extract-skill（提取技能）

Extract Skill from Conversation

This skill helps you extract knowledge, patterns, and workflows from the current conversation and persist them as a reusable skill.

When to Use

• The user says "extract a skill", "extract this as a skill", or "save this for next time"
• A useful workflow, pattern, or piece of domain knowledge emerged during the conversation
• The user wants to update an existing skill with new learnings
• The conversation uncovered non-obvious steps, gotchas, or best practices worth preserving

Extraction Process

Phase 1: Identify What to Extract

Review the conversation for:

1. Workflows: Multi-step processes that were figured out through trial and error
2. Domain knowledge: Non-obvious facts, configurations, or constraints discovered
3. Gotchas and fixes: Problems encountered and their solutions
4. Patterns: Code patterns, command sequences, or templates that worked well
5. Decision rationale: Why certain approaches were chosen over alternatives

Summarize what you plan to extract and confirm with the user before proceeding.

Phase 2: Decide Destination

If the user already specified a skill or the destination is obvious from context, just proceed. Otherwise, use the AskQuestion tool (or ask conversationally) to clarify:

1. New or existing skill?

- If existing: Which skill to update? (list relevant skills from ~/.cursor/skills/ and .cursor/skills/) - If new: What should it be named?

1. Storage location (for new skills):

- Personal (~/.cursor/skills/) — available across all projects - Project (.cursor/skills/) — shared with the repository

Phase 3: Draft the Skill Content

When extracting into a new skill:

1. Choose a descriptive name (lowercase, hyphens, max 64 chars)
2. Write a specific description including WHAT and WHEN (third person)
3. Distill the conversation into concise, actionable instructions
4. Include concrete examples drawn from the conversation
5. Add any utility scripts or commands that were used

When updating an existing skill:

1. Read the existing SKILL.md
2. Identify where new learnings fit (new section, updated steps, additional examples)
3. Integrate without duplicating existing content
4. Preserve the existing structure and voice

Phase 4: Distillation Guidelines

The goal is to transform a messy conversation into clean, reusable instructions.

Do:

• Extract the final working approach, not the failed attempts (unless gotchas are instructive)
• Generalize from the specific case discussed (replace hardcoded values with placeholders)
• Include the "why" behind non-obvious steps
• Add context the agent wouldn't know without this conversation
• Keep it under 500 lines

Don't:

• Include conversation artifacts ("as we discussed", "you mentioned")
• Repeat information the agent already knows
• Include overly specific details that won't transfer to other situations
• Add verbose explanations where a code example suffices

Phase 5: Write and Verify

1. Create/update the skill file(s)
2. Verify the SKILL.md is under 500 lines
3. Check that the description is specific and includes trigger terms
4. Confirm with the user that the captured content is accurate

Example: Extracting a Debugging Workflow

If a conversation involved debugging a tricky deployment issue, the captured skill might look like:

---
name: debug-k8s-deployments
description: Debug Kubernetes deployment failures including CrashLoopBackOff, image pull errors, and resource limits. Use when pods are failing to start or deployments are stuck.
---

# Debug K8s Deployments

## Diagnostic Steps

1. Check pod status: `kubectl get pods -n <namespace> | grep -v Running`
2. Get events: `kubectl describe pod <pod> -n <namespace>`
3. Check logs: `kubectl logs <pod> -n <namespace> --previous`

## Common Issues

### CrashLoopBackOff
- Check if the entrypoint command exists in the container
- Verify environment variables are set (especially secrets)
- Look for OOMKilled in `describe` output → increase memory limits

### ImagePullBackOff
- Verify image tag exists: `docker manifest inspect <image>`
- Check imagePullSecrets are configured for private registries

Note how this extracts the diagnostic sequence and common solutions without any conversation artifacts.

Handling Edge Cases

Conversation had multiple topics: Ask which specific learning to extract, or suggest creating separate skills for distinct topics.

Learning is too small for a skill: Suggest creating a Cursor rule (.cursor/rules/) instead, which is better suited for single-line or short guidelines.

Existing skill needs major rewrite: Confirm with the user whether to restructure the existing skill or create a new one that supersedes it.