judge-with-debate

Overview of judge-with-debate skill

The judge-with-debate skill is for evaluating a solution with structured, multi-agent disagreement instead of a single-pass opinion. It is best when you need a defensible judgment on quality, correctness, or tradeoffs and want the judge-with-debate skill to force evidence, counterarguments, and convergence before final scoring.

What judge-with-debate is for

Use judge-with-debate when the job is not “write an answer,” but “decide whether this answer, design, or implementation is actually good.” It is a strong fit for code review, solution ranking, rubric-based assessment, and any Multi-Agent Systems workflow where bias from one model pass would be risky.

Why it is different from a plain prompt

A generic evaluation prompt usually asks for one opinion. judge-with-debate adds a meta-judge, shared evaluation specification, and repeated debate rounds so the result is harder to hand-wave. That makes the judge-with-debate skill more useful when accuracy matters more than speed.

Best-fit readers

This skill is a good fit for agents, reviewers, and builders who need repeatable evaluation criteria, not just a verdict. If you are comparing multiple candidate solutions, or you need the judge-with-debate guide to produce consistent scoring across cases, this skill saves setup time and reduces guesswork.

How to Use judge-with-debate skill

Install and inspect the skill first

Use the repository install flow in your skill manager, then read the skill file before trying to apply it. A typical judge-with-debate install path is to locate plugins/sadd/skills/judge-with-debate/SKILL.md, then confirm the surrounding repo conventions so you know how this skill expects inputs and outputs to be organized.

Give it the right input shape

The skill works best when you provide a solution path or artifact plus explicit evaluation criteria. A strong judge-with-debate usage prompt says what is being judged, what “good” means, and what constraints matter. For example: Judge this PR against correctness, maintainability, and spec compliance; prioritize evidence from the diff and call out any missing edge cases.

Start with the files that define behavior

Read SKILL.md first, then look for nearby repo conventions that affect execution. In this repository, the main thing to inspect is the skill body itself; there are no helper scripts or extra reference folders, so the install decision depends on understanding the task flow, the debate phases, and the output expectations from the single source of truth.

Use it in a debate-friendly workflow

A practical judge-with-debate guide is: supply one target, one rubric, and any hard constraints up front; let the meta-judge shape the spec; then let the judges argue from evidence rather than rephrasing the same score. This skill is strongest when you preserve the distinction between “specification,” “analysis,” and “consensus,” because collapsing those steps reduces the value of the debate.

judge-with-debate skill FAQ

Is judge-with-debate only for code review?

No. The judge-with-debate skill is for any structured evaluation where multiple perspectives improve trust: code, prompts, plans, research summaries, or competing solutions. It becomes most valuable when the cost of a wrong judgment is higher than the cost of a longer evaluation.

When should I not use it?

Skip judge-with-debate when you need a quick heuristic answer, when the criteria are too vague to debate, or when there is no meaningful evidence to compare. If a simple rule-based check is enough, the debate overhead is unnecessary.

Is this better than a single strong prompt?

Usually yes for contested decisions, because the skill makes disagreement explicit and forces convergence around evidence. For simple tasks, though, a normal prompt may be faster and sufficiently accurate; the judge-with-debate skill is about decision quality, not minimum tokens.

Is it beginner-friendly?

Yes, if you can name the artifact and state the rubric. The main beginner mistake is giving a broad request like “judge this” without specifying what counts as success, which leaves the debate underpowered.

How to Improve judge-with-debate skill

Give tighter evaluation criteria

The biggest quality lever is the rubric. Instead of asking for a generic verdict, specify weighted concerns and failure thresholds: Score correctness 50%, robustness 30%, clarity 20%; fail if the solution misses an edge case or contradicts the spec. Stronger criteria help the judge-with-debate skill produce sharper disagreement and cleaner consensus.

Provide evidence-ready context

Debate works best when judges can point to concrete material: the exact solution path, relevant snippets, acceptance criteria, and known constraints. If you omit those inputs, the skill will still run, but the debate will drift toward inference instead of grounded assessment.

Watch for common failure modes

The main failure mode is overgeneralized consensus: all judges sounding aligned because the prompt was too broad. Another is rubric drift, where the discussion starts scoring different things. To improve judge-with-debate skill results, keep the target narrow, ask for explicit tradeoffs, and request a final summary that preserves any unresolved disagreement.

Iterate after the first pass

If the first output is too soft, feed back the missing decision point and rerun with a more specific rubric or stricter evidence requirements. For judge-with-debate for Multi-Agent Systems, the best improvements usually come from clarifying the decision boundary, not from asking for more rounds.