How The Philadelphia Inquirer built Scrape, an AI tool to surface hyperlocal news leads

The problem 

The Philadelphia Inquirer publishes hyperlocal newsletters covering specific communities across the Philadelphia area. The team needed a way to reliably surface newsworthy stories from each community from a fragmented ecosystem of sources, such as municipal meetings, school calendars, and Facebook pages. 

One of their editors was spending about 12 hours a week just finding items for newsletters, leaving little time for writing and shaping the voice of the publication. 

Gathering news for these regions was hard because it comes from “a lot of really small sources, and so it’s a lot to sift through,” said Kevin Hoffman, The Lenfest Institute’s AI Fellow who is embedded at The Inquirer.

The solution 

The Inquirer’s AI tool, Scrape, was created to save time and raise the signal-to-noise ratio of that work.

Scrape worked because it was built around The Inquirer’s workflows, with tight feedback loops from the editor, rather than as a purely technical experiment.

First, the team grounded the tool in the existing manual process. The lead editor had already compiled a curated list of sources and was producing daily tip sheets by hand. Kevin turned that into a direct AI training and evaluation loop.

Their hypothesis was that AI could be used to find newsworthy information. To test it, the team would “run AI every day over the sources that this editor cared about, produce a tip sheet, and then also compare this tip sheet to what this editor produced and understand, ‘Okay, is it actually extracting the same insights? Is it maybe finding new insights that the editor missed, or is it or is it missing in some of those areas?’” Kevin said. 

“And that is exactly what we did over the span of three or four months. It was pretty intense. Every other day, our editor would give me feedback on how Scrape was doing.”

The team would run Scrape and produce an e-mail digest that the editor would annotate, pointing out what was useful and what was not. 

This iterative prompt refinement, where bullet points were added and removed and newsworthiness was defined, gradually shifted the output from noise (such as rescheduled meetings and routine road closures) to genuinely impactful community news, Kevin said.

The second major success factor was embedding editorial context directly into the system. The team used newsletter-specific descriptions, such as the newsletter’s audience, relevant geographical regions, school districts and topics, rather than simply using geography as a guideline for hyperlocal news. In this way, Scrape’s outputs were aligned with what newsletter readers actually cared about.

“For instance, Chester County is one of our newsletters. The editor for the project composed a description for this newsletter that included information on what they were looking for – readers are interested in everything from municipal and local governments to localized state news relevant to them, especially things that will impact their lives, like restaurants and retail businesses,” Kevin said. The prompts explicitly state what not to include, such as straight business news on topics like the stock exchange or personnel changes. They also include information about who the readers are. 

“They are trying to help the models understand what these readers might care about, where their perspectives lie, the relevant communities, and also communities that we actually would not want to include in this newsletter, as well as the relevant school districts,” he said. “We’re passing all this information into Scrape now.”

Editors are pleased with the change, Kevin said. They are now refining the descriptions to produce more actionable tips.

The result: Scrape evolved from a small experiment to “load-bearing and critical infrastructure” for six (and soon, eight) newsletters, with nearly 20 journalists across the newsroom following its output.  The editors’ skills and audience demand has made the local newsletters a success. Scrape is helping the Local team meet demand without overworking the desk. 

What we’d do differently 

One of the hardest problems was generalizing from one editor and geography to many. The original design philosophy was region-based: scrape everything about, say, Lower Merion, then filter using a single “newsworthiness” prompt. That worked “manageably” for one editor and a few newsletters, but did not scale, Kevin said.

Instead of treating “newsworthiness” as a generic, abstract concept, they now anchor it in specific newsletter briefs that editors author themselves – who the readers are, what they’re interested in, and where the coverage line stops.

Another lesson involves cost and search strategy. Early on, the team tried to balance quality with the high cost of searching many small, scattered sources. They discovered that prompting can implicitly control search depth and behavior, but only after trial and error.

In hindsight, Kevin suggests being more explicit, earlier, about acceptable cost-quality trade-offs and designing prompts and architecture around those constraints from day one. 

Here’s how you can apply this 

Scrape’s journey offers a practical blueprint for newsrooms and organizations trying to use AI for finding news in messy local data.

• Start with a real workflow, not an abstract idea. Identify a specific person or team who is already doing the work manually and measure their effort (e.g., “12 hours a week just to find news”).
• Turn existing outputs into training data. Use their current products (tip sheets, newsletters, etc.) as both a target and a benchmark.
• Create a tight human-in-the-loop feedback loop. Ask editors to annotate outputs daily: what’s good, what’s noise, and what should never appear (examples from this case include routine road closures or recurring fitness-class promotions).
• Encode editorial judgment explicitly. Replace vague notions of “newsworthiness” with concrete prompts and newsletter descriptions: audience, geography in and out of scope, recurring beats, and examples of past coverage.
• Design for cost and quality together. Decide up front how much search depth you can afford, and use prompting and architecture to stay within that envelope.

The Lenfest AI Collaborative and Fellowship Program is supported by OpenAI and Microsoft.