Claude Code Dynamic Workflows: A Deep Dive and Best Practices

Claude Code's Dynamic Workflow can orchestrate dozens to hundreds of subagents working in parallel. It is highly powerful, but there are a few pitfalls you need to be aware of in advance.

Table of Contents

1. What is Dynamic Workflow
2. Enabling and Disabling
3. Trigger Methods
4. Keyword Highlighting and Cancellation
5. Permission Configuration
6. Cost and Token Consumption
7. Ultracode Mode
8. Managing Runs: /workflows
9. Saving as Reusable Commands
10. Resumption and Limitations
11. Complete Cheat Sheet

What is Dynamic Workflow

A Dynamic Workflow is a JavaScript script written by Claude that orchestrates multiple subagents in the background to execute tasks during runtime.

✅ Suitable Scenarios: Large-scale code audits, 500-file migrations, multi-source cross-validation research, multi-perspective solution comparisons.

❌ Unsuitable Scenarios: Simple Q&A, a few subtasks (use subagents directly), repetitive operations with fixed instructions (use skills).

Enabling and Disabling

Checking the Status

/config   # Open the configuration interface to check the Dynamic Workflows status

Disabling Methods (Choose One)

Re-enabling

Simply reverse the operation. Once disabled:

• Built-in commands like /deep-research become unavailable.
• The workflow keyword no longer triggers it.
• The ultracode option is removed from the /effort menu.

Trigger Methods

Method 1: Keyword Trigger

Include the workflow keyword in your prompt:

Run a workflow to audit every API endpoint for missing auth checks

Claude Code will highlight the word workflow to indicate that the trigger intent has been recognized.

Method 2: Ultracode Automatic Trigger

By setting /effort ultracode, Claude automatically plans a workflow for every substantial task.

Method 3: Saved Commands

Saved workflows become /<name> commands that can be called directly. Built-in examples include /deep-research.

Keyword Highlighting and Cancellation

When your prompt contains workflow, Claude Code highlights the word, indicating that a dynamic workflow is about to be triggered.

How to Cancel the Trigger?

Press Alt+W to cancel the highlight, reverting the current prompt to normal conversation mode.

⚠️ Important: Some terminals intercept Alt key combinations and require additional configuration for Alt+W to work.

Terminal Configurations

🟣 Ghostty

Add the following to ~/.config/ghostty/config:

macos-option-as-alt = true

🔵 iTerm2

Preferences → Profiles → Keys → Left Option Key: Set to Esc+

🟢 Kitty

Generally supported by default. If not, add this to kitty.conf:

macos_option_as_alt yes

🟠 WezTerm

In the configuration:

config.keys = {
  { key = "w", mods = "ALT", action = wezterm.action.SendKey({ key = "w", mods = "ALT" }) },
}

Restart the terminal for the changes to take effect.

Permission Configuration

Core Rules

Startup Confirmation (By Permission Mode)

⚠️ Prerequisites for Long Runs

Add required commands to the allowlist in advance, otherwise the workflow will pause and get stuck on a prompt mid-execution!

Method 1: Tell Claude directly:

Allow git and npm commands without prompting

Method 2: Manually edit settings.json:

{
  "permissions": {
    "allow": [
      "Bash(git *)",
      "Bash(npm *)",
      "WebSearch",
      "WebFetch"
    ]
  }
}

Configuration File Locations:

Cost and Token Consumption

Cost Characteristics

• A single workflow spawns a large number of agents, consuming significantly more tokens than normal conversations.
• All agents use the current session's model by default.
• Counts towards plan usage and rate limits.

💰 Cost-Saving Tips

Model Cost Comparison (Rough Estimate)

Haiku  <  Sonnet  <  Opus
💰 Cheap      💰💰 Medium      💰💰💰 Expensive

You can specify different models for different phases within the workflow script:

agent('Simple classification task', { model: 'haiku' })
agent('Complex reasoning task', { model: 'opus' })

Ultracode Mode

What is it?

Ultracode = xhigh Reasoning Depth + Automatic Workflow Orchestration

Enabling

/effort ultracode

Behavior

• Automatically plans a workflow for every substantial task.
• A single request may trigger multiple sequential workflows: understand → design → implement → verify.
• Skips startup confirmation under the auto permission mode.

Trade-offs

• 🔴 Extremely high token consumption; each task may spawn dozens or hundreds of agents.
• 🟡 Longer execution time, but yields more comprehensive and reliable results.
• 🔵 Valid only for the current session.

Exiting

/effort high    # Downgrade to normal high effort
/effort medium  # Downgrade to medium effort

Suitability Guide

Managing Runs: /workflows

/workflows

Opens the workflow management interface to view running and completed workflows.

⌨️ Operation Shortcuts

Progress Information

Each phase displays:

• Number of agents
• Token consumption
• Elapsed time

Saving as Reusable Commands

How to Save

1. /workflows → Select a satisfactory run
2. Press s
3. Press Tab to toggle the save location
4. Press Enter to confirm

Save Locations

Usage

Once saved, it becomes a /<name> command:

/my-audit          # Call directly

It appears in the / autocomplete list, alongside built-in commands (like /deep-research).

Priority

In case of a name collision, 📦 Project-level > 👤 Personal-level.

Resumption and Limitations

Resuming Execution

• Can be resumed within the same session after pausing.
• Completed agents return cached results and do not rerun.
• How to resume: /workflows → p, or ask Claude to restart using the same script.

⚠️ Not Retained Across Sessions

Workflows are not retained after exiting Claude Code. They must be rerun in the next session. However, saved command scripts remain and can be re-executed.

Runtime Limitations

Need Phase Sign-offs?

Break it down into multiple independent workflows, reviewing each upon completion before running the next.

Complete Cheat Sheet

# 🔄 Enable/Disable
/effort ultracode          # Enable ultracode (xhigh + auto workflow)
/effort high               # Exit ultracode
/config                    # Toggle Dynamic Workflows switch

# 🚀 Triggers
Include "workflow" in prompt # Keyword trigger
/<saved-name>              # Call a saved command
/deep-research <question>  # Built-in command

# 📋 Management
/workflows                 # Open management interface
Alt+W                      # Cancel keyword highlight (prevents workflow trigger)

# ⌨️ Interface Operations
p = Pause/Resume
x = Stop
s = Save as command
Enter = View details
Esc = Go back

# 🔐 Permissions
Say "allow git commands without prompting" → Claude updates the allowlist
Or manually edit the permissions.allow array in settings.json

# 💰 Cost Saving
/model                     # Check current model
Specify haiku for simple phases # In script: { model: 'haiku' }
/workflows → x             # Promptly stop unnecessary runs

Compiled based on official Claude Code documentation, last updated: 2026-05-30