From Hackathon First Prize to a Codex Skill: My Real‑World Playbook

Winning the Hackathon While Sick

I participated in the 小宿科技环球黑客松 Beijing station, entered the "Beyond Prompt: Agents in Action" track solo, and surprisingly took first prize. The whole experience happened while I was battling a cold, surrounded by medicine bottles, tissues, and constant sniffles.

Why the Competition Felt Different

The event was unusually pragmatic. It wasn’t a "genius kid" story but an adult‑to‑adult industry exchange with real problems, real product judgments, and real engineering delivery under tight time constraints.

Distilling the Experience into a Skill

After the hackathon I reviewed my two‑day chat log with Codex and discovered a complete trajectory: how I chose a direction, mis‑judged, cut losses, got hit by an interview insight, migrated past experience, and finally converged on a deliverable. I distilled that entire process into a reusable skill (https://github.com/kangkona/kangkona-skills).

The Reality of the 48‑Hour Sprint

The story is not a pre‑planned, flawless execution. It was a 48‑hour swing of constant pivoting, trial‑and‑error, mis‑judgments, and a sudden insight that finally let the idea converge.

Early Direction Hunting

In the first half‑day I didn’t start coding. I studied the track, the scoring rubric, and brainstormed several possible directions: a financial‑focused agent, a research‑analysis agent, and an enterprise‑workflow agent. Each had merits—commercial appeal, demo potential, alignment with my prior thinking, or quick implementation—but also drawbacks such as being too demo‑like, over‑promising, lacking depth, or requiring heavy data integration.

To filter these ideas I kept asking myself:

Can I explain the concept in 30 seconds?

Does it produce a complete closed‑loop?

Is there hard evidence a judge can see?

Is it just an API wrapper?

Can it survive Q&A scrutiny?

Experimenting with New Tech Stacks

During the hackathon I tried several new stacks for the first time:

TanStack Starter, recommended by "Fox" for modern full‑stack apps.

Astro’s Island architecture to see if it truly speeds up front‑end scaffolding.

Agent harnesses such as Mastra and Flux to reduce wheel‑reinvention.

Cloudflare’s full suite—Workers, D1, KV, R2, Durable Objects—for serverless edge deployment.

While most stacks were easy to pick up, the gap between my imagined product and the actual implementation kept widening.

The Core Gap: Product Vision vs. Reality

Many ideas looked beautiful on paper—data, agents, automatic analysis, decision recommendations, workflow loops—but when I started building, each added role, panel, or process made the prototype feel more like a growing idea collection than a concise hackathon demo.

The key realization was that a new stack should shorten the path from idea to a trustworthy demo , not inflate the distance.

Hackathon Is Not a Tech‑Stack Expo

Technology and agents must serve the narrative, the closed‑loop, and the delivery, not dominate the project. The event is still a great learning ground because it forces rapid judgment on whether a tool truly helps you finish the work.

Insights from Fiona Fung’s Interview

Fiona Fung’s interview about AI‑native engineering teams reinforced a shift I’ve seen: code is no longer the bottleneck; verification and measurement are. This mirrors my ten‑year experience with AB testing at ByteDance, where AB testing is essentially loop engineering for massive organizations.

Re‑thinking the Coding Agent Track

Initially I feared the Coding Agent track would be crushed by Codex or Claude Code. I later realized the real opportunity lies after the coding agent—how teams manage, verify, and trust the many local agent loops that appear.

I coined the term TeamLoop to describe a system where each engineer’s local coding agent produces work that the team can observe, verify, and hand off.

From Idea to Playbook

To capture the lessons I built a Hackathon Playbook skill that guides participants through:

Reading the scoring criteria.

Selecting and evaluating project ideas.

Running multiple projects in parallel ("project racing").

Compressing the narrative for a 4‑minute pitch.

Building a demo with a credible closed‑loop.

Balancing technical depth with demo clarity.

Defending Q&A.

Installation command: npx skills add kangkona/kangkona-skills Usage example:

Use $hackathon 帮我选择黑客松项目、打磨叙事、规划执行，并准备 demo/PPT/Q&A。

Why Open‑Source the Skill

Open‑sourcing lets the reusable judgment structures—reading scoring tables, avoiding sunk‑cost fights, evaluating new stacks, distinguishing real agents from fixed workflows, answering in four minutes, and honestly stating what’s done—benefit the community and prevent repeated pitfalls.

Personal Reflection at 35

The hackathon also proved that being 35 as an internet engineer isn’t a death sentence. My experience—sick, juggling medicine, trying new stacks, and constantly iterating with an AI agent—showed that experience helps recognize chaos, know when to persist or stop, and keep learning.

As long as you’re willing to learn and build, you’re still in your prime.