prose-gen（散文生成）

Generate valid OpenProse code from the user's plain English description.

Your Task

1. Understand what the user wants to accomplish
2. Generate valid OpenProse code following the syntax below
3. Output the code in a fenced code block with prose language tag
4. Suggest a filename (e.g., workflow.prose)
5. Explain how to run it: prose run <filename>

OpenProse Syntax Reference

Comments

# This is a comment
session "Hello"  # Inline comment

Strings

"Single line string"
"""
Multi-line string
preserves whitespace
"""
"Hello {name}"  # Interpolation with {varname}

Agent Definitions

agent name:
  model: sonnet          # sonnet, opus, or haiku
  prompt: "System prompt"
  persist: true          # true, project, or path string
  skills: ["skill1"]
  permissions:
    read: ["*.md"]
    write: ["output/"]
    bash: deny           # allow, deny, or prompt
    network: allow

Session Statements

# Simple session
session "Do something"

# Session with agent
session: agentName
  prompt: "Override prompt"
  model: opus
  context: previousResult
  retry: 3
  backoff: exponential

Variables

let result = session "Get data"        # Mutable
const config = session "Get config"    # Immutable
result = session "Update"              # Reassign let only

# Context passing
session "Use previous"
  context: result                      # Single
  context: [a, b, c]                   # Multiple
  context: { a, b, c }                 # Object shorthand

Composition Blocks

# Sequential block
do:
  session "First"
  session "Second"

# Named block (reusable)
block review-pipeline:
  session "Review"
  session "Fix"

do review-pipeline  # Invoke

# Block with parameters
block process(item, mode):
  session "Process {item} in {mode} mode"

do process("data.csv", "strict")

# Inline sequence
session "A" -> session "B" -> session "C"

Parallel Blocks

parallel:
  a = session "Task A"
  b = session "Task B"

session "Combine"
  context: { a, b }

# Join strategies
parallel ("first"):      # Race - first wins
parallel ("any"):        # First success
parallel ("any", count: 2):  # Wait for 2

# Failure policies
parallel (on-fail: "continue"):   # Let all complete
parallel (on-fail: "ignore"):     # Ignore failures

Fixed Loops

# Repeat N times
repeat 3:
  session "Generate idea"

repeat 5 as i:
  session "Process item {i}"

# For-each
for item in items:
  session "Process"
    context: item

for item, i in items:
  session "Process {i}"
    context: item

# Parallel for-each (fan-out)
parallel for topic in ["AI", "ML", "DL"]:
  session "Research"
    context: topic

Unbounded Loops

Use **...** discretion markers for AI-evaluated conditions:

loop (max: 50):
  session "Process next"

loop until **the task is complete** (max: 10):
  session "Continue working"

loop while **there are items to process** (max: 20) as i:
  session "Process item {i}"

Pipeline Operations

let items = ["a", "b", "c"]

# Map
let results = items | map:
  session "Transform"
    context: item

# Filter
let filtered = items | filter:
  session "Keep this? yes/no"
    context: item

# Reduce
let combined = items | reduce(acc, item):
  session "Combine"
    context: [acc, item]

# Parallel map
let fast = items | pmap:
  session "Process in parallel"
    context: item

# Chaining
let final = items
  | filter:
      session "Keep?"
        context: item
  | map:
      session "Transform"
        context: item

Error Handling

try:
  session "Risky operation"
catch as err:
  session "Handle error"
    context: err
finally:
  session "Always cleanup"

# Throw
throw "Something went wrong"
throw  # Re-raise in catch block

# Retry
session "Flaky API"
  retry: 3
  backoff: exponential  # none, linear, exponential

Choice Blocks

choice **which approach is best**:
  option "Quick fix":
    session "Apply quick fix"
  option "Full refactor":
    session "Do full refactor"

Conditionals

if **code has security issues**:
  session "Fix security"
elif **code has performance issues**:
  session "Optimize"
else:
  session "Proceed"

Program Composition

# Import programs
use "@handle/slug"
use "@handle/slug" as alias

# Inputs and outputs
input topic: "The subject to research"

let result = session "Research {topic}"
output findings = session "Synthesize"
  context: result

# Call imported program
let data = research(topic: "quantum computing")
session "Use findings"
  context: data.findings

Persistent Agents

agent captain:
  model: opus
  persist: true           # Dies with execution
  persist: project        # Survives across runs
  persist: ".prose/custom/"  # Custom path

# First call creates memory
session: captain
  prompt: "Review plan"

# Resume continues with memory
resume: captain
  prompt: "Continue review"

Example Patterns

Research Pipeline

agent researcher:
  model: sonnet
  prompt: "You research topics thoroughly"

agent writer:
  model: opus
  prompt: "You write clear documentation"

let research = session: researcher
  prompt: "Research quantum computing"

session: writer
  prompt: "Write summary"
  context: research

Parallel Review

parallel:
  security = session "Security review"
  perf = session "Performance review"
  style = session "Style review"

session "Synthesize reviews"
  context: { security, perf, style }

Iterative Improvement

let draft = session "Write initial draft"

loop until **draft is polished** (max: 5):
  draft = session "Improve draft"
    context: draft

session "Finalize"
  context: draft

Output Format

Always output:

1. The .prose code in a fenced code block
2. Suggested filename
3. How to run: prose run <filename>

Ask clarifying questions if the request is ambiguous.