A threat modeling skill for coding agents: enables Claude Code or Codex to perform read-only inspections of your repository, systematically analyzing code, trust boundaries, threats, existing protections, and mitigation recommendations through a structured workflow, with conclusions required to cite files and line numbers as evidence.
Stars 310 (+57 / 24h)
Rationale
- 36 architectural workflows (Web, API, Mobile, Microservices, Multi-tenant SaaS, RAG, Agentic AI, IoT, Blockchain, etc.), each with a quick version and a six-stage deep-dive version; 29 methodology prompts, including STRIDE, PASTA, LINDDUN, STPA, MITRE ATLAS, CVSS, etc.
- The runbook mandates read-only inspections; builds, tests, installations, and network access require explicit authorization from you; results are written to the threat-model/ directory, including review scope, reports, and JSON, allowing you to later use change mode to compare branch modifications against existing models
- For MCP servers and agent tools, it maps tools to backend credentials and actual permissions, checks the blast radius of shared tokens, and assesses whether modified tool descriptions could mislead other trusted tools
Usage
- git clone https://github.com/jmanico/grim-prompter.git ~/grim-prompter, then symlink skills/threat-model to ~/.claude/skills/ (for Codex, use ~/.agents/skills/), and start a new session
- In Claude Code, enter /threat-model, /threat-model quick api, /threat-model deep web, or /threat-model change main...feature (use $threat-model for Codex); for other agents, add the snippets from the README to AGENTS.md
Risks
- The author emphasizes that the prompts are drafts, threat models cannot prove system security, and risk ratings and approvals are your responsibility; reports may contain sensitive details, so review them before committing to version control
- Author Jim Manico also sells commercial secure coding prompts; the license is MIT, and the repository was created on 10 28