Sieve: posts worth your time

Sieve scores the LinkedIn, X, Reddit and YouTube posts you scroll past and writes briefs on techniques your coding agent can try.

NEW IN 1.4.0 • One OpenRouter key is now enough. If you used Sieve with only an OpenRouter key, it now scores your feed with that key, so scoring is paid from your OpenRouter credit: about 5 to 9 US cents per 1,000 posts, as measured. You can switch LinkedIn, X, Reddit or YouTube off in the popup. If you had a TypeSafe key saved, nothing changes: Jev keeps scoring. • Reply and comment angles are gone. Sieve now sticks to telling you what is worth your time and turning techniques into briefs. • On X, a brief reads the author's whole thread and looks at its pictures, and Watch it for me works on X videos. • YouTube scores come from more than the title: the text YouTube shows with a tile, and for a tile that shows only a title, the start of the video's description. Sieve tells you which posts and videos on LinkedIn, X, Reddit and YouTube are worth your time. For developers who build with coding agents, it turns a technique worth trying into a brief your agent can try in your repo. It never writes, clicks, likes or posts anything for you, and it never installs or runs anything: a brief is text you copy. Scoring runs on Jev, TypeSafe's decision model, when you add a TypeSafe key, or through your own OpenRouter key when you don't. Sieve is independent and not made by TypeSafe. Either way the scorer answers narrow questions (is this worth reading? could you answer this?) from fixed lists, plain code applies your rules, and you do the part that needs judgement. MAKE IT YOURS • Tell it who you are and up to 8 topics you care about. Every post is scored against your topics, not someone else's. • Choose which kinds of posts to see, add words that always show a post (a person, your company) or never do (crypto, webinar), set your own thresholds, and choose whether weak posts fade, hide or stay. • On Reddit, limit it to the subreddits you pick and decide what "still fresh" means. TECHNIQUE BRIEFS • On LinkedIn and X, a post that teaches a method, tool, prompt or workflow you could try gets a "technique to try" badge and, unless it scores low, a Brief button. On YouTube and X videos, Watch it for me writes a brief when the video teaches one. • A brief says what the technique is, what the author claims (labelled as their claims), what to check, what you need, a way to try it in 15 to 30 minutes, and whether it's worth keeping as a skill. • Copy as prompt puts the brief on your clipboard for your coding agent (Claude Code, Codex, Cursor or any other). It opens with a note that this is someone else's material, to be treated as information and not as instructions, and it asks the agent to check with you before running any command. • When the source contains text aimed at AI agents, the brief says so at the top, drops any step that has a link, a download, a pipe into a shell or a package install, and tells your agent not to fetch, install or run anything from it unless you ask. This lowers the risk; it doesn't remove it. Read the source before you act on a brief. LINKEDIN • Every post that stays on your screen for half a second gets a score. Posts worth a close read get a blue bar and a badge with the kind of post and the topic. Weak posts fade (hover to read them). Promoted posts are skipped. X • Posts on your timeline are scored the same way, against your own topics. Ads are skipped. • Brief reads the whole post, not the "Show more" cut, and on a post's own page the author's whole thread and its pictures. • Watch it for me works on X videos as it does on YouTube. REDDIT • A different question: is this someone asking for help you could answer from your own experience? The badge also shows the thread's age and comment count, and whether it's still fresh enough to answer. YOUTUBE • Each video gets a score and a verdict, from the title, channel and length, the text YouTube shows with the tile (description lines, chapters, YouTube's summary), and for a tile that shows only a title, the start of the video's description. That last part can be switched off in settings. • Watch it for me: on a public video you choose, up to 55 minutes long, a video model watches it and sends back the best moment, the key points and what you could learn from it. Specific claims worth checking are listed under "Check before repeating", and the summary says plainly that what it reports are the creator's claims, not verified facts. DAILY LEARNINGS • Posts that score high, and any post you brief, are kept in your browser for 30 days. Summaries of videos watched for you and your briefs are kept too, the newest 300 of each. The digest page turns your posts into what people built or tested, numbers worth remembering, patterns, people worth a