Using LLMs for web search

I have two main use-cases for LLMs: writing code and searching the web. There’s a lot of discussion online about LLM-assisted programming but very little about LLM-assisted search, so here are some unstructured thoughts about just that.

OpenAI calls their web search feature ChatGPT Deep Research, Google has Gemini Deep Research, and Anthropic just uses the word Research. Regardless of the label, all these products work the same way:

  • You enter a prompt, just like in any other LLM workflow.
  • The LLM may ask you a clarifying question (for some reason it’s always a single clarifying question, never more than that).
  • The LLM searches the web for pages matching your prompt using a traditional search engine.
  • It uses the web pages it found to generate a long report, citing its sources wherever it makes a claim within the text of the report.

I like this feature a lot, and I lean on it heavily when I’m looking for high-quality long-form writing by actual human beings on a topic that’s new to me. I use Kagi for most of my search queries, but it helps to turn to an LLM when I’m completely unfamiliar with a topic, when I’m not even sure what keywords to search for.

Or, sometimes, when I just don’t feel like sifting through thousands of words of SEO content to find a few good results.

I rarely ask LLMs factual questions because I don’t trust the answers they return. Unless I have a way to verify their output, I consider anything an LLM says to be a hallucination. This is doubly true when I’m learning something new. If I’m new to Rust, how can I be sure that the Rust best practices Claude is telling me about are truly the best of all Rust practices?

I find it much easier to trust LLM-generated output when it cites web pages I can read myself. This lets me verify that the information comes from an entity I can trust, an entity that’s (hopefully) a real person or institution with real expertise. Grounding the LLM in web search also means I’m likely to find more up to date information that’s not in its training data yet (though, in my experience, this is not always guaranteed).

Whenever I get curious about something these days, I write down a detailed question, submit it to Claude, and go off to get some work done while it churns away browsing hundreds of web pages looking for answers. Sometimes I end up with multiple browser tabs with a different search query running in each of them.

For most questions, Claude is able to write me a detailed report with citations in five to ten minutes, and it looks at about a hundred pages in the process. Sometimes it decides to spend a lot more time browsing, reading multiple hundreds of web pages. For one question, it churned away for almost seventeen minutes and looked at five hundred and fifty sources to generate its report.

I don’t actually care for the report Claude produces at the end of its research process. I almost never read it. It’s a a whole lot of LLM slop: unreadable prose, needlessly verbose, often misrepresenting the very sources it quotes. I only care about the links it finds, which are usually entirely different from what I get out of Kagi.

My workflow is to skim the report to get an idea of its general structure, open all the links in new tabs, and close the report itself. I wish there was a mode where the “report” could just be a list of useful links, though I suppose I could accomplish that with some clever prompting.

The web pages Claude surfaces always surprise me. I can’t tell what search index it uses, what search keywords it uses under the hood, or how it decides what links are worth clicking. Whatever the secret sauce is, I regularly end up on web pages that I’d never be able to find through a regular search engine: personal websites that were last updated 20 years ago, columns from long-defunct magazines, ancient Blogger and LiveJournal blogs, pages hidden deep inside some megcorporation’s labyrinthian support website, lecture notes hosted on university websites, PDFs from exposed wp-content directories, and other unexpected wonders from a web that search engines try their best to hide away.

Sometimes Claude links to pages that aren’t even online anymore! How is it able to cite these pages if it can’t actually read them? Unclear. I often have to pull up a cached version of such pages using the Internet Archive. For example, one of the reports it produced linked to On Outliners and Outline Processing and Getting Started Blogging with Drummer, both of which no longer exist on the web.

I can’t tell whether any major LLM providers take their web search products seriously (outside of Perplexity, which is not technically an LLM provider). I certainly haven’t seen any new changes made to them since they were introduced, and nobody seems to talk about them very much. For me, though, web search is one of the main reasons I use LLMs at all. That’s partly why I’m giving Anthropic that $20 every month.

I have a long wishlist of features I want to see in LLM-powered search products:

  • Stop hiding away web search inside a menu! Let me directly click “New search” in the same way I click “New chat” or “Code” in the Claude sidebar.
  • Let me edit the LLM’s research plan before it starts searching (Gemini lets me do this to some degree).
  • Let me edit the keywords the LLM will use to start its research process. Or, let me ask it to automatically refine those keywords before it begins.
  • What if the LLM finds new information online that re-contextualizes my original query? In those cases, allow it to interrupt its research process and ask for clarifications.
  • Let me look at the raw search results for each keyword the LLM searched for.
  • Add a mode where the LLM picks the best search results for me and only returns a list of links, like a traditional search engine.
  • Let me use “lenses” in the same way I use them on Kagi. Allow me to place limits on my sources (e.g social media, personal blogs, news websites, or academic journals).
  • Let me uprank/downrank/ban certain sources in the same way Kagi allows.

