B cache reads · 113M fresh input
- 90M output · 423M cache writes
- ~14B tokens / month total
- spread across 11 models
* Token usage extrapolated from 7 days of real Copilot data across all integrations (CLI, app, mobile, PR reviews); 95% is cache reads.
🚀 The mindset shift
You're not the developer, you're the manager now.
🧑✈️ You · the manager
🚀 Feature 1
the winning demo
🧩 Feature 2
more depth to win
🎨 UI
polish & landing
🏗️ Infra
deploy & plumbing
Many agents in parallel. You brief, review, merge.
You don't type faster. You ship wider.
🏁 The rules
⏱️ The constraint
- 2 days, hard stop
- You need to ship a real product
- You can't code all night and still ship on time
⚡ Your advantage
- You have GitHub Copilot
- A fleet of agents in parallel, not one chat window
- They keep building while you sleep
Same tools, same clock for everyone. The team that uses them best ships wider and wins.
🗓️ The 24-hour playbook
🌙 while you sleep
6pm10pm12am7am10am12pm
PrepScope it, brief your agents
LaunchSplit into parallel tasks, go
🌙 OvernightAgents build, test, open PRs
☀️ MorningMerge wins, drop the rest
Final pushPolish, integrate, demo-proof
🏁 ShipSubmit & present
Evening
2 hours of prep buys you 8 hours of unattended building.
Overnight
Agents work in parallel during the night. Your runway doubles.
Morning
Triage the overnight haul, merge the wins, re-brief for the day.
🛠️ Step 1: Prep your agents · evening
📜 Write it down · AGENTS.md
- Set the house rules once
- What you're building · conventions · how to build & test
- Every agent reads it first, no re-explaining
✅ Give them a test harness
- A real build + test command they can run
- They verify their own work before reporting back
- No tests = you can't trust the output
🛡️ Set guardrails
- One branch / worktree per task
- Small scope, clear goal
- Nothing hits main without your review
🧩 Split the work
- Break the build into independent pieces
- One task each: two features, UI, infra
- Parallel-ready = night-ready
Tip: the better the brief, the less you babysit at 2am.
☀️ Step 3: Review over coffee · morning
Your morning routine
- Triage the overnight haul: tens of PRs
- Merge the good ones fast
- Drop or re-task what didn't land
- Cherry-pick the good parts of the rest
- Re-brief and relaunch for the day
Wake up to progress
Tens of PRs are waiting as every agent looped all night. The build moved while you didn't. Then the loop repeats: brief → build → review.
📋 The one-page playbook
1Write the house rules (AGENTS.md) before you start a single agent.
2No agent without a test harness: give it a build + test it can run itself.
3One task per agent. Small scope, clear goal, own branch.
4Run them in parallel: that's the whole point. Don't babysit one.
5Let them build overnight while you sleep.
6Review in the morning: merge wins, drop failures, re-task fast.
7Pick the right model for each task; save what works as reusable skills.
If there's one slide to screenshot, this is it.
⚠️ Watch out for
Scope creep
"Build the whole app" makes an agent wander.
Fix: one tight goal per task.
Fix: one tight goal per task.
No tests, no trust
You can't read every line of tens of PRs.
Fix: harness first, green build before merge.
Fix: harness first, green build before merge.
Giant PRs
A 2,000-line PR is impossible to review.
Fix: small, reviewable chunks.
Fix: small, reviewable chunks.
Wrong model
A weak model won't solve the hard tasks; a strong one is slow and costly.
Fix: choose the right model for the job.
Fix: choose the right model for the job.
Most failed runs aren't the agent's fault. They're a briefing problem.
That's the playbook.
Now go build. 🚀
Prep in the evening · build all night · merge over coffee.
Brief them well, run them wide, and out-ship the room.
🌐 julien-dubois.com · 𝕏 @juliendubois · 🐙 github.com/jdubois
Find my conference talks
Scan the QR code for current talks, slides and workshop information.
julien-dubois.com/conferences