The Pipeline That Reads the News So You Don't Have To
Most AI content tools solve the wrong problem. They make it faster to write. They don't make it easier to have something worth saying.
The problem underneath is knowing what's actually happening in your industry, every day, without spending an hour reading news that mostly doesn't matter. Most marketing teams skip this entirely — and content starts to feel disconnected from what's real.
This is the pipeline I built to solve that.
The signal layer: Google Alerts doing the filtering work
The pipeline starts with eight Google Alerts, each targeting a specific kind of signal. An alert for "AI marketing" returns thousands of vendor press releases. These return actual research:
These surface pain-signal language — problems described in the reader's own words:
Same topic, completely different quality of signal. All eight deliver to a single Gmail label. The inbox never sees them.
The extraction layer: turning email into content
An automation scenario runs daily, triggered by the Gmail label. It extracts every article URL from the email HTML, strips Google's redirect links, then fetches the full text of each article.
This is the step most people skip. Summarizing a headline is easy. Summarizing the actual article — the specific finding, the named source, the concrete example — is what produces content that's credible rather than vague.
The fetched content goes into Claude alongside active campaign context: who the audience is, what content pillars apply, and what's been overused in the last three posts.
The selection layer: disqualifying before generating
Before Claude writes anything, it runs a disqualification check. Press releases, vendor blog posts, sponsored listicles, and tool reviews get discarded. The prompt names specific URL patterns to reject outright — prnewswire, businesswire, globenewswire — and content types to discard on sight.
Google Alerts are not curated. On any given day, the majority of results are content marketing written to appear like journalism. The disqualification step is what separates signal from noise disguised as signal.
What remains gets ranked by priority: marketing and AI operations first, product and founder stories second, engineering content only as a last resort.
The generation layer: one signal, one implication, one post
Claude doesn't summarize the article. It extracts one verifiable fact and builds a post around a single implication. The prompt looks roughly like:
The rotation rules exist because a reader who sees the same pattern three days running has learned to predict it. Prediction kills engagement before the first line is read.
The format layer: same signal, different shape per platform
The same signal goes into two separate Claude calls — one for LinkedIn, one for X:
LinkedIn:
X/Thread:
The feedback layer
Post stats come back weekly. Posts that name specific tools outperform posts that describe categories. Posts built on a single concrete finding outperform broad observations. The pipeline adjusts — rotation rules update, overused angles get flagged, next week's generation starts from a cleaner baseline.
What the pipeline can't do is tell you whether you have something worth saying. The signal layer finds what's happening. The generation layer shapes it into a perspective. The part in the middle — deciding what actually matters — remains a human judgment, made once per day, taking about five minutes.
That's not a limitation of the pipeline. It's the point of it.
How are you handling the signal problem on your team — separating what's worth knowing from the noise? Genuinely curious what's working.
