Workflow Execution Infographic Prompt

Target

Embedded in docs/workflow_spec.md after the opening paragraph and before ## Table of Contents.

Intended Alt Text

Workflow execution model showing inputs, environment, five execution step types, prompt injection, per-step control modifiers, named outputs, artifacts, parallel branch joins, HITL gates, and finally cleanup.

Final GPT Image Prompt

Use case: infographic-diagram Asset type: documentation PNG for docs/images/workflow-execution-infographic.png in a technical Markdown manual.

Primary request: Create a revised 16:9 landscape architecture infographic for the SASE YAML workflow execution model, addressing these correctness requirements: five real step execution types are agent, prompt_part, bash, python, and parallel; hidden is a modifier badge, not a step type; HITL is a hitl: true approval gate attached after agent / bash / python steps, not a step type; prompt_part expands into a calling prompt at the #name(args) reference site; named outputs and artifact: stdout flow forward; parallel branches join and then feed a downstream next step; finally: cleanup runs even after failure or HITL rejection.

Style/medium: crisp software documentation architecture diagram, light neutral background, clean vector-like blocks, restrained but distinct accent colors, readable at GitHub Markdown width, no logos, no screenshots.

Composition/framing: 16:9, generous spacing, three zones. Left zone has two distinct source blocks titled Inputs and Environment. Center zone has an ordered workflow lane with five horizontal step cards in order: agent, prompt_part, bash, python, parallel. Put small per-step control brackets near example cards labeled if:, for:, while:, repeat/until; make the bracket scope visibly per-step. Attach a small hidden: true badge to one example card. Add a reusable use: import badge on one step. Place a HITL approval gate diamond on the downstream side of the agent / bash / python area with short labels Accept, Edit, Reject and an approved -> downstream hint. Right zone shows parallel fan-out into branches A, B, C, then a join panel with four modes and default notes, then a downstream next step tile consuming a named field. Bottom strip shows finally: cleanup steps with a dotted connector from failure/reject paths and the note always runs.

Text constraints: use only short labels from the workflow spec, including Inputs, typed params, validated at invocation, Environment, Jinja2 env vars, os.environ session, output: { field: type }, {{ step.field }}, artifact: stdout, {{ step._artifact }}, #name(args), calling prompt, expanded inline, hitl: true, Accept, Edit, Reject, approved, object default for parallel, array default for for, text opt-in, lastOf opt-in, {{ parallel_step.branch.field }}, finally:, cleanup, always runs, after failure or reject, and no nested for/while/repeat, parallel, or HITL.

Constraints: Do not show hidden or HITL as rows in the execution-type lane. Do not draw control wrappers around the whole workflow lane; show per-step brackets only. Do not make the parallel join the terminus; it must feed a next step. Keep labels large and legible, dark text on light blocks. Avoid dense paragraphs and avoid decorative gradients. No watermarks.

Post-Processing Notes

Generated with the built-in image generation tool on 2026-05-10, copied into docs/images/workflow-execution-infographic.png, then lightly post-processed with ImageMagick to correct generated text defects in the title and environment block while preserving the generated composition.

The regenerated image addresses the critique in docs/images/workflow-execution-infographic.critique.md: hidden is a badge, HITL appears as approval gates rather than a step row, control wrappers are scoped to example steps, prompt_part is shown expanding into a calling prompt, parallel join feeds a downstream next step, and a finally: cleanup strip is present.