manage-artifact
为创建或修复Claude Code指令工件加载方法上下文(发现:实现清晰度)。当工件类型已知(命令、技能、钩子、协议、外部工具配置)且实现方法需要方法指导时,在实现清晰度处进行评估。
name: manage-artifact description: Load methodology context for creating or fixing Claude Code instruction artifacts (discovery: implementation-clarity). Evaluate at implementation-clarity when artifact type is known (command, skill, hook, protocol, external tool config) and implementation approach needs methodology guidance.
Manage Artifact
Provides methodology context for creating or fixing instruction artifacts.
Artifact Type Routing
After rubber-duck diagnoses the artifact type, load the appropriate methodology:
| Artifact Type | Reference Files | Fix Tool |
|---|---|---|
| Command/Prompt | command.md + persuasion-principles.md | /stage |
| Skill | skill.md + persuasion-principles.md | skill-creator |
| Hook | (not yet implemented) | settings.json |
| Protocol | protocol.md + persuasion-principles.md | CLAUDE.md edit |
| External Tool Config | external-tool-config.md | hippocampus |
Note: Command/Prompt covers ALL prompt types: slash commands, agent prompts, prompt templates. Any engineered prompt follows the same methodology.
Note: Hooks use settings.json (declarative config) - no persuasion principles needed. External tools use hippocampus for documentation - no persuasion principles needed. Commands, skills, and protocols require persuasion techniques to ensure AI compliance.
Workflow
- Confirm artifact type from rubber-duck diagnosis
- Read the appropriate reference file(s) for methodology context
- For Command/Skill/Protocol: Also read persuasion-principles.md
- Apply the methodology during implementation using persuasive language (Authority + Commitment + Social Proof)
- PRE-COMMIT PAUSE before finalizing (see verification workflow below)
Verification Workflow
After implementing the fix, do NOT commit immediately. Follow this flow:
PRE-COMMIT PAUSE
Announce: "Fix implemented. Verify in a new project session before committing."
Accept Verification Session Path
After user provides Session C conversation path, START MONITORING IMMEDIATELY.
User providing a path = start monitoring loop. No confirmation needed.
Start Monitoring Loop (MANDATORY)
When user provides a Session C path, YOU MUST start the monitoring loop:
sleep 30 && ~/.claude/skills/manage-artifact/scripts/monitor_session.sh "<path>" 5
Repeat this cycle until you see verification results (success or failure). Do NOT do a single extraction and stop - keep watching Session C progress in real-time.
Script arguments:
path: Conversation .jsonl file path (required)num_messages: Messages to display (default: 5)
Commit Decision
Based on monitoring results:
- Success observed: Proceed to commit and PR
- Failure observed: Return to rubber-duck with new context, iterate on fix
Current Coverage
Command, Skill, Protocol, and External Tool Config are supported. Hook support will be added when needed.
For commands/prompts (slash commands, agent prompts, templates):
- Read references/command.md for ADR 001-004 methodology, /stage usage, and Anthropic template guidance
- Read references/persuasion-principles.md for Authority + Commitment + Social Proof techniques
For skills:
- Read references/skill.md for skill-creator methodology
- Read references/persuasion-principles.md for persuasion techniques
For protocols:
- Read references/protocol.md for ADR-008 (WHEN/WHY patterns), CLAUDE.md structure guidance
- Read references/persuasion-principles.md for persuasion techniques
For external tool config:
- Read references/external-tool-config.md for checklist workflow
- No persuasion principles needed - external tools are configured, not instructed