Defensive tools for working with AI. Designed for OpenPackage.
opkg install gh@pm0u/edc| Agent | Description |
|---|---|
adversarial-plan-reviewer |
Stress-tests implementation plans. Pokes holes in assumptions, identifies risks, and challenges feasibility before work begins. |
adversarial-code-reviewer |
Hunts for AI-generated failure modes in code changes. Targets reckless completion patterns, fabricated assumptions, scope violations, and engineering judgment issues. |
distill-learnings |
Extracts reusable lessons from completed work. |
| Command | Description |
|---|---|
/build-n-break |
Execute a task then adversarially review the result. You review disputes, not the full diff. |
/review-plan |
Run an adversarial review against a plan or design doc. |
/review-code |
Run an adversarial code review against recent changes. |
/implementation-summary |
Generate a summary of what was built and why -- intent and decisions, not a diff listing. |
/ton-of-bricks |
Reality-check validation. Forces real-world verification instead of trusting tests and types. |
/resolve-pr-feedback |
Fetch and address unresolved PR review comments. |
/derive-spec |
Derive a specification from a task description. |
/reverse-spec |
Reverse-engineer a specification from existing code. |
/trace-flow |
Trace a data or execution flow through the codebase. |
/mermaid |
Generate Mermaid diagrams from code or descriptions. |
/un-ai |
Strip AI slop from a document. Extract what it actually says in plain language. |
AI tools are useful for bounded, verifiable tasks. They are bad at judgment, taste, and knowing when they're wrong. These tools are built around that reality:
- Adversarial by default. Assume AI output is wrong until proven otherwise.
- Verification over generation. The hard part isn't writing code -- it's knowing the code is correct.
- Human judgment where it matters. Route human attention to disputes and decisions, not scanning diffs.
Stay safe out there, meatbags.
MIT