Tasks

Tasks are single-turn LLM calls with prompt templates. They are the basic building blocks for AI-powered operations in runpiper.

What is a Task?

A task defines:

Which model to use (provider and model name)
System prompt - The AI’s role and instructions
User prompt template - How to structure the input
Optional settings - Temperature, max tokens, etc.

Task Configuration

Tasks are defined in TOML files with the following structure:

[task]
name = "summarize"

[model]
provider = "anthropic"
name = "claude-3-haiku"

[prompts]
system = "Summarize the given text concisely"
user = "{{input.text}}"

Required Fields

Field	Description	Example
`task.name`	Unique task name	`"summarize"`
`model.provider`	LLM provider	`"anthropic"`, `"openai"`
`model.name`	Model name	`"claude-3-haiku"`, `"gpt-4"`
`prompts.system`	System prompt	`"You are a helpful assistant"`
`prompts.user`	User prompt template	`"{{input.message}}"`

Optional Fields

Field	Description	Default
`model.temperature`	Sampling temperature (0-2)	Model default
`model.max_tokens`	Maximum response tokens	Model default
`model.top_p`	Nucleus sampling (0-1)	Model default

Input Templates

Use {{input.<field>}} to reference input data in your prompts:

[prompts]
system = "You are a {{input.role}} assistant"
user = """
Topic: {{input.topic}}
Tone: {{input.tone}}

Please write a response.
"""

When calling the task, pass input as JSON:

rp task test summarize --input '{
  "text": "Long text to summarize...",
  "role": "summarization",
  "tone": "professional"
}'

Supported Providers

Anthropic

[model]
provider = "anthropic"
name = "claude-3-haiku"

Supported models:

claude-3-haiku
claude-3-sonnet
claude-3-opus
claude-sonnet

OpenAI

[model]
provider = "openai"
name = "gpt-4"

Supported models:

gpt-4
gpt-4-turbo
gpt-3.5-turbo

OpenAI Compatible

[model]
provider = "openai-compatible"
name = "custom-model"
base_url = "https://api.custom.com/v1"

Task Examples

Text Summarization

[task]
name = "summarize"

[model]
provider = "anthropic"
name = "claude-3-haiku"

[prompts]
system = "Summarize the given text concisely in 2-3 sentences."
user = "{{input.text}}"

Data Extraction

[task]
name = "extract"

[model]
provider = "anthropic"
name = "claude-3-sonnet"

[prompts]
system = "Extract structured data from the text. Return JSON."
user = """
Text: {{input.text}}
Fields to extract: {{input.fields}}
"""

Translation

[task]
name = "translate"

[model]
provider = "openai"
name = "gpt-4"

[prompts]
system = "Translate the following text from {{input.from}} to {{input.to}}."
user = "{{input.text}}"

Managing Tasks

Create a Task

rp task init my-task

This creates a my-task.toml template file.

Validate a Task

rp task validate my-task.toml

Push a Task (Preview)

rp task push my-task.toml

This creates a preview version you can test.

Test a Task

rp task test my-task --input '{"message": "Hello"}'

Deploy a Task

rp task deploy my-task

Deploy the latest preview version to production.

List Tasks

rp task list                    # All tasks
rp task list --live            # Production only
rp task list --preview         # Preview only

Pull a Task

rp task pull my-task

Download the Taskfile for a deployed task.

Delete a Task

rp task delete my-task

Generate Test Input

rp task gen-test my-task

Generate a sample input file for testing.

Best Practices

Keep prompts focused - Single-purpose tasks work best
Use templates wisely - Make tasks reusable with input variables
Version your prompts - Deploy preview versions before production
Test thoroughly - Use rp task test before deploying
Choose the right model - Use faster models for simple tasks

Next Steps

Capabilities: Connect to external services
Agents: Build autonomous multi-turn agents
CLI Reference: Complete CLI command documentation