permission-patterns
Rules for evaluating, classifying, and deduplicating AI tool permissions
When & Why to Use This Skill
This Claude skill provides a comprehensive framework for evaluating, classifying, and deduplicating AI tool permissions. By implementing a standardized safety classification system—categorizing actions into ALLOW, ASK, or DENY—it ensures that AI agents operate within secure boundaries. The skill optimizes agent performance and security by identifying redundant permissions and enforcing strict rules against accessing sensitive local resources or irreversible operations.
Use Cases
- Automated Safety Classification: Categorize AI tool commands in real-time to distinguish between safe read-only operations and risky modifications that require user confirmation.
- Permission Deduplication: Optimize agent workflows by identifying when a specific permission request is already covered by a broader existing pattern, reducing redundant prompts.
- Security Guardrails: Prevent unauthorized access to sensitive files (like .env or private keys) and restricted local network addresses (localhost) through hardcoded DENY rules.
- Domain-Level Access Control: Manage granular access to cloud services and APIs by defining clear boundaries for root domains and subdomains (e.g., GitHub vs. GitHub Pages).
| name | permission-patterns |
|---|---|
| description | Rules for evaluating, classifying, and deduplicating AI tool permissions |
| version | "1.0.0" |
| author | "JacobPEvans" |
Permission Patterns
Unified patterns for permission safety classification and deduplication. Use these rules to evaluate permissions consistently.
Safety Classification
Classification rules for evaluating permission safety. Use these criteria to categorize permissions consistently.
Classification Rules
ALLOW - Read-Only and Safe Operations
Keywords: list, ls, show, info, view, get, describe, inspect, status, doctor, ping, check, --version, --help
Safe domains: github.com, docker.com, kubernetes.io, python.org, npmjs.com, official documentation sites
ASK - Modifications and Risky Operations
Keywords: update, set, edit, patch, modify, apply, rm, delete, remove, prune, clean, exec, run, eval, push, publish, deploy, kill, stop
Requires user confirmation before execution.
DENY - Irreversible Damage or Security Bypass
Keywords: sudo, chmod 777, dd, file patterns like **/.env, **/*_rsa, **/*.key, **/*secret*
Local addresses: localhost, 127.0.0.1, private IP ranges
Decision Criteria
- Read-only query + no secrets → ALLOW
- Modifies resources + reversible → ASK
- Irreversible or security risk → DENY
- Uncertain → ASK (conservative default)
Domain Coverage
Root domains cover their subdomains, but different root domains or TLDs are separate:
github.comcovers:api.github.com,docs.github.com,status.github.comgithub.iois a separate root domain (different TLD), does NOT covergithub.comand vice versagithub.comdoes NOT covergithubusercontent.com(separate root domain)localhostis separate fromlocalhost:3000(ports are distinct entities, not subdomains)
Local/private addresses always DENY:
localhost,127.0.0.1,192.168.x.x,10.x.x.xranges
Pattern Deduplication
Rules for detecting when a specific permission is already covered by a broader existing pattern.
Coverage Rules
WebFetch Domains
Domain coverage follows the same rules as the Domain Coverage section above. Ports are distinct:
localhostdoes NOT coverlocalhost:3000
File Paths
Broader wildcards cover more specific patterns:
Read(**)covers any Read permissionGlob(**/*)coversGlob(**/*.js),Glob(**/package.json)
Root Domain Recommendations
For well-known vendors (GitHub, Docker, Google, Apple, Microsoft), prefer root domain over individual subdomains.
If multiple subdomains found → suggest adding root domain instead.
Related Permission Suggestions
When discovering a safe permission, suggest related safe commands in the same family:
docker volume ls→ suggestdocker volume inspectaws s3 ls→ suggestaws s3 sync --dryrunnpm list→ suggestnpm outdated,npm audit
Commands Using This Skill
permissions-analyzeragent - Uses classification and deduplication to filter permissions during discovery/sync-permissionscommand - Indirectly uses this skill through the permissions-analyzer agent