Workflow Engineering
Design the path, not just the answer.
A prompt requests an output. A workflow creates a dependable result.
A workflow turns inputs into an accepted, reusable deliverable through observable stages. The point is not to automate everything. The point is to make the work, review, and recovery visible.
Six observable stages
- Define: Owner, outcome, audience, and acceptance criteria.
- Ground: Governing sources, versions, permissions, and data boundaries.
- Produce: One bounded transformation in a specified format.
- Test: Criteria, representative cases, and failure cases.
- Decide: Approve, revise, escalate, or stop.
- Record: Output, evidence, version, changes, errors, and next owner.
Preserve observable evidence.
Keep inputs, source references, outputs, tool actions, test results, and approvals. Do not treat a model's private reasoning as the audit trail.
Place review where consequences enter.
External communication, sensitive data, spending, commitments, system changes, and professional judgments need explicit decision rules and appropriate human authority.
Map one repeated task.
Mark deterministic steps, judgment steps, one likely failure, and the recovery path. Run the workflow manually before automating it.
Automation is an implementation choice.
The workflow comes first. A good workflow remains understandable even when tools change.
The real skill is turning good judgment into a process that can be inspected and improved.
Next: Agentic AISource checks: Anthropic, Building effective agents; NIST AI RMF Core.