Skip to content

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Independent candidates moving through anonymization and blind judging to a winner

Tournament Judge

Agent Skills License: MIT Markdown

A domain-neutral Agent Skill that creates or evaluates independent candidates, pressure-tests each one, and uses blind rubric-based judging to select the strongest option.

Use it for

  • architecture and implementation proposals
  • product, UX, and process designs
  • research plans and written deliverables
  • tool, library, service, or vendor comparisons
  • migration, operations, and project strategies
  • any substantial best-of-N decision with comparable artifacts

Skip it for factual lookups, trivial preferences, or decisions where the evidence is too incomplete to judge responsibly.

At a glance

1. Prepare 2. Compete 3. Decide
Freeze one brief, rubric, and evidence policy Build or normalize isolated candidates, then review each equally Remove team identities, score against the frozen rubric, and publish the audit trail

Quick start

git clone https://github.com/alextverdyy/tournament-judge.git tournament-judge
mkdir -p .claude/skills
cp -R tournament-judge .claude/skills/tournament-judge

This example uses Claude Code. Codex, OpenCode, Cursor, Prime Agent, and global install paths are listed in the setup guide.

Then ask:

Use tournament-judge to produce three independent approaches for this task. Freeze a rubric first, review each approach twice, judge anonymized finalists, and return the winner with the full scorecard.

Your agent may auto-load the skill from its description. You can also explicitly ask it to read tournament-judge/SKILL.md.

What happens

flowchart LR
    accTitle: Tournament workflow
    accDescr: A frozen brief feeds three isolated candidates. Each receives adversarial review before anonymization, blind judging, and a final report.
    B["Frozen brief<br/>and rubric"] --> A[Candidate A]
    B --> C[Candidate B]
    B --> D[Candidate C]
    A --> R1[Adversarial review]
    C --> R2[Adversarial review]
    D --> R3[Adversarial review]
    R1 --> X[Anonymize]
    R2 --> X
    R3 --> X
    X --> J[Blind judge]
    J --> W[Winner and report]

    classDef brief fill:#172554,stroke:#38bdf8,color:#fff;
    classDef candidate fill:#312e81,stroke:#a78bfa,color:#fff;
    classDef review fill:#064e3b,stroke:#34d399,color:#fff;
    classDef judge fill:#0f172a,stroke:#f472b6,color:#fff;
    classDef winner fill:#78350f,stroke:#fbbf24,color:#fff;
    class B brief;
    class A,C,D candidate;
    class R1,R2,R3 review;
    class X,J judge;
    class W winner;
Loading

Repository map

Path Purpose
SKILL.md Skill entry point and invariant rules
modules/ Intake, rubric, orchestration, judging, and reporting procedures
prompts/ Candidate, reviewer, and judge role prompts
templates/ Brief, candidate dossier, rubric, and final report forms
docs/ Setup, usage, modules, and companion-skill guides
examples/ Ready-to-copy tournament requests
assets/ Accessible SVG identity and workflow illustrations

Documentation

Design guarantees

  • Rubrics are fixed before candidate generation.
  • Candidates receive equal inputs and review budgets.
  • Teams remain isolated before judging.
  • Judges see anonymized finalists, not debate performance.
  • Hard-constraint failures remain visible.
  • Scores include reasons and evidence gaps.
  • A final report explains why the winner won and why the others lost.

These guarantees are procedural, not magical. A single-context runtime cannot provide perfect independence, and anonymization cannot remove authorship clues intrinsic to an artifact. The report must disclose those limitations.

License

MIT