Hackathon Kickoff

A single Claude that reawakens on a schedule (5 min, 30 min, daily) for repetitive side work.

• ·Runs locally — it's a single Claude that re-enters on the interval you set.
• ·Boris has dozens running at once; uses it daily for work that isn't the main thing being built.
• ·Example loops he runs: one that babysits a PR (auto-rebase, fix CI, address review comments until it ships), one that cleans up CI by fixing or deleting consistently failing tests every 30 minutes.
• ·Fits any repetitive task that doesn't need the builder's direct attention.

A loop that runs server-side, so it keeps going after you close your laptop.

• ·Runs on the server rather than on your machine — closing the laptop doesn't stop it.
• ·Designed for work that should happen across a whole codebase or team, not per-engineer.
• ·Can hook into MCPs — GitHub, Slack, data queries — to compose automations.
• ·Framed as 'Agent SDK on steroids' for scheduled work.
• ·Boris notes this surface as one no one has fully explored yet.

Permission prompts are routed to a separate Claude that approves or rejects them using conversation context.

• ·Activate by pressing Shift+Tab roughly three times.
• ·Requires a Max plan or API credits — not available on Pro.
• ·Every permission request inside Claude Code is sent to a second Claude that sees the conversation and decides automatically.
• ·Removes the need to sit and press 'Yes, continue' for long-running tasks — give Claude a job, come back later, task done.
• ·Boris positions this as distinct from dangerous mode; he doesn't use dangerous mode because the model occasionally does something unintended.
• ·Makes it practical to run a second Claude on another task in parallel.

4.7 follows instructions more literally than 4.6 — articulate the goal more clearly and it'll stay on track.

• ·4.7 listens more precisely than 4.6 and follows instructions more tightly.
• ·The practical consequence: vague prompts that 'worked' with 4.6 underperform here; sharpening the prompt yields the bigger step-up.
• ·Paired with its long-running stamina, a well-specified goal will run to completion without drifting.
• ·Boris's framing: if you use 4.7 the same way as 4.6, you won't feel the jump.

4.7 stays on the rails for 30 minutes, hours, or multi-day runs where earlier models drifted.

• ·Boris reports running 4.7 for multiple days continuously with the model staying on task.
• ·The combination that enables this in practice: /loop or routines for cadence, auto mode so permissions don't gate progress, precise prompts so the goal doesn't drift.
• ·Same capability that shipped Claude Design internally.

Start a second Claude on a different task while the first one runs in auto mode.

• ·With auto mode removing the need to babysit permission prompts, you can hand off one task and open another Claude on something else.
• ·Combines with /loop and routines: background work happens on schedule while you work interactively on something else.

• ·Aimed at real-world problems, niche interests, things that take weeks today and should take hours.
• ·Jason's note: the last hackathon's winners skewed toward domain specialists who code on the side, not full-time engineers — the pattern is to build from what you know best.

• ·Aim for where the models are going, even if they're not quite there yet.
• ·Jason's heuristic: if it's easier to demo than to explain, you're on the right track.

• ·Most Creative Opus 4.7 Exploration — using 4.7 as a creative medium, not just a tool.
• ·Keep Thinking — for projects whose idea visibly evolved; judges want to see how the thinking pushed forward past the first idea.
• ·Best Use of Claude Managed Agents — graded on how well the Claude platform and its managed agents were leveraged; describe this explicitly in the submission.

• ·Code must be open source and the repo accessible to judges.
• ·New work only — start from scratch at the hackathon; submit only what was built during the event.
• ·Three-minute demo video is required. Ivan suggests Loom, OBS, or CapCut; upload to YouTube or a public drive as MP4.
• ·Written description should detail how you used Claude, how managed agents worked, how you approached creativity, and how you pushed the thinking.
• ·No projects that violate legal, ethical, or platform policies, or use code/data/assets without rights.
• ·Team size: solo or up to two people.

• ·Impact — real-world potential of the project and who benefits.
• ·Demo — is it impressive, is it cool to watch; Ivan calls this out as key, which is why the video deserves real time.
• ·Opus 4.7 use — how well 4.7 was used in the project; detail this in the write-up.
• ·Depth and execution — how the team pushed past the first idea.

• ·Submissions due April 26 at 8:00 PM EST.
• ·Round one: asynchronous review by a panel from the Claude Code and Anthropic teams.
• ·Round two: top six demos replayed on April 28 at noon EST; judges including Boris and Jason give feedback and deliberate the podium + side prizes live.

• ·Office hours channel: 5:00–6:00 PM EST daily, staffed by a Claude Code team member.
• ·Wed 12–1 PM ET: AMA with Thariq.
• ·Thu: Claude Managed Agents overview with Michael Cohen.
• ·Fri: session with Mike Brown, winner of the previous hackathon.
• ·Sun: session with Michal, third-place winner of the previous hackathon, on going viral post-event.

• ·$500 in Anthropic API credits emailed individually to each accepted applicant.
• ·Provisioned 'in a few hours' after kickoff; Ivan's team is in Discord with CV status if anything goes wrong.
• ·Usable throughout the event.