Four markdown skills to stop your AI agent from losing the plot

What it does

COMPASS Skills is a set of four SKILL.md files that install into AI coding agents such as Claude Code and OpenAI Codex. They add four local workflows: task-clarifier aligns goals before expensive execution, task-forest maintains a repo-local task DAG with progress and decisions, session-handoff-prompt compresses conversation state into a paste-ready continuation prompt, and user-profile-keeper stores auditable collaboration preferences and risk boundaries. All data stays in the workspace; the README explicitly promises no upload of task or profile data.

The interesting bit

Rather than building a new framework, the project treats agent instructions as a lightweight OS layer. It encodes long-running state—user context, project context, goal context, and handoff context—into Markdown and YAML that agents already understand. The task forest even exports to an interactive HTML DAG view, turning opaque agent sessions into a visual trail of what happened and why.

Key highlights

• Four installable skills covering clarification, task memory, session handoff, and user alignment
• Repo-local storage under .agent-workbench/ and .compass-skills/; runs fully offline with no cloud upload
• Compatible with Claude Code, Codex, and other agents via the skills CLI or manual copy
• Task forest tracks goals, subtasks, dependencies, deviations, and decisions in a structured DAG
• Session handoff generates concise continuation prompts so new conversations don’t replay entire transcripts
• Alpha status; local scripts use only Python standard-library components

Caveats

• Explicitly alpha status, so rough edges are expected
• User profiles stored in local plaintext without encryption; the README warns against storing secrets or credentials
• Effectiveness depends on the agent actually following the SKILL.md instructions—if the agent ignores them, the guardrails are just polite Markdown

Verdict

Useful if you run long, multi-session agent workflows and want persistent context without vendor lock-in. Skip it if you prefer a fully managed memory service or don’t want to curate local task state yourself.