Generate an ephemeral workflow tailored to a user's goal and immediately execute it, enabling single-delegation dynamic workflow orchestration

Keywords: workflow · dynamic · generation · execution · orchestration · ephemeral · single-delegation

Usage

To run this workflow, start your prompt with ~workflow-create-run followed by your goal, or select it with the /workflow command.

When to use

when a task needs a multi-step workflow but no existing workflow matches

when the manager agent needs to create and run a workflow in a single delegation

when the user describes a goal that would benefit from a custom pipeline but wants results, not a workflow artifact

When not to use

when an existing workflow already handles the task (use it directly)

when the user wants to keep the workflow as a reusable artifact (use workflow-creation-iterative instead)

when the task is simple enough for a single agent without orchestration

Configuration

Property	Value
Goal	What task or goal should be accomplished? Describe what you need done — a workflow will be created and executed to achieve it
Referenced agents	`workflow-creator`

Pipeline

This workflow solves the problem of needing to both create and execute a workflow in a single delegation. It uses the workflow-creator agent to generate an ephemeral child workflow under the fixed name workflow-create-run-temporary, validates it, and then executes it via workflow= composition.

The Create node instructs the workflow-creator agent to design a workflow appropriate to the goal — including agent review stages, human gates, iterative loops, and branching where the task warrants them. The whole point of generating a workflow (rather than just running a single agent) is to get multi-stage orchestration with quality checks. The agent stores the child workflow's goal into pipeline context via workflow_set_context so the Execute node can pass it through. A shell node validates the generated workflow, and on failure the validation error is stored and fed back to the creator for a retry. On success, the child workflow is executed, and finally cleaned up.

The creator agent has context-writable=true so it can store the child goal via workflow_set_context. The Execute node passes the creator-decided goal to the child workflow. The Cleanup shell node removes the ephemeral child directory after execution regardless of outcome.

The Create node uses persist="full" so the workflow-creator agent retains its workspace exploration, agent inventory, and draft across validation-failure retries, avoiding redundant file reads on each loop. A graph-wide max-session-turns default of 10 caps context growth.

`create-prompt`

Generate an ephemeral workflow tailored to this goal:

$goal

CRITICAL INSTRUCTIONS — read all of these before starting:

The workflow MUST be created with the exact name "workflow-create-run-temporary" as an ephemeral workflow in the workspace. This means:

Directory: .stencila/workflows/workflow-create-run-temporary/

File: .stencila/workflows/workflow-create-run-temporary/WORKFLOW.md

A .gitignore file containing exactly "*" in .stencila/workflows/workflow-create-run-temporary/

Design the workflow to match the complexity of the goal:

The whole point of a workflow is multi-stage orchestration — include agent review stages, human approval gates, iterative revision loops, and branching where the task warrants them

Do not optimize for reusability (this is a one-shot ephemeral workflow), but do design for quality — review and iteration stages are what make workflows valuable over single agents

For substantial tasks (coding, writing, design), include create-review-revise patterns

For simpler tasks, a shorter pipeline is fine — but it should still have at least two meaningful stages, otherwise a single agent would suffice and no workflow is needed

Use human gates when the goal involves consequential actions (deployment, publishing, committing) or when the user would reasonably want to approve intermediate results

Set the child workflow's goal in the frontmatter to a clear, actionable description of what the workflow should accomplish. Then store this same goal string into workflow context using workflow_set_context with key "child_goal" so the parent workflow can pass it through during execution.

Check for validation feedback from a previous attempt using workflow_get_output. If validation errors are present, fix the issues rather than starting from scratch. Also use workflow_get_context with key "validation_result" to see the specific validation output.

Design the pipeline to use available workspace agents where appropriate. Use list_agents to see what agents are available. Fall back to general-purpose agents for tasks that do not match a specialist.

Keep the workflow's description concise and specific to the goal.

This page was generated from .stencila/workflows/workflow-create-run/WORKFLOW.md.