Iris triages the firehose. Flagged signal becomes tracked work. A loop, not a feed.
Every day brings hundreds of items: newsletters, podcasts, GitHub releases, X threads, HN stories, Reddit posts, YouTube uploads. Most of it is engagement bait. Some of it matters. Manually scanning the firehose eats the hours that should go toward building.
Persek OS turns the firehose into something useful through a loop, not a feed. Iris filters and synthesizes. Items I flag become tracked actions, research requests, knowledge updates, drafts, or follow-ups. Over time, taste compounds. The system learns what I value and gets quieter.
Most "AI news" workflows fail in one of three ways. Persek OS is built to defeat all three.
01 · The firehose
Hundreds of items publish daily across newsletters, YouTube, podcasts, Reddit, X, HN, GitHub. Most is engagement bait. Manual scanning eats hours that should go toward building. Time is the scarce resource. The filter must be automated, ruthless, and per-item.
02 · No feedback loop
Most readers don't learn what you value. You re-filter the same noise daily. Without a feedback loop, taste never compounds. The system stays naïve forever, and your time keeps getting eaten by the same low-signal items.
03 · Brief without action
Reading isn't doing. A great brief that doesn't convert into work is just better noise. Without an explicit conversion path, intel becomes background. Signal lands, brief is read, the moment passes, and nothing happens.
Three phases run every morning. Hundreds of items go in; a tight brief comes out, sharp enough to read on a phone in line for coffee.
Phase 1 · Ingest
A broad set of public sources gets deduplicated and filtered before anything reaches me.
Phase 2 · Extract & filter
An LLM extracts the substance of each item, then runs it through a three-question signal test. The remaining items get a quality score multiplied by a per-source weight, and only items above threshold pass through. Hundreds of items get cut down to a few dozen the brief actually surfaces.
Phase 3 · Synthesize & deliver
The surviving items get synthesized into three outputs: a deep read, a phone-friendly version, and a short summary for context on the go. Three formats, one pipeline.
A reader without a feedback loop is just a glorified RSS list. The signal that makes this work is what happens after I read the brief.
When I rate items in the brief, my verdicts feed back into per-source weights. Sources that consistently surface gold rise. Sources that publish noise sink. Over weeks, the same firehose becomes more relevant, without me curating sources by hand.
That is the loop that actually runs today: source-level. Source weights compound; the system gets quieter and more pointed as it learns which voices have earned attention. By month six, the brief looks nothing like a stranger's.
A brief that doesn't convert into action is better noise. The flagging loop closes the gap.
When I read a brief item that should produce work, I flag it. Cal may pick it up, or I may send it directly to the specialist who should own it. Either way, it becomes a build task, research prompt, draft, knowledge update, or follow-up.
The conversion path matters more than the conversion itself. The signal lands, the brief is read, the item is flagged, the work appears in the right place. The moment doesn't pass. Nothing has to be remembered.
Build → Linear
A build flag turns the item into a build task with the source link and enough context to act.
Research → Rex
A research flag turns the item into a deeper research prompt with the relevant context attached.
Knowledge → update
A knowledge flag turns the item into a durable update for a topic the brief surfaced.