The morning report: what happened in your field, with evidence

Feeds deliver more than anyone can read. The problem is not new, and the usual answer is a list: the day's articles, sorted by relevance, each with a snippet. A list tells you what was published. It does not tell you what it adds up to.

The morning report tries the second. It is written the way a colleague would tell you in the morning what happened — with the difference that every sentence says where it comes from. Pith is a reading memory for consultants: it reads your feeds, writes you a report every morning in which every sentence cites its source, and takes into your wiki only what you choose.

What happens every morning at six

At 6:00 German time, Pith looks at the articles that appeared in your feeds over the last 24 hours and scores them against your interest profile. The most relevant, up to twelve, make the selection. On top come the hits of your watch list: terms whose articles always appear, however the model rates their relevance. They have their own block and take nothing away from the twelve places.

For each article Pith fetches the text and summarises it — from the article itself, not from the feed's snippet. If a text cannot be fetched, for example because the publisher has opted out of text mining, the item says so and the summary relies on the snippet. The articles are then grouped by theme.

The report is written from these summaries. It is generated once and stored, so it stays the same through the day. The report itself is written in German for now — it was built for IT consultants in the German-speaking region — while each article's summary stays in the article's own language.

How the report is structured

At the top is the situation: two or three sentences on what matters today and how the most important items connect. Below it, three to five key points, one sentence each, most important first.

Then comes a section per theme. The first sentence names what ties the items together — a trend, a contradiction, a deadline. The following sentences back it up with the items. Below each section the sources appear as compact rows, and if your wiki already has pages on the subject, the section names up to three of them.

Each section has a for you line: whom or what this concretely affects in consulting work. For example: "This mostly affects manufacturers with their own brand." Or: "Anyone running coding agents with write access has the gap in their own repository." That is an observation, not a recommendation. The report does not say "check", "you should" or "it's worth it". If there is nothing concrete, the line stays empty.

The same restraint applies everywhere else: no rhetorical questions, no judgements without content such as "fascinating" or "gaining importance", no sentences about the report itself. News first, short sentences, verbs over noun chains.

Every sentence cited, and that is checked

Every statement in the report carries, at the end of the sentence, the numbers of the items it rests on: [3] or [2, 5]. The model is instructed not to add numbers, names or dates that are not in the items.

Pith does not rely on that alone. After writing, the citations are checked: numbers that do not exist that day are removed. A key point or section left without a valid citation is dropped. Wiki pages the report names are matched against your wiki; unknown names disappear. So the report cannot name a source or a page that does not exist.

New publishers, without crowding out your own feeds

A report that reads only your subscribed feeds mostly confirms what you already chose. A publisher you do not know yet never gets in.

That is what search feeds are for. You can add a Google Alert as an RSS feed; Google Alerts has no API, so you create it yourself. In addition, Pith can keep Bing News searches for your watch terms and your strongest interests, at most six, if you switch that on. For both, Pith unwraps the redirect links so the actual article can be fetched, and records the publisher.

Search results fit the profile by construction — so well that in one real run they would have taken eleven of twelve places. So for articles from publishers you do not follow, exactly two places are reserved and capped, provided they are relevant enough. They are marked in the report as a new source. Press-release portals never get one of these places.

One story, several publishers

Search feeds bring the same news from many sites, and press portals often republish one announcement several times. Left alone, the same story takes several places and repeats itself in the report.

So before the selection, the day's articles are grouped by story. Similar titles and snippets produce candidate pairs, and a model checks whether two articles report the same event, not merely the same topic. A group becomes one item: preferably from a publisher you follow, a press portal only when there is nothing else. The others appear beneath it as "also reported by". A decision on that item applies to every article in the group.

Nothing enters the wiki without your decision

Below each section the articles are listed with three options:

• Into the wiki — the article is saved, and its full text creates or extends wiki pages.
• Keep — saved and searchable, but not part of the wiki.
• Skip — gone.

The summary and the report itself never go into the wiki. They help you choose; they are not evidence. Why that choice stays with you is described in Who actually maintains your LLM wiki?.

What about Pocket, Readwise or Matter?

Read-later services and the morning report start at opposite ends. A read-later app manages a list you fill: you save what you want to read and read it later. The morning report starts from your feeds and tells you what of it matters today.

• Pocket was a clean read-later queue with a calm reader, and Mozilla shut it down in 2025. It kept what you had chosen; there was no summary across several articles.
• Readwise Reader is an excellent reader that collects articles, newsletters, RSS and PDFs in one place, with first-class highlighting. Concept pages or briefings across many sources are not part of its core.
• Matter is a particularly pleasant reader with a newsletter inbox and RSS; its AI assistant works within the article you are reading. It has no report output.

If you want to read long pieces in peace, these tools serve you well. The morning report does something else: it selects, summarises from the articles themselves, connects the items of a theme, cites every sentence, and through search feeds brings in publishers you do not follow yet. What of it stays is your decision, article by article.

The one-to-one comparisons: Pith vs Pocket, Pith vs Readwise and Pith vs Matter. Overviews: Pocket alternatives and Readwise alternatives.

Looking back

Every day stays stored: report, articles and your decisions. The Reports page lays the days out as a timeline — date, situation, the first key points, themes and how many articles you have already decided. A click opens that day's report as it looked that morning.

An article from the day before yesterday can still go into the wiki today. And if you want, the stored reports turn into an audio roundup every few days. More on that in the next post.

If you have connected an AI assistant over MCP, it can fetch today's report as well, as text with the same citations.