compass-skills
Long-running agent work loses user context, project context, goal alignment, and handoff state at every conversation boundary.
COMPASS Skills gives AI coding agents four local SKILL.md skills: task-clarifier (aligns goals, scope, and acceptance criteria before ambiguous or risky work), task-forest (repo-local task graph with goals, dependencies, and progress), session-handoff-prompt (compresses the current conversation into a paste-ready prompt for a new session), and user-profile-keeper (a local, correctable collaboration-preferences profile). Works with Claude Code and Codex.
- Clarify ambiguous or high-risk tasks into concrete decision criteria before the agent executes
- Maintain a repo-local task graph that survives AI conversation boundaries
- Generate a paste-ready handoff prompt so a new session continues work without replaying the transcript
Long-running agent work loses user context, project context, goal alignment, and handoff state at every conversation boundary. COMPASS stores all four locally — a correctable user profile, a repo-local task graph, a continuation prompt, and a clarification gate before ambiguous or risky execution.
Engineering teams building or deploying AI agents that require consistent task execution and goal alignment.
https://github.com/dongshuyan/compass-skills
By dongshuyan
How to Get It
npx skills add dongshuyan/compass-skills --skill '*' -a claude-code
Tip: Paste this into a Claude Code conversation. Verify command matches your Claude Code version.
After installing, paste this into Claude:
Help me clarify ambiguous or high-risk tasks into concrete decision criteria before the agent executes
Trust Signals Auto-scanned
Community Pulse New
No community discussions found yet. This doesn't mean the tool isn't good — it may be new or serve a niche use case.
Reviewer notes
Auto-scanned review. These are observations, not a security certification.
Scored from trust signals (evidence-eval-v1): 325 GitHub stars; contributors unknown; last commit 0d ago; license MIT.
Things to check
- Alpha-status project (self-labeled). Skills are local files — keep references/, scripts/, and agents/ subdirectories intact when installing manually. Evaluated from public trust signals, not hands-on tested.
- Single maintainer. Consider the risk if this person stops maintaining the project.
How to evaluate tools before deploying →
Data shown here comes from public APIs and automated scanning. Reviewer notes reflect one person's experience. This is not a security certification or legal recommendation. Always evaluate tools according to your own organization's policies.
Evaluation
Scored from trust signals (evidence-eval-v1): 325 GitHub stars; contributors unknown; last commit 0d ago; license MIT.