Knowledge Publishing.
An inspectable, multi-stage workflow for preparing self-help content, with a person directing each step and reviewing the result.
Team contribution — output testing & feedback
From uneven drafts to useful knowledge.
HUMAN DIRECTION → CONSIDERED SYSTEMSMake each transformation inspectable, with a human choosing the next step and approving publication.
Useful knowledge needs consistent structure
Preparing self-help content requires consistent formatting, useful search metadata, and material that can be rendered as HTML. Those are related problems, but they are different transformations. Combining them in one broad prompt makes it harder to see which task caused an output problem.
Casey collaborated with a service-desk coworker on three specialized GPT workflows: standardizing guide formatting, extracting search metadata, and converting material into HTML for a ServiceNow catalog. The work is described here without reproducing internal guides or employer configuration.
A team contribution centered on testing
A service-desk coworker led the multi-stage design. Casey contributed alongside that work, running content through the stages, examining the output, and supplying feedback about formatting, metadata, and HTML. The direct account establishes collaboration, not sole ownership or project leadership.
The GPTs performed specialized transformations, while people directed the sequence and assessed the results. This is AI-assisted knowledge work with a clear human role. It is not an autonomous integration that publishes directly into ServiceNow.
Standardize, extract, generate, review
Each specialized GPT had a narrow role: standardize the source, extract metadata, or generate HTML. Stage-specific documentation expectations made the intermediate outputs easier to inspect and gave the next transformation a more predictable input.
Metadata had its own stage because discoverability depends on more than the body of an article. HTML generation was also separate from content preparation. A person selected the next action and reviewed the result before publication, so model-generated output did not acquire publication authority simply by completing a stage.
Keep intermediate output visible
Specialized stages trade manual handoff for inspectability. A formatting issue can be examined before metadata or HTML generation compounds it. The reviewer can identify the transformation responsible for a defect instead of evaluating one large response with several intertwined tasks.
The operator also retains control of the sequence. That limits autonomy and requires a person to decide what the document needs, but it keeps the workflow’s authority small. The design is a human-directed LLM workflow, not a state-aware agent that chooses its own tools.
Turn vague quality concerns into stage-specific feedback
Casey's documented testing identified breakdowns in formatting, metadata, and generated HTML and fed them back to the specialist leading the work. The findings identified concrete refinement targets for the prompts and stage boundaries. No private article examples or unsupported before-and-after claims are needed to explain the diagnostic method.
The useful question is where the output stopped meeting the expected contract. Checking each stage against the source gives the team a clearer correction target. Human review before publication supplies the change-control boundary; inspectable intermediate output supplies a way to detect problems before they propagate.
Applied workflow, human-reviewed output
Casey’s account confirms collaboration on the three specialized GPTs and their roles in preparing ServiceNow catalog material. The public project account also describes stage-output review and feedback to the coworker leading the work.
There is no supported claim of automatic ServiceNow publication, measured time savings, search improvement, or adoption. The concrete outcome is the human-directed staged workflow and Casey’s contribution to evaluating its outputs.
Understand who chooses the next action
Running the stages and examining their output gave Casey practical experience with task decomposition, structured handoffs, metadata extraction, and generated HTML review. Reporting failures at the stage where they occurred demonstrates a more precise testing practice than asking whether the overall AI response seemed good.
The workflow also exposed the boundary between a pipeline and an agent: a person was still the scheduler. Casey’s subsequent Publishing Agent design addresses that specific limitation by proposing state classification and approved tool selection while retaining human publication approval.