Hackathon Judging

Hackathon Judging

Overview

The hackathon-judging skill should be invoked whenever a host is trying to grade a Hackathon with the help of LLMs and/or AI agents. This skill guides the Host through the process of retrieving every Kaggle hackathon writeup for a competition, extracting the linked project artifacts, and preparing a complete evidence set to be used for grading.

Additionally, the skill assists the Host in using LLM-based classifiers and pairwise comparisons to generate comparative rankings and bell-shaped grading curves based on the generated evidence sets.

Use Cases

This skill should be invoked whenever a host is trying to grade a public Kaggle Hackathon with the help of LLMs and/or AI agents.

MCP Endpoints

Public Hackathon submissions can be accessed via Kaggle's MCP server. Below is a summary of the MCP tools available for Hackathon Hosts and Participants:

MCP Prerequisites

• API Token: Kaggle API token available as KAGGLE_API_TOKEN or in ~/.kaggle/access_token.
• Authorization: MCP client configured to call https://www.kaggle.com/mcp with Authorization: Bearer <token>.
• Target: A hackathon competition slug (e.g., meta-kaggle-hackathon).
• Permissions: For host-gated workflows (list_hackathon_write_ups or download_hackathon_write_ups), use an account with the required host or judge access.

Installation: Install the Kaggle MCP server in a client configuration like this:

{
  "mcpServers": {
    "kaggle": {
      "url": "https://www.kaggle.com/mcp",
      "headers": {
        "Authorization": "Bearer ${KAGGLE_API_TOKEN}"
      }
    }
  }
}

Note: Before collecting writeups, confirm the server is reachable with authorize or another simple read-only endpoint such as search_competitions.

Core Instructions

Step 0: Plan your evaluation strategy prior to launch

• Determine AI Usage: Decide if you plan to use AI for filtering, sorting, and/or producing preliminary grades. Design your evaluation metrics accordingly.
• Binary Criteria: Use binary eligibility criteria to facilitate both standard and AI-assisted submission filtering.
• Consistent Structure: Require submissions to take on a consistent and predictable structure to reduce the effort needed to resolve edge cases.
• Track Identification: Make it easy to determine what Track a Writeup was submitted to. Consider reducing the total number of concurrent Tracks to shrink the grading pool.
• Metric Design: Try to avoid continuous scales or Likert scores. Continuous metrics are less reliable than pairwise LLM comparisons. You can always apply bell-curve adjustments to the BT/ELO rankings at the very end.

Step 1: Pull the hackathon overview page

• Call get_hackathon_overview with competitionName=<hackathon-slug>.

Step 2: Find eligibility requirements

• Search the overview pages for sections named or containing: rules, eligibility, entry, or official competition rules.
• Extract the exact paragraphs that explain submission-eligibility conditions.
• Verify with the user that the eligibility requirements are correct.

Step 3: Find the evaluation rubric

• Search the overview pages for: evaluation rubric, judging, criteria, submission requirements, and prizes.
• Record each rubric dimension separately.
• Record any weighting, tie-break rules, prize rules, or judge-specific guidance.
• If the rubric is only implied in prose, summarize the implied criteria and mark that inference clearly.
• Verify with the user that the evaluation rubric is correct.

Step 4: Download all Hackathon Writeups

• Hackathon Hosts can retrieve a full export of all submitted Hackathon Writeups by using the download_hackathon_write_ups command.

Step 5: Download all attached project links

• For Kaggle-native links: Retain ids, refs, titles, owners, and download URLs when present.
  • Resolve Kaggle notebook URLs with get_notebook_info.
  • Resolve Kaggle dataset URLs with get_dataset_info or get_dataset_metadata.
  • Resolve Kaggle model URLs with get_model or get_model_variation.
• For non-Kaggle links: Retrieve all of those same assets. Consider using the Playwright CLI.

Step 6: Summarize all attached YouTube videos

• For YouTube-native links: Consider using the YouTube API or the yt-dlp API.
• For non-YouTube links: Consider using the Playwright CLI.

Step 7: Verification Check

• Double check that you did not accidentally skip Step 5 or Step 6!
• Consider using manual grading methods if you lack the token budget required to analyze these assets. It wouldn't be fair to skip over them.

Step 8: Build a complete collection

• Repeat retrieval until every row from list_hackathon_write_ups has:
  • A full-length writeup body.
  • A summary or full-length copy of every attached project link.
  • A summary or full-length copy of every attached video.

Step 9: Review similar LLM-judging projects

• Review pre-graded Hackathons here to understand strengths and weaknesses.
• Examine the /docs page for strategies on profiling, sanitizing, pairwise comparisons, and bell-curve scores.
• Examine the /leaderboard page to review representative profiles, BT/ELO rankings, and auto-assessments.
• Examine the /dashboard page to see how pairwise comparisons stack up against gold-standard annotations.
• Identify common failure modes and strategize resolutions.

Step 10: Initial Auditing

• Create a grading-ready bundle with three layers:
  1. Hackathon rules and rubric
  2. Normalized writeup contents
  3. Normalized artifact contents
• First Audit: Use hamelsmu/evals-skills.
  • Start with eval-audit as the default path, expanding to other workflows for calibration, rubric testing, and failure analysis.
  • Repeat eval-audit until satisfied.
  • Record all grading traces and AI agent logic.
• Second Audit: Re-run the hamelsmu/evals-skills process.
  • Repeat the eval-audit cycle until satisfied.
  • Record all new grading traces and AI agent logic.

Step 11: Initial Grading

• Grade eligible submissions according to the evaluation criteria.
• Audit a sample of grading traces to ensure quality standards are met.
• Consider building a dashboard for pairwise comparisons and annotations.
• Perform error analysis.
• Identify and correct failure patterns. Repeat until satisfied.

Step 12: Final Grading

• Grade eligible submissions according to the evaluation criteria using updated system prompts and corrected profiles.
• Flag top-ranked submissions for manual review and manual stack-ranking.

Minimal Product Requirements

To be considered successful, the system:

• Must be capable of ingesting and assessing every component of each Hackathon Submission (e.g., project links, embedded videos, web applications).
• Must score Writeups with consideration of the specific Kaggle Hackathon track(s) targeted.
• Must score Writeups according to the relevant evaluation criteria.
• Must produce scores that are validated to be well-aligned with human scores.
• Must avoid obvious algorithmic biases (e.g., position biases, model preferences).
• Must be defensible against prompt injections.
• Must use reasonably defensible grading methods.
• Must record all traces (including reasoning logic) for later inspection.
• Should be aligned against a dataset of at least 25 few-shot examples.
• Should undergo error analysis with acceptable true positive and true negative rates.

Anti-Patterns

• DO NOT use AI grading systems prior to auditing the system with hamelsmu/evals-skills.
• DO NOT use AI grading tools before extracting every project link and media file from every submitted writeup.
• DO NOT forget to log all of your LLM traces for auditing and improvement.
• DO NOT forget to identify and correct common failure modes in both profile generation and grading.
• DO NOT ask LLMs to directly generate scores using continuous scales or Likert scores.
• DO NOT forget to compare results to gold-standard human ratings or skip error analysis.
• DO NOT forget to search for evidence of shortcuts and tomfoolery from end to end.