Learn from Conversation

Critical Reminder

Skills are inserted into the agent's context window. Always be extremely concise to save tokens. Every word counts.

Skill Locations

• •Primary: ~/.cursor/skills/ (user home) - most skills are here
• •Secondary: .cursor/skills/ (project root) - rare, but check both locations

Workflow

Step 1: Assess (Do Not Modify Yet)

1. •Review conversation history for:

• •Times the user didn't like something and requested changes
• •Patterns or preferences that should become automatic
• •Recurring corrections or clarifications

1. •List existing skills (check both locations):

• •Get skill names/descriptions: head -n 3 ~/.cursor/skills/*/SKILL.md
• •If already in your context window, use that instead
• •Do not read full skill files yet

1. •Identify candidates:

• •Which existing skills were involved in the conversation?
• •Would a new skill be needed? (New skills are rare)

Step 2: Report to User

Present findings:

• •Which existing skills should be updated (and why)
• •If a new skill is needed, describe its purpose
• •Do not make changes yet

Step 3: Get Confirmation

Wait for user approval

Step 4: Execute (After Confirmation)

For existing skills:

• •Read only the skills that need updates
• •Make minimal, targeted additions
• •Preserve existing structure and style
• •Avoid verbosity - add only essential information

For new skills:

• •Follow the create-skill workflow or mimic other skills
• •Keep it concise

Principles

• •Minimal changes: Extend existing skills slightly, don't rewrite them
• •Token efficiency: Skills consume context - every addition must justify its cost
• •Selective reading: Only read skills that clearly need updates
• •New skills are rare: Most learnings fit into existing skills