Work Verification Skill

Critical Requirement: This skill forces the AI agent to rigorously verify every claim, statement, and output. No assumptions. No skipping. Everything must be double-checked with concrete evidence.

Applies to: Code, writing, analysis, research, documentation, data interpretation, summaries, reports, recommendations, and all other work products.

Usage Modes

Important: When the user invokes this skill, check what they want verified:

• •If they mention file paths → verify those files
• •If they mention specific claims → verify those claims
• •If they say just /verify → verify recent session work
• •If unclear → ask them what they want verified

Mode 1: Verify Recent Session Work

When invoked without arguments (/verify), verify all work produced in the current session.

Example: /verify

Mode 2: Verify Specific Files

When the user provides file path(s), verify all claims in those specific files:

• •/verify path/to/report.md
• •/verify analysis.txt data-summary.md
• •/verify reports/*.md
• •User message: "Can you verify the Q4 report?" → Read the file and verify it

This mode is useful for:

• •Fact-checking existing documents
• •Verifying reports written outside this session
• •Auditing analysis done by others
• •Checking legacy documentation
• •Reviewing work done by colleagues

Mode 3: Verify Specific Claims

When the user provides specific claims to check:

• •/verify "User engagement increased 25%"
• •/verify the statistics in section 3
• •User message: "Check if the revenue numbers are accurate" → Find and verify revenue numbers

Verification Protocol

You MUST follow this protocol:

1. Identify All Claims

Go through everything you produced and list every factual claim or assertion you made. Include:

For Code Work:

• •Statements about what code does
• •Claims about file locations or contents
• •Assertions about how systems work
• •Descriptions of changes made
• •Explanations of behavior
• •Documentation of features

For Analytical & Writing Work:

• •Statistical claims or data points
• •Quotes or paraphrases from sources
• •Numerical figures or calculations
• •Date/time references
• •Citations or attributions
• •Causal relationships ("X causes Y")
• •Comparisons ("A is larger than B")
• •Historical facts or events
• •Definitions or explanations
• •Summary statements
• •Conclusions drawn from data

2. Verify Each Claim

For EVERY claim identified above:

1. •

   State the claim explicitly

2. •

   Determine verification method: What evidence would prove or disprove this?

3. •

   Gather evidence: Use tools to get concrete proof

   For Code/Technical Work:

   • •Read actual files (don't rely on memory)
   • •Run actual commands (don't assume output)
   • •Check actual git history (don't guess)
   • •Search actual code (don't presume)

   For Analytical/Writing Work:

   • •Re-read source documents/files
   • •Re-calculate numbers and statistics
   • •Check original sources for quotes
   • •Verify dates and times mentioned
   • •Cross-reference claims across multiple sources
   • •Search for contradicting information
   • •Validate logical inferences
   • •Check data sources and methodology
4. •

   Compare: Does the evidence match the claim?

5. •

   Document result:

   • •✅ VERIFIED: Evidence confirms claim
   • •❌ INCORRECT: Evidence contradicts claim
   • •⚠️ PARTIAL: Evidence partially supports claim
   • •❓ UNVERIFIABLE: Cannot gather evidence (state why)

3. Output Verification Table

At the end of your verification, provide a table with this format:

4. Corrections

If any claims were incorrect or partially correct:

1. •Acknowledge the error: Be explicit about what was wrong
2. •Provide correct information: Based on actual evidence
3. •Explain the discrepancy: If known, why was the original claim wrong?
4. •Update any affected work: Fix documentation, code, or explanations

Verification Examples

Code Work Examples

Example 1: Code Change Verification

Claim: "I added error handling to the process_data() function"

Verification:

1. •Read the file containing process_data()
2. •Search for try/catch or error handling patterns
3. •Compare with git diff to see what was actually added

Example 2: File Location Verification

Claim: "The configuration is stored in config/settings.json"

Verification:

1. •Use Glob to find all config files
2. •Read the file at the claimed path
3. •Verify it contains configuration data

Example 3: Behavior Explanation Verification

Claim: "The system retries failed requests 3 times"

Verification:

1. •Find the retry logic in code
2. •Read the actual implementation
3. •Check for the retry count constant or configuration

Analytical & Writing Work Examples

Example 4: Statistical Claim Verification

Claim: "The report shows a 25% increase in user engagement"

Verification:

1. •Re-read the source report/data file
2. •Locate the actual engagement metrics
3. •Recalculate the percentage change
4. •Verify the baseline and comparison periods

Example 5: Quote Verification

Claim: "According to the document, 'performance improved significantly'"

Verification:

1. •Read the source document
2. •Search for the exact quote
3. •Verify surrounding context
4. •Check if paraphrasing is accurate

Example 6: Date/Historical Fact Verification

Claim: "The policy was implemented in Q3 2024"

Verification:

1. •Read meeting notes or policy documents
2. •Search for implementation dates
3. •Verify against timeline documents
4. •Cross-reference with related events

Example 7: Calculation Verification

Claim: "The total cost is $15,420 across 3 departments"

Verification:

1. •Re-read the source data
2. •Manually recalculate the sum
3. •Verify all departments are included
4. •Check for any rounding or currency conversions

Example 8: Summary Verification

Claim: "The main findings are X, Y, and Z"

