Skip to content

Template Variables

Reference for {{inputs.X}} and {{steps.Y.output}} template syntax.

Overview

Templates are placeholders in your workflow that get replaced with actual values at runtime:

prompt: "Analyze {{inputs.files}} and compare to {{steps.baseline.output}}"
#              ↑                           ↑
#        Input template              Step output template

Input Templates

Reference workflow inputs defined in the inputs section.

Syntax

{{inputs.INPUT_NAME}}

Example

inputs:
  files:
    description: "Files to analyze"
    default: "*.go"
  mode:
    description: "Analysis mode"

steps:
  - id: analyze
    prompt: "Analyze {{inputs.files}} in {{inputs.mode}} mode"

Run: 

goflow run -w example.yaml --inputs files='src/*.go' --inputs mode='detailed'

Resolved prompt: 

Analyze src/*.go in detailed mode

Step Output Templates

Reference the output from a previously completed step.

Syntax

{{steps.STEP_ID.output}}

Example

steps:
  - id: analyze
    prompt: "List the main functions in the code."

  - id: review
    prompt: |
      Review these functions:

      {{steps.analyze.output}}
    depends_on: [analyze]

What happens: 1. analyze runs and produces output (e.g., "Found: main(), init(), handleRequest()") 2. review runs with the prompt: "Review these functions:\n\nFound: main(), init(), handleRequest()"

Template Locations

Templates work in these locations:

Location Example
Step prompts prompt: "Review {{inputs.files}}"
Agent inline prompts prompt: "You specialize in {{inputs.language}}"
Condition values contains: "{{inputs.keyword}}"

Examples

Step prompt: 

steps:
  - id: review
    prompt: "Review {{inputs.files}}: {{steps.analyze.output}}"

Inline agent prompt: 

agents:
  specialist:
    inline:
      prompt: "You are an expert in {{inputs.language}} development."

Condition: 

condition:
  step: classify
  contains: "{{inputs.expected_type}}"

Resolution Rules

Order of Resolution

  1. Input templates are resolved first
  2. Step output templates are resolved based on dependency order
  3. Templates are resolved before sending to the AI

Missing Values

Scenario Behavior
Missing input (no default) Error: "missing required input: X"
Missing input (has default) Uses default value
Missing step output Empty string (step was skipped or not yet run)
Invalid template syntax Kept as literal text

Whitespace

Templates preserve surrounding whitespace:

prompt: |
  Before:
  {{steps.analyze.output}}
  After:

If analyze.output is "Hello", the resolved prompt is: 

Before:
Hello
After:

Multi-line Outputs

Step outputs often contain multiple lines. Use YAML multi-line syntax:

prompt: |
  Review this analysis:

  {{steps.analyze.output}}

  Focus on the key findings.

The | preserves newlines in the prompt and in the injected output.

Multiple Step References

Reference multiple steps in one prompt:

- id: summary
  prompt: |
    Combine these reviews:

    ## Security Review
    {{steps.security.output}}

    ## Performance Review
    {{steps.performance.output}}

    ## Accessibility Review
    {{steps.accessibility.output}}
  depends_on: [security, performance, accessibility]

Truncation

Large step outputs can exceed model context limits. Configure truncation:

Global Truncation

config:
  truncate:
    strategy: "lines"
    limit: 100

Output-Specific Truncation

output:
  steps: [summary]
  truncate:
    strategy: "chars"
    limit: 5000

Strategies

Strategy Description
lines Keep last N lines
chars Keep last N characters

Example: If steps.analyze.output is 500 lines and truncation is lines: 100, only the last 100 lines are injected.

Handling Skipped Steps

When a step is skipped due to a condition, its output is empty:

steps:
  - id: deep-dive
    depends_on: [initial]
    condition:
      step: initial
      contains: "CRITICAL"

  - id: summary
    prompt: |
      Initial review: {{steps.initial.output}}
      Deep dive (if performed): {{steps.deep-dive.output}}
    depends_on: [initial, deep-dive]

If deep-dive is skipped: 

Initial review: [actual content]
Deep dive (if performed):

Tip: Structure prompts to handle empty values gracefully: 

prompt: |
  Review findings: {{steps.review.output}}
  Additional findings (if any): {{steps.optional-review.output}}

Error Handling

Invalid Template Syntax

prompt: "Hello {{invalid syntax}}"  # Missing dots

Invalid templates are kept as literal text: 

Hello {{invalid syntax}}

Circular References

goflow's dependency system prevents circular references:

# ✗ Error: Circular dependency
- id: step-a
  prompt: "{{steps.step-b.output}}"
  depends_on: [step-b]

- id: step-b
  prompt: "{{steps.step-a.output}}"
  depends_on: [step-a]

Error: 

circular dependency detected in workflow DAG

Reference Before Completion

You can only reference a step if it's in your depends_on:

# ✗ Bad: no dependency declared
- id: step-a
  prompt: "Task A"

- id: step-b
  prompt: "Use {{steps.step-a.output}}"  # No depends_on!

The template might reference an incomplete or empty output. Always declare dependencies:

# ✓ Good: dependency declared
- id: step-b
  prompt: "Use {{steps.step-a.output}}"
  depends_on: [step-a]

Best Practices

1. Use Clear Labels

prompt: |
  ## Security Analysis
  {{steps.security-review.output}}

  ## Performance Analysis
  {{steps.perf-review.output}}

2. Handle Optional Content

prompt: |
  Core findings: {{steps.main-review.output}}

  Extended analysis (if available):
  {{steps.extended-review.output}}

3. Keep Templates Simple

# ✓ Good: Simple references
prompt: "Review: {{steps.analyze.output}}"

# ✗ Avoid: Complex expressions (not supported)
prompt: "Review: {{steps.analyze.output | uppercase}}"

4. Document Inputs

inputs:
  files:
    description: "Files to analyze (glob pattern, e.g., 'src/**/*.go')"
    default: "*.go"

See Also