用 Jev 做内容研究 agent 的视频筛选判官

jaffa @dsqjaffa原推发布于 2026.09.26

在 X 上打开原推

every marketer needs a content research agent that plugs into a social media brain an agent that watches TikTok, Instagram and YouTube Shorts for you, pulls from a database of 13.2M+ viral videos, and uses Jev to decide in 200 milliseconds if each one is worth your time i used it to get: - 500K views on social media in 3 days - $3,057 in a SINGLE day - just under 1,000 new leads here's exactly how to set it up for content marketing: 1️⃣ understand what your judge actually does Jev is a judge. you give it a video plus the answers you allow, and it hands back a decision with a confidence score attached. three ways to ask it: - Choice: pick one from your list. "is this a tutorial, a review, or a storytime format?" - Score: rate it on your scale. "how strong/viral is this hook, from 1 to 10?" - Noul: yes or no, as a probability. "is this video about the niche it's hashtagged under?" ask all three in one request and the response time barely moves 2️⃣ give your content research agent a niche your judge can only rate what your agent collects. set the agent up across TikTok, Instagram and YouTube Shorts with: - your niche - your brand name - your competitors' names - the platforms you post on brand and competitor names pull the strongest results. start there, then add intent phrases 3️⃣ make every video pass 2 checks before you trust it hashtags lie. one off-topic video in the sample and every benchmark after it is off. - does it match the niche? judged from the caption, the hashtags AND the transcript - are the hashtags lying? a video can tag its way into a niche it has nothing to do with this is the step most setups skip, because a reasoning model makes it expensive. i ran the same 15,200 batches on both: $115.07 on Gemini, $4.54 on Jev 4️⃣ benchmark the survivors on 80 intelligence signals for example: a hook only means something if it's performing better than the other hooks in its niche. so i group the signals into 6 intelligence categories: - hooks - formats and topics - visual - production - sound (transcript + video audio) - product and sales (CTA type and frequency) 📌 pro tip: Jev decides which videos get counted, and the tagging is a separate job. mixing the two up is the most common mistake in every Jevslop post this week 5️⃣ route every call by confidence - high confidence: act on it - shaky: escalate it to a stronger model or a person Jev is trained to be honest about its confidence. say 90%, be right 90% of the time, which is why you can run it on AUTOPILOT 6️⃣ write the judge's question like a brief this ONE sentence is the best thing i learned. early on, Jev treated every detail of my research brief as a hard rule: the format, the tone, even the audience. - mirroring the brief's exact wording back at it: 16 of 44 right - one line saying the topic is the only requirement and everything else is a preference: 33 of 44 same test set, one sentence, and 16 became 33. 7️⃣ turn the top result into a script a ranked list still needs turning into something you can film. give Claude the top results and have it: → group them by how each one opens: a question, a bold claim, or a before-and-after → pick the outlier that fits your brand best → write a beat-by-beat script modeled on its opening, pacing and structure → adapt it, never copy it word for word run step 3 first, so every script starts from a video that's actually in your niche 9️⃣ test it against your current judge before you hand it the keys before switching from Gemini, i ran mine head-to-head: - 44 human-labeled videos across 6 research intents: Jev 33, my existing judge 30 - on the videos where the two disagreed, Jev was right 3 out of 4 times 🔟 start with one niche and one brand you already post for set this process up below👇

这条是引用着另一条写的：x.com/dsqjaffa/status/2102054148925526206