AgentSkillsCN

improve

Claude Code 辅导——对照最佳实践分析使用模式,识别利用率不足的功能,并就如何更高效地使用 Claude 提供建议。

SKILL.md
--- frontmatter
name: improve
description: Claude Code coaching - analyze usage patterns against best practices, identify underutilized features, and suggest improvements to how you work with Claude.

Optimize

Coach on Claude Code usage - are you using it effectively?

What to Analyze

Skills vs Sub-agents vs Hooks

Check if the user is using the right mechanism:

MechanismPurposeUse when
SkillWorkflow instructionsOrchestration, multi-step processes, templates
Sub-agentIsolated Claude instanceTool restrictions, fresh context, parallel/background work
HookAutomated scriptDeterministic automation, no LLM reasoning needed

Signs a skill should be a sub-agent:

  • Needs tool restrictions (read-only, no writes)
  • Needs isolated context (fresh perspective)
  • Should run in background

Signs something should be a hook:

  • Deterministic (same input → same output)
  • Should run automatically on events
  • Doesn't need LLM reasoning

Skill Naming Conventions

  • Actions: validate-*, generate-*, analyze-* (verb-first for things that do something)
  • Workflows: *-workflow, *-process (for multi-step orchestration)
  • Analysis: reflect, optimize, review (introspective skills)

Avoid generic names like helper, util, misc.

Prompting Patterns

  • Are instructions clear and specific?
  • Could tasks be batched or parallelized?
  • Are they using plan mode for complex tasks?
  • Are they over-specifying or under-specifying?

Tool Usage

  • Are they using the right tools? (e.g., Glob vs Bash find)
  • Are sub-agents being used for parallelizable work?
  • Are hooks being used for automation?
  • Is the scratchpad being used for temp files?

Repository Setup

  • Is there a CLAUDE.md with project conventions?
  • Are skills defined for repeatable workflows?
  • Is the .claude/ structure well-organized?
  • Are there hooks for common operations?

Workflow Efficiency

  • Are they context-switching unnecessarily?
  • Could doc-validator prevent rework?
  • Are they using /init, /compact, /clear appropriately?
  • Is work being persisted properly between sessions?

Process

  1. Use /ask-claude to get current best practices for the patterns you're evaluating (e.g., /ask-claude when should I use sub-agents vs skills?)
  2. Compare best practices against observed usage in this conversation
  3. Identify gaps and underutilized features
  4. Prioritize suggestions by impact

Note: This skill composes with /ask-claude. Don't skip step 1.

Output Format

code
## Usage Analysis

### What You're Doing Well
- [effective pattern observed]

### Opportunities

#### High Impact
- [significant improvement]
  - Current: [what you're doing]
  - Better: [recommended approach]

#### Quick Wins
- [small change, easy to adopt]

### Repository Improvements
- [specific setup suggestion]

### Commands/Features to Try
- [underutilized feature]