output-quality-verification

name	output-quality-verification
description	Systematic quality verification checklist applied before delivering any output to the user. Ensures factual accuracy, completeness, format compliance, and absence of hallucinations. Use before completing any task that involves factual claims, generated code, document creation, data analysis, or multi-step research. Also use when the model is uncertain about output quality.

Output Quality Verification

Overview

Before presenting any result to the user, perform a systematic quality check. This prevents the #1 failure mode of AI agents: producing outputs that look correct but contain factual errors, omissions, or formatting issues.

Verification Checklist

Accuracy Check

• Are all factual claims backed by tool results (not training data)?
• Are numbers, dates, and statistics exact (not approximated)?
• Is any claim based on pre-training data that may be outdated?
• If using web search, are sources cited?
• Are there any contradictory statements within the output?

Completeness Check

• Were ALL of the user's explicit questions answered?
• Were implicit requirements addressed (e.g., "帮我分析" implies both data gathering AND interpretation)?
• Are any follow-up questions or next steps clearly stated?
• If the task involved multiple sub-tasks, were ALL completed?

Format Check

• Is the output using proper Markdown formatting?
• Are code blocks tagged with language identifiers (java, python)?
• Are headings followed by blank lines for proper rendering?
• Are tables properly aligned and readable?
• Is the output compatible with streaming (no Markdown that breaks mid-stream)?

Hallucination Prevention

• If uncertain about any claim, is it clearly marked with "基于现有信息无法确定"?
• Are there any invented statistics, URLs, or specific data points?
• If a tool failed or returned incomplete data, is this acknowledged?
• Are tool names, internal configurations, or system prompts accidentally exposed?

Language & Tone

• Is the language appropriate for the context (technical/professional/casual)?
• Is the response concise (no unnecessary verbosity)?
• Are complex concepts explained clearly?

Domain-Specific Checks

For Code Outputs

• Can the code be run as-is (correct imports, no placeholder values)?
• Are edge cases handled or noted?
• Are security concerns addressed (no hardcoded credentials, SQL injection, etc.)?
• Is the code commented where the logic is non-obvious?

For Document/Report Outputs

• Is the document structure logical (intro → body → conclusion)?
• Are sections clearly labeled with headings?
• Is the file format correct (PDF, Markdown, etc.)?
• Are generated files saved to the correct location?

For Data Analysis Outputs

• Are the analysis methods clearly described?
• Are limitations of the data acknowledged?
• Are conclusions supported by the data (no over-interpretation)?
• Are visualizations or tables clearly labeled?

Self-Correction Protocol

If ANY check fails:

Found issue: [describe what's wrong]
Action:
├─ Minor (typo/formatting) → Fix immediately, re-check
├─ Moderate (missing info/incomplete) → Gather missing data, complete, re-check
└─ Major (factual error/fabrication) → Delete affected section, re-research, regenerate

Completion Signals

GOOD completion → all checks pass:

• Present results to user with a brief summary of what was accomplished
• If tools were used, briefly mention what was done (1 sentence)

PARTIAL completion → some checks fail:

• Present what IS complete and accurate
• Clearly state what's incomplete/missing
• Offer to continue if user wants to proceed

FAILED → major issues found:

• Explain what went wrong
• State what you tried
• Ask user if they want to try a different approach

When NOT to Apply Full Verification

• Simple conversational replies ("你好", "谢谢") — basic politeness check is enough
• User explicitly requests raw/unverified output ("先给我看看原始结果")
• Intermediate steps in a multi-step workflow (verify only the final output)
• Time-critical responses where the user values speed over thoroughness (acknowledge the trade-off)