SF AI Hackathon Playbook

Researched 2026-07-05. Goals: connect with builders, learn how others build, meet hiring managers/recruiters, and ship demos that become talking points. Winning is a side effect.

Companion docs: IDEAS.md (duration-tiered backlog) · INGREDIENTS.md (reusable code pantry + pre-event checklist). Kickoff skill: /hack <idea> <duration>.

---

1. Define your win condition

You're not optimizing for prizes. A successful event =

1. One working demo shipped and pitched (the artifact + the rep).
2. Three real conversations where you learned how someone else builds agents.
3. One follow-up sent within 24h to someone worth knowing (builder, judge, or hiring-adjacent).
4. One talking point captured in DEMO.md → future Byline content + interview material.

If all four happen and you place last, the event paid for itself.

2. The scene: who runs what

Finding them: Luma (lu.ma — follow the "SF Hackathon Collection"), cerebralvalley.ai/events, agihouse.ai, events.ycombinator.com, Partiful for informal ones, Devpost for submission-style. Popular events waitlist fast and screen with a short application — your portfolio one-liners (jim-agent's x402 economy, grocery-buddy on Temporal) are exactly what those forms want.

"Loop engineering" / "harness" / "agent swarms" naming: real and current — the scene pivoted from "build an AI app" to "build/orchestrate autonomous agent systems." This is literally your thesis. You are walking into events themed on the thing you've spent two waves building.

3. Blunt expectations (first-timer, 2026 edition)

• Everyone ships now. AI codegen made "working demo" the floor — at recent events ~100% of teams demo something functional. You will not impress anyone by merely having it work. Differentiation moved to problem selection, product judgment, and demo clarity.
• Judging is 90 seconds to 3 minutes, demo-first, no slides. A confusing demo of something deep loses to a clear demo of something simple. Judges want to feel the problem in 10 seconds and see the money moment within a minute.
• Solo is normal and sometimes an advantage. "Solo or squad" is literal event copy; solo wins at Anthropic events are documented ($15k solo win in 8 hours). Small teams (1-2) win on coordination speed. Don't stress about finding a team — but the team-formation window (often ~10am-12:30) is itself a networking device; use it even if you build solo.
• Domain experts are beating engineers. A lawyer won an Anthropic hackathon; Opus 4.6 winners were domain experts, not senior engineers. The meta: bring a vertical you understand, don't out-code the room on a generic idea. (Your verticals: payments/trust, energy, civic, music.)
• The rooms at flagship tier skew pro. Expect founders, staff engineers, and repeat builders, not students. Smaller/vendor events are mixed.
• Recruiter/VC presence is concentrated, not universal. YC events (interview-as-prize), Afore's "Return of the Agents" ($500k–$2M pre-seed access), and model-provider events (product team as judges) are where hiring-adjacent people actually are. AI Tinkerers explicitly bans it. Pick events to match the goal of the week.
• Wrapper fatigue is real. Judges are openly tired of "ChatGPT wrapper + sponsor API call." The 2026 bar for "AI-native" is multi-step agentic behavior — tool use, verification, recovery. Your gate-catches-the-failure demos are precision-targeted at this fatigue.
• The cynical take exists. A circulating critique (European student event, HN-amplified) says AI hackathons reward vibe-coded front-end vaporware with nothing behind it. Not SF-verified, but the mood is industry-wide — which is exactly why a visibly real mechanism (a gate that deterministically blocks a bad action live) reads as refreshing.
• Operational reality: bring your own hotspot fallback, pre-pull models/deps, never depend on a live login or fresh API quota during the demo. Credits run out; wifi at 200-person events is wifi at 200-person events.

4. Event selection ladder

1. Calibrate (this month): one AI Tinkerers SF night or a Cursor/vendor build night. Zero pressure, learn the room's vocabulary, watch demos.
2. Rep (next 4-6 weeks): 2-3 mid-tier themed hackathons (agent/harness/MCP-themed Cerebral Valley or AGI House events are less oversubscribed than headline model launches). Ship from IDEAS.md.
3. Stretch (apply now, continuously): Anthropic/Cerebral Valley flagship series + YC events. Sub-4% acceptance means apply broadly and early; the application essay is your portfolio one-liner.

5. Gameplan

Before (T-7 → T-0)

• Read the event page rules: boilerplate/pre-existing-code policy (norms vary and are getting explicit — your INGREDIENTS pantry is normally fine if declared, but check), team cap, sponsor-prize criteria, judging format.
• Pick the idea from IDEAS.md that matches the event theme + duration. Have a backup one tier smaller.
• Run the INGREDIENTS pre-flight (env template, credits, cached deps, logged-in sessions).
• Prepare the 30-second self-intro (§6) — you'll say it 20 times.

