What Better TV Recommendations Taught Me About Giving AI Context

I asked an LLM for a TV recommendation.

It suggested The Diplomat.

On paper, I could see why. It is a popular Netflix series, and from previous chats the model probably had enough context to know I have some interest in politics.

I watched about 20 or 25 minutes of it and thought: absolutely not.

That was when the problem became obvious.

The model had not necessarily made a bad recommendation from the information it had. I had simply given it nowhere near enough information about my actual taste.

“Give it context” can sound more complicated than it is

One of the most common bits of advice when using LLMs is to give them more context.

That is good advice, but it can also feel quite abstract.

What does that actually mean in practice?

In this case, I could have sat there trying to remember every television show I have ever liked, explain why I liked each one, list genres I enjoy, genres I avoid, how much comedy I want, how serious I want something to be, and so on.

I was sitting on the sofa looking for something to watch.

I did not want to write an essay.

So I made the model do more of the work.

Make the model draw the context out of you

I asked it to start listing TV series to me in groups of five.

For each one, I would reply with either:

• a score out of 10;
• or X if I had not seen it.

I specifically asked for only five at a time because I was doing it on my phone and wanted the process to stay manageable.

I also told it not to start drawing conclusions too early. I wanted it to work through 50 shows first so it had a decent body of preferences to work with.

That made the whole process very low effort from my side.

The model supplied the prompts.

I supplied the judgement.

By the end, it had far more useful information about my taste than I would ever have bothered typing out from scratch.

The first batch of context was still incomplete

Once we got through the 50 shows, I noticed something else.

The model had naturally gone down quite a serious route: prestige series, big HBO-style dramas and similar shows.

There was nothing wrong with that. Those types of epic series are often amazing.

But they were not the whole picture.

I also like much lighter things, particularly British comedy, so I added a few examples manually to widen the picture it had of me.

That was another useful part of the process.

Context does not have to be perfect from the beginning.

You can see where the model's understanding is becoming too narrow and correct it as you go.

The Peep Show moment

There was a small moment where I knew the process was starting to work.

While adding some comedy examples, I considered typing the famous UK comedy, 'Peep Show'.

I decided I had probably given it enough and pressed enter without mentioning it.

The model came back and said it was surprised I had not mentioned Peep Show.

That was a good sign.

It had started to build a picture of my taste well enough to infer something I had not explicitly told it.

At that point, it felt less like a generic recommendation engine and more like something that had actually learned enough about what I enjoy to make a sensible judgement.

Then it recommended something I would never have found

One of the final recommendations was Nathan for You.

I had never seen it.

More importantly, it is exactly the sort of thing I would probably never have discovered by scrolling through the usual streaming menus or looking at generic lists of popular shows.

I tried it.

Since then, I have watched it nearly every night and I am just about at the end of the fourth series.

That is quite a turnaround from watching 20 minutes of the original recommendation and giving up.

The lesson is not really about television

The useful part of this example is the mechanism.

My first prompt was effectively:

Recommend me something to watch.

The model had a little bit of context and returned something plausible.

The better process was:

Help me build enough context about my preferences that you can make a genuinely personalised recommendation.

That is a very different job.

The same idea can apply almost anywhere.

If you want better recommendations, better writing, better planning or better decisions from an LLM, one option is to give it a huge prompt full of background information.

Another is to make the model help gather that information from you.

Context can be built interactively

The broader point is that context can be built interactively.

You can ask the model questions.

You can make it present options.

You can score things, reject things, rank things or correct the direction it is taking.

You can add missing context once you notice what it has failed to understand.

That can be much easier than trying to front-load everything into one perfect prompt.

And in this case, it made the idea of “giving the model context” feel much less abstract.

I did not need to become better at describing my taste.

I just needed to give the model a better way to learn it.