autocritic

Agents often produce code too quickly, ignore architectural trade-offs, or clutter the terminal with verbose commentary. This skill enforces a partner-like collaboration model where the agent investigates first, checks in before any modification, and performs an invisible self-review to ensure high-quality output without the "theater" of fake tests.

What it does

• Ambiguity classification categorizes requests to decide whether to stop and ask for clarity or proceed with best-practice assumptions.
• Pre-modification check-ins ensure the agent never edits files or creates scaffolding without first sharing a plan for user approval.
• Invisible self-review runs an internal senior-engineer perspective shift to catch edge cases, security flaws, and logic gaps before delivery.
• State persistence maintains a .project/ directory to track session goals, architectural decisions, and deferred uncertainties.
• Refinement reporting provides a concise note only when substantive corrections were made during the internal review process.

How it works

1. Analyze the request for ambiguity and either ask up to three targeted questions or proceed to investigation.
2. Investigate the codebase or environment freely without needing permission to read files or run diagnostics.
3. Check in by presenting a concrete action plan and waiting for user confirmation before modifying any project state.
4. Execute and review the task internally, applying domain-specific lenses (Code, Content, or Architecture) to prune errors.

Frameworks & tools

This is a model-agnostic behavioral layer designed for terminal-based AI environments like Claude Code, Gemini CLI, Aider, and Cursor. It works across any programming language or stack by governing the agent's interaction logic rather than specific syntax.

Why this beats prompting it yourself

Prompting an agent to "be careful" or "check in" is often ignored during long sessions. This skill provides a rigid execution framework with specific ambiguity routing and a mandated internal checklist that forces a perspective shift, preventing the agent from agreeing with its own initial mistakes.

Use cases

• Preventing agents from refactoring the wrong files in large, complex repositories.
• Ensuring architectural decisions are approved before a multi-file migration begins.
• Stripping out "AI fluff" and fake test blocks from CLI-based coding sessions.
• Maintaining a durable log of technical debt and deferred questions in the .project/ folder.

Known limitations

The skill requires the agent to have file-system access to maintain project state. Performance may degrade if the model lacks an extended reasoning trace for the internal review step.