How we get from 300+ NYC stories a day down to five.

The promise

The homepage says NYC Daily TL;DR is "all signal, no noise." This page is the receipt — the actual pipeline that produces every weekday's briefing, the actual sources we read, the actual filters that weed out the rest, and the actual editorial principles that decide what makes the cut.

If you can't see the methodology, you can't trust the output. So here it is.

The funnel

Every weekday morning the pipeline runs and produces a single email. The stages, in order:

1. Ingest — 67 monitored feeds 59 article feeds + 8 podcast feeds, including the NYC.gov press wire and newsletters from named NYC beat reporters. Podcast entries stay labeled as episode descriptions or show notes; the pipeline doesn't consume audio or treat them as transcripts. Truncated article feeds can be enriched through the article scraper.
2. Pre-filter — language, length, date, dedup English-only, minimum article length, published within the last week, and de-duplicated against recent issues. A story we already covered won't run again unless there's genuine news since. Firehose feeds (silive.com, jta.org, news10, pix11) require an NYC keyword in the title before they advance.
3. Categorize + relevance — LLM pass #1 An LLM reads each article and decides: relevant for an NYC resident? If yes, which of seven categories — Politics & Policy, Housing/Transportation/Infrastructure, Culture & Lifestyle, Business & Economy, Civic Services & Government, Civic Engagement, Events. Tourist content, real-estate ads, and clickbait get cut here.
4. Score — LLM pass #2 Each surviving article gets scored on seven dimensions — significance, NYC-specificity, novelty, evergreen value, source quality, narrative arc, and reader actionability. Scoring is batched, cached in a Cloudflare D1 database, and skips known-low-quality domains entirely.
5. Human editorial — the final cut The top-scored candidates land in front of a human editor (the operator, Farzad). The editor picks the five that actually go in the briefing, writes the framing, and ships. The pipeline never decides what gets sent — it only decides what gets considered.
6. Compose + send HTML rendered, sent via Buttondown to subscribers, archived to /archives. An operational digest goes to the editor's inbox alone (a Buttondown filter restricts it to a single email address by construction — there's no scenario in which it accidentally goes to the list).

Roughly 300+ articles enter the funnel each weekday. About a third clear pre-filter. About half of those clear the LLM relevance pass. The scored shortlist usually runs 20–40. The editor picks five.

The sources

The 67 monitored feeds — 59 article and 8 podcast — are grouped roughly as follows. We update the roster regularly; the current list lives in the open-source pipeline config (config/feeds.py).

What this list deliberately leaves out: tourism content, partisan-aggregator sites with no original reporting, "best restaurants in NYC" listicles, and national stories with no specific consequence for New Yorkers.

What gets cut

The seven scoring dimensions exist to push back against the failure modes that produce noisy newsletters:

• Significance — does this affect actual New Yorkers' lives, money, mobility, or representation? "Mayor said something" is not significance. "City Hall is changing how it does X" is.
• NYC-specificity — irreducibly NYC, or generic urban-issue content with NYC mentioned in passing? Generic gets cut.
• Novelty — is this new information, or a recap of yesterday's news? Recaps get cut.
• Evergreen value — will this matter next week, or only today? Both can earn space, but pure-news cycle items have a higher bar.
• Source quality — original reporting from named sources, or a press release wearing a journalism costume? Press releases get cut unless they're the news themselves.
• Narrative arc — does this fit a thread the briefing has been tracking (rent freeze, MTA capital plan, mayoral control sunset), or is it disconnected? Threaded items earn space.
• Reader actionability — can a reader do something with this (vote, attend a hearing, change a habit, file a complaint), or is it pure observation? Both are valid; the bar shifts based on the slot.

The hardest cuts are the ones that score well on most dimensions but fail one. A well-reported Mayor-press-conference recap can score high on source quality and narrative arc but fails NYC-specificity if the substance is national. A breaking neighborhood story can score high on novelty and significance but fails source quality if it's secondhand. Both get cut.

What also gets cut, on principle:

• Tourist-coded content. "Things to do in NYC this weekend." Not the audience.
• Best-of listicles without genuine curation. Filler.
• Crime stories that exist mainly for shock. Real public-safety policy is in. Sensational individual incidents without policy or pattern context are out.
• Celebrity sightings. Not a thing this newsletter covers.
• Real-estate sales-funnel content. The 80-photo new-development announcement masquerading as news. Out.
• National political news that happens to involve a NYC name. The bar is whether NYC residents need to know it as NYC residents.

What we get wrong

Some of this pipeline is automated and that means it has failure modes worth being honest about.

The relevance and scoring passes use an LLM. LLMs miss things — sometimes a small but important City Council vote scores low because it doesn't read as significant in isolation. The human editorial pass is the backstop, but the editor only sees what the pipeline surfaces; if the pipeline drops something, the editor never sees it. We mitigate this by (a) keeping the pre-filter cuts narrow and the relevance bar permissive, (b) reviewing dropped-article logs in the operational digest, and (c) periodically auditing what got cut and why.

The 67-feed roster is a curation choice. It's biased toward sources we trust, which means it's underweighted on sources we haven't heard of yet. We add sources when they prove themselves; we drop sources when they degrade. The current list is in the public repo and changes are visible in git history.

The scoring weights and the prompt language are also choices. They're tuned for our reader — the plugged-in NYC resident who reads Gothamist and THE CITY and wants the synthesis. They are not tuned for the city's institutional power-broker, the academic researcher, or the journalist on a beat. Those readers should subscribe to other things in addition.

What we promise

• Daily, weekday mornings. Five stories. Under five minutes to read. In your inbox before 9am Eastern.
• Free. Email only. No paywall, no upsell, no premium tier.
• One human accountable. The byline on the long-form explainers and the editorial selection on the daily briefing are both Farzad Khosravi. That's the trust.
• Open methodology. This page exists. The 67-feed roster is in the public repo. The scoring framework is described above. If we change something material, we update this page.
• Honest framing. The pipeline is LLM-assisted; the editorial decisions are human. We don't pretend the AI doesn't exist, and we don't pretend it makes the calls. Both are true.
• Cross-pillar consistency. When a long-form explainer corrects a number or a framing, every page that referenced the old version gets updated in the same commit.

How we change

The pipeline gets better when readers tell us what's missing. Three useful signals:

• Tell us a source we should be reading — reply to any briefing email with the URL. Especially valuable: NYC-beat substacks, neighborhood blogs, immigrant-press outlets in any language, and policy-shop newsletters.
• Tell us what we missed — if a story you think should have been in the briefing wasn't, that's the highest-signal feedback we get. It either means the pipeline missed it (a real bug to fix) or it means the editorial bar disagreed (a calibration question worth surfacing).
• Tell us what shouldn't have been there — if something in the briefing felt like filler, we want to know. Cuts get easier when readers vote against the marginal items.

How to subscribe

To learn more: about NYC Daily TL;DR, all explainers, the archive of past briefings.

Last updated: July 13, 2026. The pipeline architecture (4 phases, 7 scoring dimensions, 67 monitored feeds: 59 article and 8 podcast) is stable; the feed list and scoring weights are tuned regularly and the changes are visible in the open-source repo. Material changes to this methodology page will be dated.