My content pipeline runs on its own every day. A local model drafts, a stronger model reviews, and nothing goes out without my approval. Two design rules keep it safe.
Rule one: stop loudly, never guess quietly
Every pipeline reads a config file, and the risky moment is when that file can't be read. An invisible character (a BOM, or byte order mark) is enough to break it. Code that falls back on default values keeps running on the wrong settings, and in a publishing system that can put the right video on the wrong account.
So my pipeline halts when its config is unreadable and sends me a warning. Stopping costs a minute. Guessing can cost an account.
Rule two: quality control is not fact-checking
A review step judges the writing: clarity, tone, structure. It can't tell whether a claim is true. An AI draft can read well and still get the facts backwards, for example saying a project's games are unplayable when four of them are public.
The fix is to separate the jobs. The writing step now works only from a short fact sheet that I verify and date, and it refuses to run when that sheet is stale. Review checks the style. The fact sheet checks the facts.
The stack
Text drafts run on Ollama with Qwen3 14B, and images come from Flux.1 through ComfyUI. Both run on my own PC, so each extra draft costs nothing. The reviewing model is a stronger cloud model.
Your turn
I'm documenting the project here as it develops. If you run automations of your own, what does yours do when a config breaks: stop, or carry on with defaults?
