verify-references

name	verify-references
description	Use when adding research, citations, or factual claims to course materials — verifies all URLs work and all numeric claims have proper sourced references

Verify References

Overview

Audit course materials for broken URLs and unsourced factual claims. Every specific number, percentage, or data point must have a clickable source URL.

When to Use

• After adding or updating research content (docs/research/)
• After adding or updating class readings or slides
• After adding factual claims with specific numbers to any course file
• After editing readings/resources sections in schedule, slides, or handouts
• On-demand with /verify-references to check the full repo

Core Pattern

For each file in scope:
  1. Extract NEW URLs → WebFetch each → report broken ones. Save last date of checking.
  2. Extract ALL numeric claims (percentages, counts, multipliers)
     → check each has a [source](url) within 5 lines
     → flag unsourced claims
  3. For sourced claims: spot-check that the linked page
     actually contains the claimed data point

Quick Reference

Check	What to look for	Tool
Broken URL	HTTP 404, 403, redirects to wrong page	WebFetch
Unsourced claim	Number/% without a nearby [source](url) or **Source:** block	Grep + Read
Mismatched claim	Source page doesn't contain the stated number	WebFetch with targeted prompt
Stale reference	"2024 report" linked to page now showing 2025 data	WebFetch

Scope

Default file patterns to check:

docs/research/**/*.md
course/**/*.md
slides/**/index.md (readings sections only)

Output Format

Report using this structure:

## Reference Verification Report

### Broken URLs
| File:Line | URL | Error |
|-----------|-----|-------|

### Unsourced Claims
| File:Line | Claim | Attributed to | Fix needed |
|-----------|-------|---------------|------------|

### Mismatched Claims
| File:Line | Claim in file | What source says |
|-----------|--------------|-----------------|

### Summary
- X URLs checked, Y broken
- Z numeric claims found, W unsourced

Common Mistakes

• Checking only new content — older sections accumulate link rot. Check the full file.
• Skipping non-Anthropic URLs — external references (Veracode, OWASP, HumanLayer) break more often than first-party links.
• Accepting named attribution as "sourced" — "Veracode 2025 found 45%" is NOT properly sourced without a clickable URL. The standard is: reader can click and verify.
• Not spot-checking content — a working URL that doesn't contain the claimed number is worse than a broken URL (implies fabrication).