Verification:

1. •Re-read the entire source material
2. •Identify all key findings mentioned
3. •Verify no major findings were omitted
4. •Check if X, Y, Z accurately reflect priorities

Example 9: Causal Relationship Verification

Claim: "The new feature caused user retention to increase"

Verification:

1. •Read the analysis methodology
2. •Check if other factors were controlled for
3. •Verify timeline (correlation vs causation)
4. •Look for conflicting evidence or alternative explanations

Important Rules

1. •Never skip verification: Every single claim must be verified
2. •Don't rely on memory: Always read/check actual sources
3. •Document uncertainty: If you can't verify, say so explicitly
4. •Be honest about errors: If you were wrong, admit it clearly
5. •Verify verifications: Double-check your verification methods are sound
6. •Update as you verify: If you find errors, fix them immediately

Common Verification Pitfalls

General:

• •❌ "I'm pretty sure this is correct" → ✅ Verify with evidence
• •❌ "This should be right" → ✅ Check the actual source
• •❌ "I recall that..." → ✅ Re-read the original material
• •❌ "Approximately..." (when exact figures exist) → ✅ Get exact numbers

Code Work:

• •❌ "The file is probably in X directory" → ✅ Use Glob/Grep to confirm
• •❌ "Based on common patterns..." → ✅ Check this specific codebase
• •❌ "I added X feature" → ✅ Read the actual diff to confirm
• •❌ "This should work" → ✅ Test or read actual implementation

Analytical Work:

• •❌ "The document says something like..." → ✅ Find and quote exact text
• •❌ "Around 50 people..." → ✅ Get the exact number from source
• •❌ "This probably correlates with..." → ✅ Check if data actually shows correlation
• •❌ "The main point is..." → ✅ Re-read to verify it's actually the main point
• •❌ "According to studies..." → ✅ Cite the specific study/source
• •❌ "The data shows..." → ✅ Verify what the data actually shows (not interpretation)

When to Use This Skill

Use this skill:

• •After making significant code changes
• •After writing documentation or explanations
• •After completing analytical work or research
• •After writing reports, summaries, or recommendations
• •After processing data or creating visualizations
• •After making claims based on sources
• •When the user asks you to verify your work
• •Before marking tasks as complete
• •When accuracy is critical
• •Before presenting findings or conclusions
• •After any work where facts, figures, or claims were stated

Workflow

For Session Work (Mode 1)

1. •Review your recent work (messages, code changes, files written, analysis, reports)
2. •Extract all factual claims (make a comprehensive list)
3. •Verify each claim systematically (use tools, gather evidence)
4. •Create the verification table (show all results transparently)
5. •Make corrections (fix any errors found)
6. •Summarize findings (how many verified, how many errors, what was fixed)

For Specific Files (Mode 2)

1. •Read the specified file(s) completely
2. •Extract all factual claims from the content (comprehensive list)
3. •For each claim, gather evidence from:
   • •Other files referenced in the document
   • •Source data mentioned
   • •Original sources for quotes
   • •Re-calculations for numbers
4. •Create the verification table showing what was checked
5. •Report findings: Which claims are accurate, which need correction
6. •Do NOT modify files unless explicitly asked - just report findings

For Specific Claims (Mode 3)

1. •Identify the specific claims the user wants verified
2. •Locate these claims in files or recent work
3. •Verify each claim with concrete evidence
4. •Create focused verification table for just those claims
5. •Report findings with details

Types of Work This Applies To

• •Code & Development: Changes, features, bug fixes, architecture
• •Documentation: Technical docs, user guides, README files
• •Analysis: Data analysis, performance analysis, competitive analysis
• •Research: Literature review, fact-checking, source verification
• •Writing: Reports, articles, summaries, presentations
• •Data Processing: Calculations, aggregations, transformations
• •Recommendations: Decisions based on data or analysis
• •Communication: Emails, messages with factual claims

Practical Usage Examples

Example: Verify an Existing Report

User: /verify reports/Q4-analysis.md

Agent should:

1. •Read reports/Q4-analysis.md completely
2. •Extract all claims (statistics, dates, conclusions)
3. •Find source data files referenced
4. •Re-calculate any numbers mentioned
5. •Verify all quotes and attributions
6. •Create verification table
7. •Report: "Found 15 claims, 13 verified, 2 incorrect"

Example: Verify Specific Statistics

User: /verify the revenue numbers in financial-summary.txt

Agent should:

1. •Read financial-summary.txt
2. •Extract all revenue-related numbers
3. •Locate source data (spreadsheets, databases)
4. •Recalculate each figure
5. •Create verification table for revenue claims only
6. •Report findings

Example: Verify Multiple Documents

User: /verify docs/*.md

Agent should:

1. •Read all markdown files in docs/
2. •Extract claims from each file
3. •Verify systematically (may take multiple tool calls)
4. •Create comprehensive verification table
5. •Report findings per file

Example: Check Someone Else's Analysis

User: /verify analysis/competitor-research.md (written by colleague)

Agent should:

1. •Read the file without assuming it's correct
2. •Treat every statement as needing verification
3. •Find original sources mentioned
4. •Re-verify data and calculations
5. •Report findings objectively
6. •Do NOT modify the file unless asked