Maybe Kagi Assistant will grow into this in the future? Maybe I should try using Perplexity? I’ve had meh experiences with both these products, and I’m not sure whether they can compete with the quality of results ChatGPT/Claude/Gemini surface.

Anyway, yeah. I like LLM-powered search.

Weeknote 2025-W45: Text message bankruptcy

Work

  • Once again, did zero work on the novel this week.
  • Didn’t work on the comic for Indie Comix Fest either. Blame my broken brain.
  • Ironically, I wrote and published a blog post about my relationship with work. Whatever you’re thinking right now, stop thinking it immediately.
  • Does emotional work count as work? Because I did a lot of that this week. I spent many of my days reading, taking notes, and writing in my journal. I have little to show for it in terms of output, but I feel more at peace today than I did at the end of last week.

Not work

  • I mentioned in my last weeknote that I wanted to carve out more time for myself. I’m happy to report that I managed to do just that on most days this week, spending much of my time on things that make me happy and content. As I write this, I feel a lot more like myself.
  • The orange cat I mentioned last week had to have his tail amputated because of necrosis. It’s a terrible shame, but the doctors said it was the only way he could have a life free of pain and infection. He’s still recovering at the vet’s, but he should be back home this coming week.
  • I’m declaring text message bankruptcy. If you’ve texted me in the past few weeks and haven’t received a response, I apologize profusely. Abandon all hope at this point, because I’m not likely to respond back.
  • I gifted Tanvi a MagiClick AI Button as a late birthday present. The previous version of the MagiClick button was sold as a programmable macropad powered by an ESP32. This new version is sold as a sort of AI assistant, and comes with a pre-installed firmware that allows you to talk to an LLM served by XiaoZhi AI using your voice. The control panel allows you to chose between a couple of popular Chinese open-weights models, tweak a bunch of parameters, and set a custom system prompt. The default prompt tells the LLM to act like an AI girlfriend who is a “biker chick”. All I’m going to say is that some choices were made. Regardless, the button is fun to play with and can presumably be used to control something like Home Assistant.
  • I finally met fellow weeknoter Nats and her partner Rudy on Saturday. We spent an entire evening talking about stationery, Hobonichi Techos, Sakura Microns, gothic literature, Rust, LLM assisted programming, and bare-chested Stalin on a horse.
  • My new therapist had to cancel our appointment on Monday because she broke her ankle. I know I sound selfish as I say this, but I was incredibly upset when I received the cancellation message. I’d been feeling awful the previous week and really needed the support. I couldn’t have been more grateful to her when she rescheduled our appointment for later in the week, despite still being in pain from her injury. She didn’t have to do it, and I hadn’t asked for a fresh appointment. She did it anyway, not just for me but for her other clients as well. I’m filled with gratitude, even as I understand that working while sick is unsustainable in the long run. (Didn’t I already say I feel selfish? That absolves me from all guilt. Them’s the rules.)

Media diet

  • Finished playing A Short Hike
  • Finished watching Superstore
  • Started watching Nobody Wants This
  • Still reading Fingersmith and All About Love. Sigh.

The solace of work

I like my work and I like working. Would I say I love my work? I’m not sure. Maybe? I don’t know what it means to love something that’s not alive. In any case, the distinction is not important.

As far back as I can remember, I’ve enjoyed working and looked forward to my workdays. I count myself lucky in that sense. I’ve somehow stumbled into work that I find meaningful.

To me, writing code or fiction or essays does not feel significantly different from playing a video game or watching a TV show. It’s all play and it’s all work and it’s all really the same to me.

Of course, there are always some parts of any kind of work that are frustrating. I lose motivation when I’m forced to work on something I believe has no value. Work that’s not challenging enough can be fun in small doses, but starts to feel depressing after a while. Work with unreasonable deadlines or constantly changing requirements is usually a recipe for burnout (though even this nightmare scenario can sometimes present a satisfying exercise in creative problem solving under extreme constraints). And self-directed work, such as this blog post, always feels more meaningful compared to work I do for other people in exchange for money.

The idea of separating my work from the rest of my life—the idea of leaving my work at work—makes no sense to me. My work and life have always been deeply intertwined. I’ve met some of my closest friends through work. Work gave me the financial freedom to move out of my parents’ house and start a new life here in Bangalore, where I met many of the people I love most dearly today. Traveling for work allowed me to experience new places, new cultures, new ways of being.

How do I pull out the one strand marked “work” from the rich braid of my life and cast it aside, even temporarily? It’s impossible. Work, life, play, love—they’re all the same thing.

Lately work has been a source of comfort. My life has been tumultuous this year, but the pleasures of work have remained constant. With all the emotional ups and downs, there has always been blog posts to write, software to build, plot outlines to wrangle, and poetry to scribble in the back of notebooks.

