搜索

PythonReactSEOPDF
GitHub Skill

ultrathink-orchestrator

Auto-invokes ultrathink workflow for any work request (default orchestrator)

rysweet☆ 16 仓库 Stars7 仓库 Forks

安装与使用

请根据来源仓库和你的 AI 工具说明安装此技能。先阅读技能文件和依赖要求,再添加到工作流。

查看来源 ↗使用指南

技能说明

Ultrathink Orchestrator Skill

Purpose

This skill provides automatic orchestration for development and investigation tasks. It detects the task type from keywords and delegates to the appropriate workflow skill (investigation-workflow or default-workflow).

Auto-activation priority is LOW (5) to allow more specific skills to match first. When activated, this orchestrator selects between investigation-workflow and default-workflow based on the user's request keywords.

This skill acts as a thin wrapper around the canonical ultrathink command, following the amplihack pattern of single-source-of-truth for command logic.

Canonical Sources

This skill is a thin wrapper that references canonical sources:

  • Primary Command: ~/.amplihack/.claude/commands/amplihack/ultrathink.md (278 lines)
  • Workflow Sources:
    • Development: ~/.amplihack/.claude/workflow/DEFAULT_WORKFLOW.md
    • Investigation: ~/.amplihack/.claude/workflow/INVESTIGATION_WORKFLOW.md

The canonical command contains complete task detection logic, complexity estimation, and orchestration patterns for both investigation and development workflows.

⛔ MANDATORY EXECUTION PROCESS (5 Steps)

When this skill is activated, you MUST follow this exact 5-step process:

Step 1: Read Canonical Command (MANDATORY)
Read(file_path="~/.amplihack/.claude/commands/amplihack/ultrathink.md")

Validation Checkpoint: Confirm ultrathink.md content is loaded before proceeding.

Step 2: Detect Task Type (MANDATORY)

Analyze user request using keywords from canonical command:

  • Q&A keywords: what is, explain briefly, quick question, how do I run, simple question
  • Investigation keywords: investigate, explain, understand, analyze, research, explore
  • Development keywords: implement, build, create, add feature, fix, refactor, deploy
  • Hybrid tasks: Both investigation and development keywords present

Validation Checkpoint: Task type must be determined (Q&A, Investigation, Development, or Hybrid).

Step 3: Invoke Workflow Skill (MANDATORY - BLOCKING)

⛔ THIS IS A BLOCKING REQUIREMENT - Session will be terminated if skipped.

For Q&A tasks:

Read(file_path="~/.amplihack/.claude/workflow/Q&A_WORKFLOW.md")

For Investigation tasks:

Skill(skill="investigation-workflow")

For Development tasks:

Skill(skill="default-workflow")

For Hybrid tasks:

Skill(skill="investigation-workflow")
# After investigation completes:
Skill(skill="default-workflow")

Validation Checkpoint: Confirm Skill tool was invoked OR proceed to Step 4.

Step 4: Fallback to Read Tool (IF Step 3 Fails)

Only if skill invocation fails, use Read tool as fallback:

Investigation fallback:

Read(file_path="~/.amplihack/.claude/workflow/INVESTIGATION_WORKFLOW.md")

Development fallback:

Read(file_path="~/.amplihack/.claude/workflow/DEFAULT_WORKFLOW.md")

Validation Checkpoint: Confirm workflow content is loaded in context.

Step 5: Execute Workflow Steps (MANDATORY)

Follow all steps from loaded workflow without skipping.

Validation Checkpoint: All workflow steps must be completed.


Enforcement

Power-Steering Detection:

  • Validates Skill tool invocation occurred
  • Validates Read tool fallback if Skill failed
  • Blocks session termination if workflow not loaded

Self-Check Protocol: Before proceeding past Step 3, verify:

  • Skill tool invoked with workflow skill name, OR
  • Read tool used to load workflow markdown
  • Workflow content confirmed in context

No Shortcuts:

  • You cannot skip workflow invocation
  • You cannot assume workflow content
  • You must use Skill or Read tool explicitly

Why This Pattern

Benefits:

  • Single source of truth for orchestration logic in canonical command
  • No content duplication between command and skill
  • Task detection rules defined once, maintained once
  • Changes to ultrathink command automatically inherited by skill

Trade-offs:

  • Requires Read tool call to fetch canonical logic
  • Slight indirection vs. inline implementation

This pattern aligns with amplihack philosophy: ruthless simplicity through elimination of duplication.

Related Files

  • Canonical Command: ~/.amplihack/.claude/commands/amplihack/ultrathink.md
  • Development Workflow Skill: ~/.amplihack/.claude/skills/default-workflow/
  • Investigation Workflow Skill: ~/.amplihack/.claude/skills/investigation-workflow/
  • Canonical Workflows:
    • ~/.amplihack/.claude/workflow/DEFAULT_WORKFLOW.md
    • ~/.amplihack/.claude/workflow/INVESTIGATION_WORKFLOW.md

使用示例

查看来源仓库中的使用示例与配置说明。

查看来源 ↗

来源历史

收录时间
2026-01-20
索引更新时间
2026-01-27
Branch
main

收录与同步日期属于本站记录。源码修改历史请查看来源仓库。

GitHub ↗

chat Comments (0)

chat_bubble_outline

No comments yet. Be the first to share your thoughts!