Autonomous content engine

A one-person consultancy can’t publish consistently, and everybody in that position knows it. The work wins, the writing slips, and the site quietly goes stale.

A pipeline that proposes topics against a defined brief, drafts to a tone of voice held in the database rather than in my head, generates the imagery, publishes to WordPress, then audits its own output and amends it. It watches performance per article through analytics and search data and feeds that back into what it writes next.

Around it, a cockpit inside WordPress that shows every client on the platform, run health, drafts waiting for review and the analytics, with a run-now button for each tool. And a small kit that makes the site legible to the answer engines: a machine-readable site summary that rebuilds itself, an instant ping to the search indexes on every publish, and the structured data the audit said was missing.

Because the interesting engineering is in the failure modes, not the generation. Deterministic gates so a half-finished draft cannot publish. Sentinel-wrapped output so parsing cannot drift. A sanitiser between the model and the live site.

And the one that got past me. For a few weeks in the summer, deduplication lived in the prompt, which meant it lived in the model’s memory of what it had already covered, which is to say nowhere. One newspaper article was drafted twelve times and published nine. I deleted eight of them by hand, plus five LinkedIn shares. It took a while. Dedup now happens in code, against the database, before the model gets a say.