llm-box Workflow Skill

When to Use

Use this skill when the user wants to:

• Automate multi-step terminal tasks — fetching data, processing it, saving results
• Create reusable pipelines — workflows that can be run repeatedly
• Chain commands — where output of one step feeds into the next
• Batch process data — transform, filter, combine multiple data sources
• Integrate LLMs into automation — use Ollama, DeepSeek, or OpenAI-compatible models
• Replace fragile bash scripts — with structured, auditable YAML workflows

How llm-box Works

Plain English description → YAML workflow → Execute with progress

llm-box generates a YAML workflow file from a natural language description.

The workflow is deterministic and reproducible — same workflow always produces

the same result. Users can edit the YAML by hand if they want to tweak things.

Quick Reference

CLI Commands

# Generate a workflow from plain English
llm-box create "<description>"

# Run a workflow file
llm-box run <workflow.yaml>

# List all available nodes
llm-box list

# Validate a workflow file without running
llm-box validate <workflow.yaml>

# Run in safe mode (disables execute node)
llm-box --safe-mode run <workflow.yaml>

# Dry run (show steps without executing)
llm-box --dry-run run <workflow.yaml>

Available Nodes

Utility Nodes:

LLM Nodes:

Control Nodes:

YAML Workflow Structure

name: my-workflow
description: What this workflow does
vars:
  api_key: "your-api-key"
steps:
  - node: fetch_url
    params:
      url: "https://api.example.com/data"
  - node: json_parse
    params:
      path: "result.items.[0].name"
  - node: file_write
    params:
      path: "output.txt"
  - node: notify
    condition: "{{.output}} != ''"
    params:
      message: "Done!"

Step Features

• condition: Go template expression, step runs only if true
• retry: Number of retries on failure
• delay: Delay between retries (e.g., "2s", "1m")
• parallel: Run multiple steps concurrently
• _timeout: Per-step timeout (e.g., "30s")

Workflow Generation Guidelines

When generating a workflow for the user:

1. Identify the steps — break the task into discrete operations
2. Choose the right nodes — prefer specific nodes over execute when possible
3. Chain with variables — use {{.steps[N].output}} to reference previous outputs
4. Add error handling — use condition and retry where appropriate
5. Keep it readable — add a description field, use meaningful step names

Common Patterns

Fetch and Save:

steps:
  - node: fetch_url
    params:
      url: "https://example.com/data"
  - node: file_write
    params:
      path: "data.txt"

Fetch, Parse, and Summarize:

steps:
  - node: fetch_url
    params:
      url: "https://example.com/article"
  - node: ollama
    params:
      model: "llama3"
      prompt: "Summarize: {{.steps[0].output}}"
  - node: file_write
    params:
      path: "summary.md"

Multi-source Aggregation:

steps:
  - parallel:
      - node: fetch_url
        params:
          url: "https://api1.example.com"
      - node: fetch_url
        params:
          url: "https://api2.example.com"
  - node: combine
    params:
      separator: "\n"
  - node: ollama
    params:
      model: "llama3"
      prompt: "Analyze: {{.steps[1].output}}"

Security Notes

• Built-in SSRF protection (URL validation, DNS rebinding checks)
• Path traversal protection (sandboxed paths, symlink resolution)
• Command injection prevention (shell metachar filtering, optional allowlist)
• Resource limits (file size, response body, step count, timeouts)
• Safe mode disables the execute node entirely

Installation

# Linux/macOS
curl -sL https://raw.githubusercontent.com/alib8b8/llm-box/main/install.sh -o install.sh
bash install.sh

# Or via Go
go install github.com/alib8b8/llm-box/cmd/llm-box@latest