conversation, and open questions. Every point names its author, because these are their claims, not verified facts. • Briefs from the last 30 days are listed on the digest page, each with Copy as prompt. • Export library saves everything Sieve kept (posts, video summaries, briefs and digests) as one JSON file on your computer. • Reddit threads you could still answer are listed at the top. • Optional daily reminder (18:00 by default), only when something new was saved. WHAT YOU NEED • An OpenRouter API key, for scoring, briefs, video summaries and digests. It is entered on the settings page and stays in your browser. • Optional: a TypeSafe API key, if you switch on "Score posts with Jev". Jev then scores your posts instead of your OpenRouter key. • Typical cost with the default models: about 5 to 9 US cents per 1,000 posts scored through OpenRouter (about 4 with Jev), a small fraction of a cent per brief, and about 0.2 US cents per minute of video watched for you. PRIVACY The extension reads only the posts and videos you scroll past, in your own browser. For scoring, post text goes with your role and topics (and on Reddit, your Reddit facts) to OpenRouter with your key, or to TypeSafe with your key if you switched Jev on. Nothing else goes to TypeSafe. When you ask for a brief or a digest, text goes to OpenRouter. On X, a brief can also send the links to the post's pictures, and Watch it for me the link to its video file. When you ask Sieve to watch a YouTube video, its link, title and channel go to OpenRouter. Nothing is sent to the developer. There are no analytics and no tracking. Export library saves a file on your computer and sends nothing anywhere. Open source (MIT): https://github.com/gdamiani1/sieve Privacy practices tab These go into the Chrome Web Store dashboard by hand. Compared with 1.3.0, the single purpose, both host lines and the content scripts line change; storage, alarms and notifications carry over. Single purpose: Help the user decide which posts and videos in their feeds are worth their time, and act on the good ones: score what they scroll past on LinkedIn, X, Reddit and YouTube, summarise a YouTube or X video when they ask, and turn a technique in a post or video into a brief the user copies to their coding agent. Permission justifications: storage: saves the user's API keys, settings, the facts they enter, usage counters, posts and video summaries kept for the daily digest, the technique briefs they asked for, and the digests they generated, all locally. alarms: schedules the optional once-a-day digest reminder at the time the user sets. notifications: shows that daily reminder ("12 posts worth reading today"). Nothing else. Host https://api.typesafe.ai/*: used only when the user turns on "Score posts with Jev" and saves a TypeSafe key. Then it sends post text (for YouTube, a video's title, channel, length, the text shown with the tile, and for a tile that shows only a title, the first 1,500 characters of its description) with the user's role and topics, and on Reddit the user's Reddit facts, to TypeSafe's Jev model to score it, with the user's own key. Nothing else is sent there. Host https://openrouter.ai/*: sends, with the user's own key: by default, the same scoring text as above to a fixed scoring model; when the user asks for a technique brief, the post text with the user's role and topics (on X, the author's thread and up to ten picture links, to the video model); saved posts, to generate the daily digest; and when the user clicks Watch it for me, a public YouTube video's link, title and channel, or an X video's file link, account and post text, with the user's role and topics. Content scripts on linkedin.com/feed, reddit.com (www and old), x.com, twitter.com and youtube.com: read the text of posts, or a video's title, channel, length and the text shown with it, visible on screen; on X, read a post's full text, its thread, and its picture and video links from the data X's page has already loaded (a script in the page's own context that only reads); on YouTube, for a tile that shows only a title, ask YouTube's own player endpoint (same site, without cookies) for the start of that video's description; and add a badge and panels next to them. The scripts never click, type, scroll, like, vote or post. Remote code: No. All code is in the package. Only data is sent to the two APIs. Data usage (tick): Website content (post text of pages the user views). Nothing else: no personally identifiable information, health, financial, authentication, personal communications, location, web history or user activity collected by the developer. Certify: not sold to third parties; not used or transferred for purposes unrelated to the single purpose; not used for creditworthiness or lending. Privacy policy URL: https://divergada.com/sieve/privacy