What is a YouTube custom feed and how does it work?
A YouTube custom feed is a saved homepage tab built from a written prompt. You describe the videos you want in plain language, Gemini interprets the request, and YouTube pins the resulting recommendation stream to the top of your home page. YouTube announced the feature on 23 September 2026 and says rollout on web and mobile begins next month.
There is a prompting technique called the curation prompt: a short brief that says what to feature, what to exclude and what to prioritise. Most people have written half of one without noticing. Almost nobody writes all three parts, and that is exactly what YouTube's new custom feeds reward.
Definition: a custom feed is a user-defined recommendation stream on YouTube. You type a description into a prompt box, Gemini builds a feed around it, the feed appears as its own tab at the top of your home page, and it keeps refreshing with new recommendations while the main feed stays untouched.
The details come from TechCrunch's report from the Made on YouTube event. YouTube's own examples include a feed of video podcasts for a 30-minute train commute and a feed of relaxing commentary videos to wind down with. Emily Moxley, YouTube's VP of Product Management for Viewer AI, framed it as a way through a corpus of more than 20 billion videos.
The part that matters for practitioners is one sentence in that announcement: the request can be lengthy, and it can spell out what the feed should feature or exclude and what content to prioritise. That is a prompt spec, and it is the same spec that fixes most inconsistent ChatGPT and Claude outputs.
Why does a feed prompt matter for anyone who writes AI prompts?
Feed prompts matter because the same three-part structure, feature, exclude and prioritise, now controls recommendations on YouTube, Spotify, X and Bluesky, and it is the structure that separates reliable AI outputs from lucky ones. Learn to write one curation prompt well and you have learned a transferable skill, not a YouTube trick.
YouTube is not alone. On the same day, Spotify launched Taste Profile in the US, which lets Premium listeners reshape their Home feed with typed requests such as more of a genre or a mood, with changes appearing within a few hours. Bluesky shipped an AI feed builder called Attie in March 2026, and X added AI-powered custom feeds in April 2026, both covered by TechCrunch.
Four platforms in one year moved from "the algorithm decides" to "describe what you want". For content creators and marketers, that changes discovery: a viewer who types "long-form interviews with Hong Kong founders, no reaction videos" is telling Gemini something the watch-history algorithm could never infer.
For power users, it is a rehearsal. A curation prompt is a constraint prompt. When your ChatGPT report drafts drift, it is usually because you specified the topic and skipped the exclusions and priorities. Feed prompts force you to write all three, and the habit carries straight back into your work prompts.
How do you write a custom feed prompt that actually works?
Write a feed prompt in four labelled parts: purpose (when and why you will watch), feature (formats, topics and lengths you want), exclude (formats and topics that must never appear) and prioritise (tie-breakers such as recency, new channels or language). Keep it under 120 words, and use full sentences rather than keyword lists.
The purpose line does more work than people expect. "For my 40-minute MTR commute" tells the model something about length, audio quality and attention level all at once. Google's own prompting guidance for its image and video models says the same thing: describe the scene in full sentences and state the context, not a pile of comma-separated tags.
Here is a complete, copy-paste-ready feed prompt you can save now and paste in when the feature reaches your account:
Try this prompt:
PURPOSE: A feed for my 40-minute MTR commute, watched with earphones, mostly listening.
FEATURE: Video podcasts and long interviews between 25 and 60 minutes about marketing, consumer psychology and Hong Kong business. Cantonese or English audio.
EXCLUDE: Shorts, reaction videos, clips under 10 minutes, anything about crypto trading, and channels that post more than one video a day.
PRIORITISE: Uploads from the past 30 days, channels I have never watched, and episodes with guests rather than solo monologues. Aim for roughly 70% new creators and 30% channels I already follow.
Notice what the exclude block does. Without it, a marketing feed fills with three-minute "5 hacks" videos because they are the most-clicked items in the category. The exclude block is where you protect the feed from the platform's default incentives.
Notice also that the prioritise block contains ratios and time windows. Models handle "mostly new creators" loosely and "roughly 70% new creators" much more consistently. Numbers are cheap to write and expensive to leave out.
How can marketers and creators use custom feeds at work?
