ask ... using

ask … using

Send a task to a reasoning provider (LLM, symbolic planner, solver). The primary step type for AI-powered reasoning within a machine. Every ask step is governed: the runtime checks permissions, enforces token budgets, records the decision in the behavioral ledger, and mediates the call through the governance interpreter.

When to use

Use ask when you need:

Natural language understanding or generation
Classification, extraction, or summarization
Decision-making that requires judgment
Structured output from unstructured input

Use compute for deterministic calculations. Use call to invoke another machine. Use decide for rule-based branching.

Three variants

Variant	Syntax	Purpose
`ask ... using`	`ask classify, using: "anthropic:claude-sonnet-4-6"`	Send a task to an LLM provider
`ask ... from`	`ask data, from: "@mashin/actions/http/get"`	Request data from an effect machine
`ask ... of`	`ask status, of: "monitor"`	Query a running machine’s state

This page covers ask ... using. See the linked pages for the other variants.

Syntax

ask <name>, using: "<provider>:<model>"
  with task "<instruction>"
  with role "<system prompt>"
  returns
    <field> as <type>
  assuming
    <field>: <mock value>

Configuration

Config	Required	Description
`using`	Yes	Provider and model: `"anthropic:claude-sonnet-4-6"`, `"openai:gpt-4.1"`, `"ollama:llama3"`
`with task`	Yes	The instruction sent to the model. Supports `input.` and `steps.` interpolation.
`with role`	No	System prompt. Sets the model’s persona or constraints.
`returns`	No	Structured output schema. Fields the model must return.
`assuming`	No	Mock values for testing. Used in test/simulate mode instead of calling the real model.
`with tools`	No	List of tool names the model can call during reasoning.
`with temperature`	No	Sampling temperature (0.0 to 2.0). Lower = more deterministic.
`with max_tokens`	No	Maximum tokens in the response.

Examples

Simple classification

machine sentiment
  accepts
    text as text, is required

  responds with
    sentiment as text
    confidence as number

  implements
    ask classify, using: "anthropic:claude-sonnet-4-6"
      with task "Classify the sentiment of this text as positive, negative, or neutral. Return a confidence score between 0 and 1."
      with role "You are a sentiment analysis expert."
      returns
        sentiment as text
        confidence as number
      assuming
        sentiment: "positive"
        confidence: 0.95

Using input interpolation

ask summarize, using: "anthropic:claude-sonnet-4-6"
  with task "Summarize this document in 3 bullet points: ${input.document}"
  returns
    bullets as list

Chaining with compute

machine email_triage
  accepts
    subject as text, is required
    body as text, is required

  implements
    ask analyze, using: "anthropic:claude-sonnet-4-6"
      with task "Analyze this email. Determine priority (high/medium/low), category, and whether it needs a response."
      with role "You are an executive assistant triaging emails."
      returns
        priority as text
        category as text
        needs_response as boolean
        suggested_action as text
      assuming
        priority: "medium"
        category: "general"
        needs_response: true
        suggested_action: "Review and respond within 24 hours"

    compute format_response
      {
        priority: steps.analyze.priority,
        category: steps.analyze.category,
        action: steps.analyze.needs_response
          ? "Respond: " + steps.analyze.suggested_action
          : "No response needed",
        triaged_at: now()
      }

With tool use

ask research, using: "anthropic:claude-sonnet-4-6"
  with task "Research this company and provide a brief summary"
  with tools ["web_search", "read_url"]
  returns
    summary as text
    sources as list

Providers

Supported provider prefixes:

Provider	Prefix	Example Models
Anthropic	`anthropic:`	`claude-sonnet-4-6`, `claude-haiku-4-5`
OpenAI	`openai:`	`gpt-4.1`, `gpt-4.1-mini`
Google	`google:`	`gemini-2.5-pro`, `gemini-2.5-flash`
Ollama (local)	`ollama:`	`llama3`, `mistral`, `codellama`
Groq	`groq:`	`llama-3.3-70b`

The model is resolved at runtime through cell settings. If using is omitted, the cell’s default model is used (configured in Cell.Settings).

Governance

Every ask step is governed:

Permission check: the machine must have reason capability
Token budget: checked before the call; denied if budget would be exceeded
Consent: in interactive mode, the user is asked to approve the LLM call
Behavioral ledger: the call, its cost, tokens used, and model response are recorded
Cost tracking: input/output tokens and estimated cost are tracked per execution

In test mode, assuming values are returned instead of calling the real model. This makes tests fast, deterministic, and free.

Testing with assuming

The assuming block provides mock return values for test and simulate modes:

ask classify, using: "anthropic:claude-sonnet-4-6"
  with task "Classify this input"
  returns
    category as text
    confidence as number
  assuming
    category: "technology"
    confidence: 0.92

When the machine runs in test mode, assuming values are returned immediately without calling the model. This lets you:

Write deterministic tests
Run CI without API keys
Verify downstream logic without LLM variability

Translations

Language	Keyword
English	ask
Spanish	pregunta
French	demande
German	frage
Japanese	質問
Chinese	问
Korean	질문