multi-model-meta-analysis_skill

Overview

This skill synthesizes outputs from multiple AI models into a single, verified assessment by cross-referencing their claims against the actual codebase. It resolves contradictions, confirms or refutes factual statements with evidence, and produces a reliable consolidated document that is more trustworthy than any individual model output. The result highlights verified findings, unresolved items, and actionable recommendations.

How this skill works

The skill parses each model's output to extract discrete claims, recommendations, and reported issues, tagging them by source. It deduplicates semantically equivalent statements, then verifies factual claims and bug/security assertions by reading relevant files and quoting code as evidence. Conflicts are resolved by inspecting source lines and documenting which model (if any) was correct. Finally, it synthesizes a structured assessment with confirmed findings, refuted claims, unverifiable items, and prioritized action items.

When to use it

• You have feedback from multiple LLMs about code, architecture, or design and need a single authoritative assessment.
• Models disagree about behavior, dependencies, or vulnerabilities and you must resolve contradictions against source code.
• You want verified bug reports and security findings before opening issues or creating a remediation plan.
• You need a consolidated list of recommendations with evidence and clear prioritization.
• You want to preserve helpful model suggestions that cannot be fully verified but are worth considering.

Best practices

• Always attach the model source to each claim so provenance is clear.
• Prioritize verifying bug/security claims and statements about code behavior.
• Use simple grep/glob/read operations to quote exact lines as evidence for each verification.
• Deduplicate similar claims into canonical phrasing and list all supporting models.
• Mark runtime, performance, and external-API claims as unverifiable without tests or logs.

Example use cases

• Consolidate code review comments from GPT, Claude, and Gemini into one evidence-backed report.
• Resolve conflicting vulnerability reports from multiple models by inspecting the implementation files.
• Produce prioritized action items for a repo after synthesizing varied model suggestions.
• Generate a short verified summary for stakeholders that cites exact files/lines for major issues.