DEV Community

Fenju Fu
Fenju Fu

Posted on

A real Astron Agent dual-skill run for a landscape photo folder

#opensource #ai #agents #design

A workflow screenshot is only useful when it proves what actually ran. For this demo, a local orchestration step read a folder of nine landscape photos and built a BATCH_MANIFEST. Astron Agent then processed that manifest through a real three-node workflow and two attached skills.

Two skills attached to the Astron Agent node

The orchestration

The workflow is intentionally small and inspectable:

Start → Agent Decision → End
Enter fullscreen mode Exit fullscreen mode

The Agent node attaches two prompt-and-QA skills:

  • Paper Signal Art Director routes preservation-first landscape and architecture work.
  • GC Minimal Zine Poster routes sparse paper-poster treatments with controlled negative space.

Astron workflow canvas

The Agent was explicitly constrained to call each read_skill_* tool once with an empty JSON object, avoid run_skill_*, and return an auditable merged result. That matters because these skills provide art direction, prompt compilation, and QA rules; they are not bitmap-generation tools.

Verified result

The live debug run completed successfully across all three nodes in 87.184 seconds with 4,605 tokens. The result included file routing, protected details, generation prompts, QA gates, and prompt_ready render status.

Successful Astron debug run

A separate downstream image pass applied the prompts to all nine photos. Eight outputs passed visual review. One remained in human review because portrait-panel count and signboard details drifted. The workflow did not hide that failure or label it high fidelity.

Reuse

The published package contains the exported Astron DSL, both Astron-compatible SKILL.md files, and import notes. Skill IDs in an Astron export are environment-local, so importers should register the two skills and re-bind them in the Agent node when necessary.

Top comments (0)