cortex-skills-loop（皮层技能循环）

Cortex Skills Loop

Overview

The cortex CLI includes an AI-powered recommendation engine that learns from usage patterns. This skill establishes the workflow for participating in that learning loop: get recommendations when context shifts, provide feedback on recommendation quality, and rate skills after use. Each interaction improves future recommendations.

When to Engage

Signals That Trigger Recommendations

Run cortex skills recommend when any of these mismatch signals appear:

• File pattern shift — git diff or the working set includes file types not covered by active skills (e.g., .tf Terraform files appear but no infrastructure skill is active, or **/auth/** paths get touched without security skills loaded)
• Agent activation change — a new agent gets activated that likely has complementary skills not yet loaded (the recommender maps agents to skill sets internally)
• Explicit domain pivot — the user switches focus to a different domain ("now let's handle the database migrations") and the current skill set is oriented elsewhere
• Skill gap felt — a task requires domain knowledge that no active skill covers, or an active skill is providing no value to the current work

Signals That Trigger Rating/Feedback

• Task completion — a skill was active during work that just finished successfully
• Recommendation acted on — a recommendation was recently followed or dismissed
• Negative experience — a skill actively misled or produced unhelpful guidance

When Not to Engage

Do not run recommendations on every session start or for routine tasks where the active skill set is clearly appropriate. The loop adds value through selective, signal-driven use.

Workflow

1. Recommend — Surface Relevant Skills

When a context change is detected, run:

cortex skills recommend

This analyzes the current git state, active agents, and historical patterns to suggest skills grouped by confidence level:

• High confidence (>=0.8): Auto-activate candidates — consider activating immediately
• Medium confidence (0.6-0.8): Worth reviewing with the user
• Low confidence (<0.6): Informational only

To check what is currently active before acting on recommendations:

cortex status

For full option details: cortex skills recommend --help

2. Feedback — Improve Recommendation Quality

After a recommendation has been acted on (activated or dismissed), record whether it was useful:

cortex skills feedback <skill-name> helpful --comment "Caught auth vulnerability early"
cortex skills feedback <skill-name> not-helpful --comment "Not relevant to this API work"

Positive feedback with context available teaches the recommender to associate the current project context with that skill for future sessions. Always provide a --comment when possible to enrich the learning signal.

For full option details: cortex skills feedback --help

3. Rate — Record Skill Effectiveness

After a task completes and a skill was active during the work, rate its contribution:

cortex skills rate <skill-name> --stars <1-5> --review "Description of experience"

Additional flags to enrich the signal:

• --helpful / --not-helpful — binary usefulness indicator
• --succeeded / --failed — whether the task the skill supported succeeded

Ratings feed back into the recommendation engine. Highly rated skills get prioritized in future suggestions; low-rated skills get demoted.

To view existing ratings before adding one: cortex skills ratings <skill-name>

To discover top-performing skills: cortex skills top-rated

For full option details: cortex skills rate --help

Closing the Loop

The recommend-feedback-rate cycle is cumulative. Each interaction updates the SQLite-backed learning database, improving four recommendation strategies:

1. Semantic similarity — embeds successful session contexts for future matching
2. Rule-based — matches file patterns to skill suggestions
3. Agent-based — maps active agents to complementary skills
4. Pattern-based — promotes skills with high historical success rates

When multiple strategies converge on the same skill, confidence gets boosted. Consistent feedback and ratings are what make this convergence happen over time.

Quick Reference