Starting constraint
The editorial team wanted faster summaries, structured metadata, draft briefs, and internal content transformations. They did not want autonomous publishing.
The main risks were data leakage, inconsistent prompts, token cost drift, hallucinated facts, and editors losing trust in the workflow.
Architecture decision
WordPress stayed the editorial system of record. Claude API was used as an assistant inside bounded tasks: summarise, classify, extract, rewrite for a specific format, and propose metadata.
Every AI output had an explicit source, a visible review step, and a human approval gate before it could affect public content.
Governance model
Prompts were versioned, named, and scoped by task. Editors did not paste arbitrary private context into a chat box. The integration passed only the fields needed for the current operation.
Cost ceilings were handled per workflow: maximum input size, maximum output length, retry limits, and logging of token usage by task type.
Reusable lesson
The valuable output was not 'AI content'. It was a repeatable editorial machine with guardrails: structured prompts, source links, review states, cost visibility, and clear ownership.
This is the pattern I would reuse before adding MCP or agent-facing surfaces. Internal governance comes before external automation.