How to use the Judgment Day skill for dual AI code review

What is an agent skill and what does Judgment Day do?

An agent skill is a folder containing a SKILL.md file and optional reference files. When you ask for a matching task, the agent loads the skill and follows its instructions instead of improvising.

Judgment Day only activates when you explicitly ask for a Judgment Day or an equivalent dual or adversarial review of a concrete target. It builds a fixed snapshot of that target, then has two reviewers, judge A and judge B, examine it in parallel with the same criteria: correctness, edge cases, error handling, performance, security, and project conventions.

• Both reviewers are blind and read-only, so they do not edit files or see each other's output.
• The parent agent merges their findings into a shared ledger and persists it to an artifact store.
• The agent asks before fixing anything. Only issues that both reviewers confirm as severe are corrected.
• Fixes are re-checked by a scoped re-judgment, with at most two fix-and-recheck rounds.
• The run ends with exactly one verdict: APPROVED or ESCALATED.

How do you install the judgment-day skill?

Download the ZIP of the judgment-day skill from this page and unzip it. Then place the folder where your agent looks for skills. For a personal setup, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal: unzip into ~/.claude/skills/
• Project only: unzip into .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload

What does the skill need to run?

Judgment Day depends on a few things in your environment. Check these before you rely on it.

It needs the host runtime to be able to launch sub-agents or parallel judges. It also expects project skills named jd-judge-a, jd-judge-b and jd-fix-agent to be resolvable, and an artifact store to save the findings ledger, which the caller selects.

• A sub-agent or parallel judge launching capability in the host runtime
• Project skills named jd-judge-a, jd-judge-b and jd-fix-agent that can be resolved
• An artifact store to persist the findings ledger, chosen by the caller
• The skill runs no scripts and uses no network access

How do you trigger Judgment Day with a natural request?

Name the target and ask for a Judgment Day or dual review in plain language. The skill's trigger phrases include judgment day, dual review, adversarial review, and juzgar.

If your request does not name a concrete target, the agent will ask one scope question and stop. Vague requests like reviewing the project will not start the process.

• Run a judgment day on the changes in src/payments/refund.ts before I rely on them.
• Do a dual review of auth/session.py, adversarial style.
• Juzgar el módulo de facturación en billing/invoice.go.
• Judgment day on the diff for this branch against main.

What happens during a run?

The agent first builds the immutable target and freezes its scope. Both judges then review it in parallel and each returns one findings result before terminating. The agent waits for both, because a partial judgment is never accepted.

Where judges disagree, the run escalates for a human decision. Where only one judge reports an issue, it is recorded as suspect and not auto-fixed. Only severe issues confirmed by both judges go to the fix step, and the agent asks you before the first round of corrections.

Re-judgment sees only the frozen ledger and the fix changes. If issues remain after round two, the run escalates and stops.

• Both judges confirmed a severe finding: ask before round-one correction, then fix.
• Only one judge reports it: record as suspect, do not fix.
• Judges contradict each other: escalate for a human decision.
• Issues remain after round two: escalate and stop.
• The output includes target identity, round, counts of confirmed, suspect, contradiction and INFO items, artifact references, and one final verdict.

What does the skill not do?

Judgment Day issues no receipt and gives no permission to commit, push, open pull requests, or release. Its verdict does not satisfy any delivery gate, and delivery stays under your normal repository policy.

It is also heavier than ordinary review because it launches multiple sub-agents. The skill says not to run it together with the standard review method for the same target. If you want ordinary negotiated review, run it as a separate step.

• Not for routine everyday code review.
• Not a substitute for approval gates, commit rules or release checks.
• Do not combine it with ordinary review on the same target.

Safety notes and tips

The skill runs no scripts and makes no network calls, according to its metadata. Its reviewers are read-only, and fixes are only applied after you confirm. Because the process spawns several sub-agents, check your runtime's usage and cost expectations before a large target.

For auditability, make sure your artifact store is set up so the findings ledger is saved, and keep the target scoped to a specific change or file.

• Give a narrow target such as one file or one diff.
• Confirm the artifact store is configured so you can review the saved ledger.
• Read the ledger before accepting a fix round.
• Keep judgment-day separate from your normal review for the same target.