AgentSkillsCN

go-watch-this

多角度深度研究的并行编排器。同时启动多个Claude子代理与Codex实例，从不同视角对同一主题展开深入探究。当用户调用“/go-watch-this”指令，或提出“深度挖掘”“多角度研究”“彻底调查”等需求，又或是希望对某一问题进行并行分析时，可使用此功能。

SKILL.md
--- frontmatter
name: go-watch-this
description: >
  Parallel multi-angle deep research orchestrator. Spawns multiple Claude sub-agents
  and Codex instances to investigate a topic from different perspectives simultaneously.
  Use when user invokes /go-watch-this, or asks to "deep dive", "research from multiple angles",
  "investigate thoroughly", or wants parallel analysis of a problem.
argument-hint: "[prompt] [sub-agent num] [codex num]"

Go Watch This - Parallel Research Orchestrator

Deep-dive into any question by dispatching multiple AI agents in parallel, each tackling a different angle.

Argument Parsing

Parse from user input: /go-watch-this <prompt> [agent_count] [codex_count]

  • •prompt: The research question or topic (required)
  • •agent_count: Number of Claude sub-agents to spawn (default: 2)
  • •codex_count: Number of Codex calls to make (default: 1)

If counts are not provided as trailing numbers, use defaults.

Execution Workflow

1. Decompose into Research Angles

Before spawning agents, generate agent_count + codex_count distinct research angles for the prompt. Each angle must be meaningfully different. Strategies:

  • •Breadth: different facets of the problem (e.g., performance vs correctness vs maintainability)
  • •Depth: surface-level overview vs deep implementation details
  • •Perspective: user's view vs system internals vs external dependencies
  • •Contrast: pros/cons, alternatives, trade-offs

2. Create Task Tracking

Use TaskCreate to create one parent task for the overall research, then one child task per agent/codex call. This gives the user visibility into progress.

3. Select Sub-agent Types

For each Claude sub-agent, pick subagent_type based on the angle's nature:

Angle involvessubagent_type
Codebase files, architecture, patternsExplore
Web search, docs, external infogeneral-purpose
Implementation planningPlan
Code review, design evaluationarchitecture-reviewer

Use best judgment. When in doubt, use general-purpose.

4. Launch All Agents in Parallel

Critical: dispatch ALL agents in a single message with multiple Task tool calls.

For each Claude sub-agent:

code
Task(
  subagent_type=<selected_type>,
  prompt="Research angle: <angle>\n\nContext: <original_prompt>\n\nInvestigate this specific angle thoroughly. Return structured findings with key insights, evidence, and any open questions.",
  description="Research: <short_angle_label>"
)

For each Codex call, use Bash:

bash
codex exec -s read-only "<angle-specific prompt based on original question>"

Launch Claude sub-agents and Codex calls together in the same parallel batch.

5. Collect and Synthesize

After all agents complete:

  1. •Mark all child tasks as completed
  2. •Read each agent's output
  3. •Synthesize a unified summary:
markdown
## Research Summary: <topic>

### Angle 1: <label>
<key findings>

### Angle 2: <label>
<key findings>

...

### Cross-cutting Insights
<patterns, contradictions, or connections across angles>

### Open Questions
<unresolved items needing further investigation>
  1. •Mark parent task as completed

Example

User: /go-watch-this "How does ByteTrack handle occlusion?" 3 1

Decompose into 4 angles:

  1. •Algorithm internals (Explore) — trace ByteTrack code in tracking source files
  2. •Academic foundation (general-purpose) — web search for ByteTrack paper, occlusion strategy
  3. •Integration context (Explore) — how occlusion affects detection pipeline in this codebase
  4. •Codex perspective (Bash/codex) — ask Codex to analyze the tracking implementation

Launch all 4 in parallel, collect results, synthesize.