About Candor

Candor is a Claude Code plugin designed to enhance AI interactions by providing twelve distinct work-mode personas. These personas aim to replace sycophantic behaviors like flattery and hedging with directness, verification, and task-specific efficiency. The plugin automatically adapts Claude's style based on the task at hand, whether it's coding, debugging, security analysis, architecture design, writing, data curation, and more.

• Core Goal: To reduce common AI shortcomings such as reflexive agreement, hollow praise, caving to pushback without new evidence, padding, false balance, and claiming success without verification, while maintaining helpfulness.

• Anti-Sycophancy: The core skill ensures directness without devolving into manufactured contrarianism.

• Twelve Skills: Includes a shared core skill and eleven work-mode personas, each grounded in established personality-science frameworks.

• Auto-Invocation: Skills load automatically based on the context of your task (e.g., candor-coding for code review, candor-security for security-related queries).

• Explicit Invocation: Users can also manually invoke specific skills using commands like /candor:candor-core.

• Prompt-Level: Operates as instructional Markdown, requiring no code execution or special permissions, ensuring safety and transparency.

• Auditable: All behavioral rules are clearly defined in plain language within the SKILL.md files.

• Status: Currently in an experimental phase (v0.5.0), with a benchmarked performance indicating a preference for candor responses in blind tests.

Candor is licensed under the MIT license.