dongshuyan/compass-skills

What it solves

COMPASS Skills addresses the loss of state and context in long-running AI agent workflows. It prevents the need to replay entire conversation transcripts by managing user preferences, project progress, and task dependencies locally, ensuring that agents remain aligned with user goals across different sessions.

How it works

The project provides a set of SKILL.md packages that can be installed into AI coding agents (like Claude Code or Codex). These skills use Markdown instructions and local Python scripts to implement specific workflows:

• State Management: It maintains a local user profile for communication preferences and a repo-local task graph (DAG) to track goals and subtasks.
• Alignment: A clarification gate ensures ambiguous or high-risk tasks are narrowed down through focused questions before execution.
• Continuity: It compresses current conversation state into a paste-ready prompt for new AI sessions.
• Meta-Programming: It includes tools to turn successful run histories into new reusable skills or upgrade existing ones based on real evidence.

Who it’s for

Developers using AI coding agents who manage complex, long-term projects and need a structured way to maintain context, track progress, and ensure the agent doesn't deviate from the original objective.

Highlights

• Local-First Privacy: All task and profile data is stored locally; no data is uploaded to external servers.
• Task Forest: A visualizable task graph that survives conversation boundaries.
• Session Handoff: A specialized prompt generator for seamless transitions between AI conversations.
• Skill Engineering: Tools to build and refine skills directly from observed agent run history.