HKPUG Opik Challenge Participant guide

Improve one prompt. Learn from every run.

Build a stronger support prompt across a focused 14-day challenge. Submit an encrypted prompt, receive a score and private submission feedback, then use Opik to decide what to improve next.

Start the mini workshop Make your first submission
Starts
Ends
Attempt limit
Eight total
Daily pace
Two maximum

Challenge background

You are tuning HarbourCloud's support copilot before it goes live.

HarbourCloud is a fictional multi-region SaaS company. Its support team handles policy-heavy customer questions where the correct response depends on current records, source authority, and whether support is allowed to act. Your team owns the reusable prompt that guides the copilot through those decisions.

  1. A ticket arrives

    A customer asks about an account, payment, policy, incident, or product operation.

  2. Evidence is supplied

    The shared handbook and a domain evidence pack provide the facts available for that case.

  3. Your prompt guides the model

    DeepSeek V4 Flash receives the question, evidence, and the system prompt submitted by your team.

  4. The run is evaluated

    Deterministic checks and Qwen 3.7 Plus evaluate the answer before the score is published.

THE QUEUE

Fifty cases across ten support domains

One prompt must work across the whole queue. A case may combine verified records with stale guidance, conflicting states, pasted customer claims, or missing approval. The goal is not to memorize answers. It is to teach the model how to choose evidence and make a safe support decision.

  • Refunds
  • Subscriptions
  • Account access
  • API limits
  • Billing
  • Data retention
  • Integrations
  • Privacy
  • Security
  • Service incidents
What you are improving

A single system prompt that consistently produces grounded, relevant, well-structured answers. After each scored run, use the returned Opik traces to find where the model trusted the wrong evidence, missed a conclusion, cited poorly, or escalated incorrectly.

The same evaluation settings apply to every team

Runs use fixed models, a fixed seed, and the same seven criteria. Only your submitted system prompt changes between attempts.

Evaluation bank
40 discovery cases and 10 hidden holdout cases
Answer model
DeepSeek V4 Flash with temperature 0 and seed 0
Semantic judge
Qwen 3.7 Plus using fixed score tiers
Score composition
40 deterministic points and 60 judge points

Your private bundle contains all 40 discovery traces. Holdout cases remain aggregate-only. The official score combines 75% discovery and 25% holdout performance.

Your challenge loop

Each route covers one participant task. Start with the helper, keep the submission small, and bring every submission feedback bundle back to your own machine.

Learn Opik first

Load the six-case mini workshop, follow its span trees and Experiments, and check your findings against the answers.

Start the mini workshop

Track your progress

Compare the official standings and use your remaining attempts where they can make the most difference.

Open the leaderboard

Keep the loop deliberate

The helper handles packaging and protection so your attention stays on prompt quality and trace evidence.

Official score
75% discovery and 25% holdout
Private by default
Your prompt and returned submission feedback remain encrypted in transit.
One file in each pull request
Submit only submission/submission.zip.