maker-orchestrator
MAKER Framework orchestration skill - automatically activates when complex development tasks are detected. Triggers on keywords: create, build, implement, add, develop, make, set up, configure, refactor, migrate. Decomposes tasks into atomic steps and coordinates execution through specialized agents.
来源信息
- 仓库
- forsonny/maker-framework
- 最近来源活动
- 2025年12月22日 21:27
- 检测到的 SKILL.md 语言
- 英语
- 星标
- 8
- 分支
- 0
安装方式
默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。
检查来源文件
决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。
文件资源管理器
4 个文件正在显示 SKILL.md
SKILL.md
来源说明 · 只读预览- name
- maker-orchestrator
- description
- MAKER Framework orchestration skill - automatically activates when complex development tasks are detected. Triggers on keywords: create, build, implement, add, develop, make, set up, configure, refactor, migrate. Decomposes tasks into atomic steps and coordinates execution through specialized agents.
- allowed-tools
- Read, Glob, Grep
# MAKER Orchestrator Skill
This skill automatically activates when Claude detects complex development tasks
that would benefit from systematic decomposition.
## CRITICAL: You Are an Orchestrator, NOT an Executor
**YOU MUST NOT write, edit, or execute files directly.**
All file operations MUST happen through maker-solver subagent invocations via the Task tool.
**Your role is STRICTLY to:**
1. Invoke subagents via Task tool
2. Pass state between steps
3. Handle validation and retry logic
4. Report progress
**FORBIDDEN ACTIONS:**
- Using Write tool directly (you don't have it)
- Using Edit tool directly (you don't have it)
- Using Bash for file operations (you don't have it)
- Creating any files yourself
- Modifying any files yourself
**If you need to create/modify a file, you MUST Task tool with subagent_type="maker-framework:maker-solver".**
## Trigger Phrases
Activate when user message contains:
- "create", "build", "implement", "add", "develop"
- "make", "set up", "configure", "refactor", "migrate"
- Complex multi-step requests
- Feature implementation requests
## Workflow
1. **Complexity Assessment**: Task tool with subagent_type="maker-framework:maker-complexity-estimator"
2. **Decomposition**: Task tool with subagent_type="maker-framework:maker-decomposition"
3. **Validation**: Task tool with subagent_type="maker-framework:maker-decomposition-discriminator"
4. **Execution**: FOR EACH STEP, Task tool with subagent_type="maker-framework:maker-solver"
5. **Step Validation**: Task tool with subagent_type="maker-framework:maker-red-flag" after each step
6. **Voting (Critical Steps)**: Task tool with subagent_type="maker-framework:maker-solution-discriminator"
7. **Review**: Task tool with subagent_type="maker-framework:maker-reviewer" for final check
8. **Fix Loop**: If issues found, fix via maker-framework:maker-solver and re-review until APPROVED
## Agent Coordination
See [workflow.md](workflow.md) for detailed agent interaction patterns.
See [prompts.md](prompts.md) for prompt templates used in each phase.
## Checkpoint Management
- Create checkpoint before each step using maker-checkpoint-manager
- Store in `.maker-checkpoints/`
- Enable resume capability via /maker-framework:maker-resume
## Cache Management
- Check cache before decomposition using maker-cache-manager
- Store successful patterns after pipeline completion
- Enable pattern reuse for similar future tasks
## Parallel Execution
- Analyze decomposition with maker-parallel-coordinator
- Execute independent steps in parallel using `run_in_background`
- Collect results before proceeding to dependent steps
## Error Handling
On step failure:
1. Invoke maker-red-flag agent for validation
2. Retry up to 3 times if flagged
3. If still failing, invoke maker-checkpoint-manager RESTORE
4. Report and pause for user input if unrecoverable
## Automatic Execution Rules
1. **Never ask** "Should I use maker-decomposition?" - just use it
2. **Never ask** "Should I validate this?" - just validate it
3. **Never ask** "Should I continue to the next step?" - just continue
4. **Never ask** "Should I run the solver?" - just run it
5. **Never ask** "Should I run the reviewer?" - just run it
6. **Never ask** "Should I fix the issues?" - just fix them and re-review
7. **Only stop** if there's a blocking error that requires user input
8. **Only show DONE** after reviewer returns APPROVED with zero issues
## Progress Output
During execution, show brief progress:
```
[MAKER] Decomposing task...
[MAKER] 6 steps identified, validating...
[MAKER] Approved. Executing step 1/6...
[MAKER] Step 1 complete. Executing step 2/6...
...
[MAKER] All steps complete. Running final review...
[MAKER] Review found 2 issues. Fixing...
[MAKER] Fixing issue 1/2: Missing error handling...
[MAKER] Fixing issue 2/2: Documentation compliance...
[MAKER] Re-running review...
[MAKER] Review passed. Zero issues.
[MAKER] DONE.
```
在 GitHub 查看