adversarial-review

Adversarial Documentation Review

Catch what manual review misses: undocumented features, unverified claims, misleading examples.

When to Use This Skill

• README/docs review: "Is this README accurate and complete?"
• API documentation audit: "Are all methods documented?"
• Performance claims: "Are these benchmarks real?"
• Migration guides: "Does this show all breaking changes?"

Red flags that trigger adversarial review:

• Quantitative claims without evidence ("5x faster")
• Selective examples (shows 1 of 8 subcommands)
• Version-dependent claims (could be outdated)

The Adversarial Difference

Manual Review	Adversarial Review
"Looks comprehensive"	82.5% coverage (33/40 methods)
"Examples seem right"	7 methods undocumented (proven via derivation)
"5x claim noted"	Unverified (no benchmark files found)
Miss check method gap	Formal proof 4 variants missing

Instructions

Phase 1: Exhaustive Enumeration (Ground Truth)

Build complete inventories from source code, not docs.

For CLI Tools:

# Extract ALL commands and subcommands
uv run tool --help | grep "Commands:" -A 100
uv run tool add --help | awk '/Commands:/,/^$/' 
uv run tool check --help | awk '/Commands:/,/^$/'

# Count total
echo "add subcommands: 8 total"
echo "check subcommands: 7 total"

For Python APIs:

# Extract all public methods from source
import ast

with open('api/table.py') as f:
    tree = ast.parse(f.read())

for node in ast.walk(tree):
    if isinstance(node, ast.ClassDef) and node.name == 'Table':
        methods = [m.name for m in node.body 
                  if isinstance(m, ast.FunctionDef) 
                  and not m.name.startswith('_')]
        print(f"Total public methods: {len(methods)}")
        for m in sorted(methods):
            print(f"  {m}")

Critical: This is your ground truth. Docs must be measured against THIS, not assumed complete.