Writing Guarddog Rules Skill for Claude

Writing GuardDog Rules

GuardDog detects supply-chain malware with a two-layer model. The full reference,

including the philosophy, metadata schema, and worked examples, lives in

WRITING_RULES.md at the repository root (the skill directory is

.claude/skills/writing-guarddog-rules/, so the doc is three levels up). This

skill is the procedural layer: it captures the mental model and the workflow,

and it points to the reference for detail. Read WRITING_RULES.md when you need

the schema specifics, field definitions, or longer examples; do not duplicate it.

The model (memorize this, skip the doc for simple calls)

• Capability = "CAN DO" — a function call that enables an action

(requests.get(, .readFileSync(, subprocess.Popen(). Match calls, not imports.

• Threat = "SUSPICIOUS" — an attacker indicator (/etc/passwd, discord.com/api/webhooks,

base64.decode + exec together). Not a bare function call.

• Risk = capability + threat in the same file with matching category. A capability

alone is benign; a threat indicator alone is often a false positive; together they are a risk.

• identifies format: {type}.{category}[.{detail}] where type is capability or threat,

category is one of network, filesystem, process, runtime, system, metadata.

• Categories must match for a risk to form. capability.process.* only pairs with

threat.process.*. General detail matches specific (threat.network + capability.network.outbound),

but conflicting details do not (...outbound + ...inbound).

• threat.runtime.* and threat.metadata.* auto-form risks without needing a capability

(obfuscation, install hooks, typosquatting, maintainer compromise).

When unsure whether a pattern is a capability or threat: does it show what code *can do* (capability)

or a *suspicious indicator* (threat)? Does it stand alone without a capability (runtime/metadata threat)?

Authoring workflow

• Decide type and category. Pick capability vs threat, then category and optional detail.

Confirm the matching counterpart exists or will exist so a risk can form.

• Write patterns as YARA (.yar). Source-code rules are YARA-only and language-agnostic

(loaded for every ecosystem). Follow the best practices in WRITING_RULES.md (word boundaries \b,

match method calls not object names, require quote context for bare strings, establish context

with private rules before matching threats). Extract shared building blocks (LOLBAS, hooks) into

.meta files instead of repeating them.

• Add metadata. Required: identifies, severity, description. Threat rules also need a

single mitre_tactics. Optional specificity/sophistication default to medium. Also

available: max_hits, path_include. See the schema section of WRITING_RULES.md.

• Name and place the file under guarddog/analyzer/sourcecode/ as {type}-{category}-{detail}.yar.

.meta files (shared private rules) are {pattern-name}.meta.

• Write test cases (see below).
• Run the checklist at the end of WRITING_RULES.md before finishing.

Testing rules (the real harness)

The harness is tests/analyzer/sourcecode/test_sourcecode_yara.py. Key points:

• The YARA rule's internal name must be the file id with hyphens replaced by underscores.

File capability-filesystem-read.yar must contain rule capability_filesystem_read. The

no-false-positive test filters matches to this exact name, so a mismatch silently skips coverage.

• Positive test: add a file tests/analyzer/sourcecode/<rule-id>.<ext> containing code the rule

should flag (e.g. <rule-id>.py, .js, .go, .rb). The harness asserts the rule matches it.

These are matched by filename prefix.

• Negative test (false positives): add tests/analyzer/sourcecode/benign/<rule-id>.<ext> with

legitimate code that must NOT trigger the main rule. Every new or changed rule should have one.

• Compilation: all .yar files must compile, including include "...meta" references resolving.

Run the suite:

make test-yara-rules
# or directly:
uv run pytest tests/analyzer/sourcecode -k <rule-id>

Scan a real package or local path to sanity-check end to end (use uv run):

uv run guarddog pypi scan <package-name> --rules <rule-id>
uv run guarddog pypi scan /path/to/package --output-format json

Common patterns and pitfalls

• Install hooks are a capability, not a threat. A hook is process-spawning ability like

subprocess.call(). The threat is the LOLBAS tool *inside* the hook. Pair

capability.process.hooks with threat.process.hooks (hook context + curl/wget). See the

Advanced Patterns section of WRITING_RULES.md.

• Split LOLBAS by purpose: lolbas-proc.meta (bash, python, node) vs lolbas-net.meta

(curl, wget, nc). YARA cannot reliably use multiple private rules from one include, so keep

them separate and compose includes per rule.

• Avoid false positives from shebangs, READMEs, and non-hook code by establishing context

with private rules rather than matching a bare keyword anywhere.

• After adding or editing rules, regenerate docs if the repo expects it: make docs.