During

• Hour 0: /hack <idea> <duration> — scaffold + T-minus schedule. Skeleton demo first, mocks everywhere, then make it real.
• Team-formation window: even going solo, work the room during it. Ask people what they're building and how — your goal is their insights, and "what harness are you using?" opens every conversation at these themed events.
• Mid-event: talk to sponsors' engineers (they're bored and technical), and to the people whose demos you'll want to ask about later.
• Demo freeze at the preset checkpoint. Non-negotiable. Rehearse twice. Pre-open every tab.
• Pitch order: problem they can feel (10s) → live money moment ≤60s in — show the gate catching a failure, not just success (90s) → the one distinctive mechanism (30s) → what you'd build next (the conversation hook).
• After demos: this is the highest-value networking window — everyone has a shared artifact to discuss. Ask the three best demos how they built theirs. Judges linger; ask them one specific question about your demo's weakest part.

After (within 24h)

• Follow up with every real conversation: one line referencing what you discussed + link to the repo/demo. LinkedIn or X, whichever they used.
• Capture talking points + lessons in the repo's DEMO.md; run /update-project if it should land on the portfolio site; consider a Byline-style post — "what I learned about X building Y in 10 hours" is exactly consumer-legible builder content.
• Log the event in this folder (what worked, who you met, what the room cared about) — the compounding asset is your read on the scene.

6. Your unfair advantages (say these out loud)

30-second intro: "I build agents where the model proposes and deterministic code disposes — the gate owns every irreversible action. I've shipped that sixteen ways: a research agent that pays for its own data over x402 and blocks any uncited claim, a pantry agent on Temporal that's never allowed to auto-buy, a procurement agent where the card issuer enforces policy so the model literally can't overspend. Now I'm working on making fleets trustworthy — verified orchestration."

• Events themed "loops/harnesses/swarms" are your home turf; most of the room discovered agent orchestration this year. You have ADRs, eval harnesses, and production hours.
• Seven demo-ready projects = seven talking points. Match project to listener: payments person → jim/x402/procurement; infra person → Temporal/grocery-buddy; product person → dj-agent/me-2.
• Your job-target map overlaps the judge/sponsor pool (Anthropic judges, Stripe-sponsored events, Temporal-adjacent infra crowds). Each event is a warm-intro channel Tier-S/1 can't give you through a portal.

7. OpenClaw / Hermes fluency (know the lingo, skip the dependency)

Verdict from 2026-07 research: neither is a gap in your stack; both are vocabulary you should own. Details:

• OpenClaw (né Clawdbot→Moltbot, Jan 2026): self-hosted personal-agent gateway (messaging front-ends, cron, memory, ClawHub skill registry). ~350k stars, 3M+ users, real SF hack-day circuit. It is a life-automation runtime, not a dev tool — it competes with your jim/grocery personal-agent layer, not with Claude Code. Security story is rough: CVEs, 135k+ exposed instances, ~12-20% of ClawHub skills found malicious in audits. Its default posture (broad allowlists, unvetted public skills) is the anti-pattern your gates exist to prevent — that contrast IS your talking point, not a reason to adopt it. Never install ClawHub skills live.
• Hermes in builder conversation = Hermes Agent, Nous Research's MIT-licensed agent framework (Feb 2026, ~40k stars) — self-improving skill loop, MCP-native, has its own Nous/NVIDIA/Stripe-sponsored hackathon. Distinct from the older Hermes model family (also Nous). Disambiguate on first mention; people conflate them.
• Job-market signal: a few postings list OpenClaw alongside Claude Code/Cursor as "know-the-ecosystem" bullets; none of your target companies require it. Anthropic's relationship with third-party harnesses is adversarial (the Apr–Jun 2026 subscription-billing whiplash) — worth knowing as an interview-current-events topic.
• Worth 30 minutes: read ClawRouter-Hermes (x402 USDC settlement wired into an agent runtime) — it overlaps your agentic-payments portfolio directly and is a better use of time than standing up either tool.

8. Watch-outs

• Application windows close fast; the Life Sciences hackathon deadline was literally same-day as this research. Check Luma/Cerebral Valley weekly, apply on sight.
• Team-size caps vary (some cap at 2) — confirm before promising to team up.
• Boilerplate rules vary per event — declare your pantry code; don't assume.
• Don't demo anything requiring live payments/money movement without a testnet/sandbox — your own gates should be the demo, not a compliance incident.