claude-code-dispatch: Spawning Claude as Its Own Sub-Agent Is Either Genius or Overkill

366 stars on SkillsMP with zero movement in the last week. That's not a skill on a rocket ship — that's a skill that found its audience and stopped there. Which, honestly, tells you something useful before you even read the SKILL.md.

I've been poking at claude-code-dispatch from LigphiDonk's Oh My Paper project, and I have thoughts. Let me save you the 20 minutes I spent figuring out what this actually is.

What This Skill Actually Does

Strip away the framing and this skill is a single bash invocation:

claude --print --dangerously-skip-permissions \
  --model claude-opus-4-6 \
  -p "<task description>" \
  --cwd {baseDir}

That's it. The skill tells Claude to use exec (or bash) to shell out to the Claude Code CLI and spin up a child agent for complex coding work. You're using Claude to call Claude.

The declared use cases are multi-file refactors, large-scale edits (50+ lines), complex debugging, and writing training/evaluation scripts. Basically: anything that would normally make you wish you had a second agent working in parallel.

Why This Pattern Exists and Why It's Not Stupid

Here's the thing — recursive agent delegation sounds like a joke until you've actually hit the context window wall mid-refactor. The parent agent is managing high-level orchestration, tracking pipeline state, coordinating across multiple research stages. You don't want it burning 40k tokens doing a line-by-line rewrite of your data loader. You want to hand that off.

This is the core insight behind the skill: context is a finite resource, and not all tasks deserve equal allocation of it. Delegating to a sub-agent via CLI means the child process gets a fresh context window, reads the project's CLAUDE.md for grounding, does its work, and returns. The parent stays lean.

In the context of Oh My Paper — which is a full research pipeline harness with 5 agent roles, 34 skills, and a structured Survey → Experiment → Publication workflow — this makes more sense. You've got an orchestrator agent managing a long-running research session. When it hits a meaty implementation task, it punts to a sub-agent rather than context-bloating itself.

As a standalone skill outside that ecosystem? The value proposition gets thinner.

Key Capabilities Worth Noting

1. Explicit task scoping by complexity The SKILL.md is upfront about when to use this: 50+ line changes, multi-file refactors, tasks requiring deep code understanding. That's a reasonable heuristic. It's not trying to delegate everything — it's trying to delegate the right things.

2. --cwd {baseDir} for project context Passing the working directory means the sub-agent inherits your project structure and will pick up CLAUDE.md automatically. This is the right call. A sub-agent that doesn't know where it is will hallucinate paths and make a mess.

3. --print mode for non-interactive execution Using --print keeps this scriptable and composable. You get output back to the parent, not an interactive session that blocks. Important if you're building any kind of automation on top of this.

4. Best practices section is actually useful The guidance to include file paths, current logic, expected outcomes, and validation criteria in the task description is practical. Garbage-in-garbage-out applies doubly when you're delegating to a sub-agent that can't ask follow-up questions mid-task.

5. It's part of a larger, coherent system The real capability here isn't the bash command — it's that this skill is one node in a 34-skill, 5-agent ecosystem. The dispatch pattern only fully pays off when you have an orchestrator that knows when and how to use it.

Who Should Install This

Install it if: - You're already using or evaluating the Oh My Paper plugin for research workflows - You're building multi-agent pipelines on top of Claude Code and need a delegation primitive - You're running long research or engineering sessions where context management is a real constraint - You're comfortable with --dangerously-skip-permissions and understand what that flag actually does (it bypasses tool-use confirmation prompts — read the Claude Code docs before you run this in a production environment)

Skip it if: - You want a standalone skill that adds obvious value to a generic Claude Code session - You're doing normal day-to-day coding work — just use Claude Code directly, you don't need a sub-agent - You're not familiar with Oh My Paper and don't plan to install it — this skill is an internal component, not a general-purpose tool - You're in a security-sensitive environment and the --dangerously-skip-permissions flag makes your eye twitch (it should)

How to Install

If you're going the standalone route (not recommended, but possible):

# Project-level
mkdir -p .claude/skills
cp claude-code-dispatch.md .claude/skills/

# User-level
mkdir -p ~/.claude/skills
cp claude-code-dispatch.md ~/.claude/skills/

If you're doing this properly, just install the whole Oh My Paper plugin:

# In Claude Code:
/plugin marketplace add LigphiDonk/Oh-my--paper
/plugin install omp@oh-my-paper

Then restart Claude Code and run /omp:setup in your project.

Concerns and Limitations — Being Honest Here

The --dangerously-skip-permissions flag is load-bearing and alarming. This flag tells the sub-agent to skip all tool-use permission prompts. In a controlled research environment where you trust your task descriptions, fine. In any other context, you're handing a sub-agent a blank check. The skill doesn't caveat this enough.

The model is hardcoded to claude-opus-4-6. That's going to rot. Model names change, versions deprecate, and a skill that hardcodes a specific model string is going to break quietly when that model is no longer available or has been superseded. There's no fallback or configuration option exposed.

This is not a polished, general-purpose skill. The SKILL.md is partially in Chinese, the description is minimal, and the whole thing reads like an internal component that got published alongside the rest of the Oh My Paper skills. That's fine — but set your expectations accordingly. You're not getting documentation, you're getting a recipe.

The recursion risk is real but unaddressed. What happens if the sub-agent, given a complex enough task, decides to invoke claude-code-dispatch itself? The skill doesn't address this. In practice it's probably fine, but in a sufficiently automated pipeline it's a footgun worth thinking about.

No error handling or output validation. The skill tells you to provide good inputs but says nothing about what to do when the sub-agent fails, times out, or produces garbage. That's left entirely to the calling agent's judgment.

Verdict

This is a 3/5 skill that becomes a 4/5 if you're in the Oh My Paper ecosystem.

As a standalone install, I can't recommend it. It's a thin wrapper around a bash command with a security flag and a hardcoded model name. You can write this yourself in 30 seconds.

As part of Oh My Paper, it makes sense. It's the delegation primitive that the orchestrator uses to offload heavy coding work without blowing its context budget. In that context, it's doing exactly what it should — being a small, focused tool that does one thing.

If you're evaluating Oh My Paper as a research harness (and you probably should be if you're doing any kind of systematic ML research or paper writing in the terminal), then this skill comes along for the ride and you don't need to think about it separately. If you're not in that world, move on.

The 366 stars suggest a real community of Oh My Paper users found value in it. I trust that signal more than I trust the sparse documentation.

SkillsMP page: https://skillsmp.com/skills/ligphidonk-oh-my-paper-skills-claude-code-dispatch-skill-md

GitHub: https://github.com/LigphiDonk/Oh-my--paper/tree/main/skills/claude-code-dispatch