Marketers can use custom feeds as living research folders: one feed per client industry, one for competitor content, one for skill-building. Creators should assume that viewers will soon describe feeds in words, so titles, descriptions and spoken content need to state clearly what a video is, who it is for and what it is not.
A practical setup for a Hong Kong marketing manager looks like three feeds. A competitor feed: "FEATURE: videos published by or about these five brands in the past 14 days. EXCLUDE: fan uploads and compilations. PRIORITISE: product launches and paid partnerships." A client-industry feed built the same way for, say, F&B or property. A learning feed restricted to 15-to-40-minute tutorials on the tools you use.
Each feed refreshes on its own, so the research happens while you commute rather than in a Friday afternoon scramble through search results. Delete the feed when the campaign ends.
For creators, the shift is subtler. Nobody outside Google knows exactly how Gemini matches a prompt to a video, so treat the following as a reasonable inference rather than a confirmed ranking factor: a prompt-driven feed can only feature your video if the system can tell what it is. A description that says "a 35-minute conversation with a Hong Kong café owner about pricing" is far easier to match to "long interviews about Hong Kong business" than a description that says "you won't believe what happened".
SQ Magazine's analysis of the launch also argues that niche creators may gain most, because a precise prompt can surface channels the main feed rarely pushes. If you serve a narrow audience, this is the first algorithm change in years that favours you.
What are the most common mistakes with feed prompts?
The five common mistakes are: writing only a topic with no exclusions, giving contradictory instructions, omitting length and format, expecting the custom feed to fix the main feed, and assuming the feature is live everywhere. Custom feeds are US-first at launch and do not replace the default homepage.
Topic-only prompts fail first. "Marketing videos" is a category, not a feed. The platform will hand you its most-engaged marketing content, which is rarely the content you meant. Every feed prompt needs at least two exclusions.
Contradictions are the second failure. "Deep, calm explainers" plus "high-energy, fast-paced" gives the model nothing to optimise. Pick one register per feed and make another feed for the other mood.
Missing format constraints are the third. If you do not name lengths and formats, Shorts will find their way in because they dominate upload volume. State minimum and maximum minutes.
The fourth mistake is expecting a custom feed to retrain the main feed. YouTube has said the main feed keeps showing all your recommendations; custom feeds sit beside it in their own tabs. Your watch history still shapes the default page.
The fifth is availability. Coverage of the Made on YouTube event describes the initial rollout as US viewers on web, mobile and TV, starting next month. Hong Kong accounts may wait longer. Draft your prompts now, save them in a note, and paste them in the day the prompt box appears.
How can you practise this today before the rollout reaches you?
You can practise the curation-prompt structure today in ChatGPT, Claude or Gemini by asking the model to draft three feed prompts from a description of your week, then applying the same feature, exclude and prioritise structure to your next real work brief. The exercise takes about ten minutes.
Paste this into any assistant:
Try this prompt:
I want to prepare three YouTube custom feed prompts. Each must have four labelled parts: PURPOSE, FEATURE, EXCLUDE, PRIORITISE, and stay under 120 words.
My week: I commute 40 minutes by MTR twice a day, I manage social media for a Hong Kong skincare brand, and I am learning Make.com automation in the evenings.
For each feed, include at least two exclusions and one numeric constraint (minutes, days or a percentage). Then list the three most likely ways each feed could go wrong and how the prompt already guards against it.
Read the output critically. If the model gave you a feed with no numbers, push back. If the exclusions are generic, name the specific format that annoys you. You are training yourself to spot the missing block, which is the whole skill.
Then take a real task from today, a report outline, a client email or a campaign brief, and add EXCLUDE and PRIORITISE blocks to the prompt you would normally write. Most people find the output tightens immediately, and the improvement is repeatable because it comes from structure rather than phrasing.
The bigger picture is worth naming. Recommendation systems spent fifteen years guessing what you wanted from your clicks. In 2026, four major platforms started asking you to say it. The people who can write a clear curation prompt will get better feeds, better research and better AI outputs from the same tools everyone else is using. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Find out how sharp your prompting instincts really are
You now have the feature, exclude and prioritise structure. The next step is knowing which other prompting fundamentals you are missing and turning them into a workflow your whole team can repeat. UD's free AI IQ Test scores your practical AI knowledge in minutes, and we'll walk you through every step from the results to a working setup.