It’s a powerful feeling. Come heartbreak or contentment or grief or celebration, I have always been able to come back to the work. The notebook and pen, terminal and editor, have always been ready for me, waiting for me to put my mark on their canvas. They have been my constants, my anchors in a turbulent sea. They have been my lighthouses, always guiding me back to myself.

I like my work. You might go as far as to say that I love my work. But most importantly, I’m grateful that it exists, that it is never-ending, that it waits for me every morning. I’m grateful that I can lose myself in it whenever I want and, eventually, hopefully, if I do it right, find myself.

Weeknote 2025-W44: Oklahoma smash

Work

  • I haven’t worked on the novel in two weeks. The week I was traveling for Diwali was going to be a wash anyway, but my inability to work during this post-Diwali week caught me by surprise. A combination of social commitments, mild sickness, physical exhaustion, and poor mental health kept me away from my desk.
  • Itihas led the session for this week’s IndieWebClub, where they talked about digital gardens and showed off their note-taking and digital gardening setup. As expected, the discussion veered off into note-taking methods, PKM, non-linear thinking and writing, and the joys of hypertext. I realized during the session that I prefer to think in neat, linear narratives and do not wish to start a digital garden (even though I’m fascinated by other people’s gardens).
  • At Writers Club, we used the submission guidelines for the Journal of Imaginary Research to write abstracts for fictional papers. My abstract was a mix of sci-fi with a subtle horror element. After two weeks of zero creative work, writing something and calling it done felt good. I’ve grown to enjoy writing using writing prompts. They challenge myself to think in different ways, and they’re especially fun to do in a group setting. I want to do a lot more of this!
  • I’m helping a friend write a comic for Indie Comix Fest. The submission deadline is next week, so I’m not sure if we’ll make it. But we’re going to try our darndest.

Not work

  • I’ve been struggling with mental health lately, especially when it comes to intimate relationships and close friendships (aren’t they the same?) To help me work through these (long-standing) issues, I’ve started going to a new therapist. The first session on Monday was promising and left me feeling hopeful, and I’m looking forward to the next one.
  • I’ve been finding it difficult to carve out time for myself, which is probably a contributing factor to my poor state of mind. In the coming few weeks I will attempt to find my way back to myself.
  • I’ve also been finding it difficult to set priorities for myself. I’ve been taking time away from my most important tasks and spending them on trivialities instead. The problem is exacerbated by my sleep cycle being completely out of whack. A good recipe for burnout and resentment.
  • The silver lining in all of this was a lovely dinner date with somebody I have come to care about deeply.
  • Guerilla Diner makes the best Oklahoma smash, don’t @ me.
  • For the last three years, an orange cat has made a home in the common areas of my apartment building. Sometime this week, he came back badly injured, with deep wounds across his body. The family on the ground floor—who take care of him and feed him—found him yowling outside their window, demanding to be fed but otherwise unable to move without pain. We tried for a few days to get him into a crate so we could take him to the vet, but he refused to let anybody come near him. Close to midnight on Thursday, my friend Shruti finally lured him into the crate and we drove him to the vet. I’m happy to report that he is pain-free and on his way to recovery!

Everything is index cards

I wrote this post as an exercise during a meeting of IndieWebClub Bangalore.

Much of the knowledge-management software we use today is modeled after stacks of index cards. Here is a growing list of applications that are directly or indirectly digital recreations of processes and techniques commonly used with index cards in the analog world.

I’ll keep appending to this list.

HyperCard

Hypercard

Image credit: https://wiki.xxiivv.com/site/hypertalk.html

More information: https://en.wikipedia.org/wiki/HyperCard

Trello

Trello

Image credit: https://www.atlassian.com/blog/trello/trello-manage-team

More information: https://trello.com

Google Keep

Google Keep

Image credit: https://zapier.com/blog/google-keep-vs-evernote/

More information: https://keep.google.com

Scrivener

Google Keep

Image credit: https://startupstash.com/tools/scrivener/

More information: https://www.literatureandlatte.com

TiddlyWiki

TiddlyWiki

Screenshot from https://zettelkasten.sorenbjornstad.com/

More information: https://tiddlywiki.com

Weeknote 2025-W42: Magical and terrifying

Work

  • Wrote this weeknote on an airplane flying to Delhi. Magical and terrifying.
  • Zero progress on the novel this week. Instead, I spent most of my time rushing to doctors’ appointments, preparing for my Delhi trip, and collapsed facedown on my bed after driving across the city for three hours straight.
  • Next week is likely to be a wash, too.
  • I should probably check my calendar before I set writing goals.
  • I’m hoping to at least get some reading done. Looking forward to finishing Fingersmith and moving on to Close to the Machine: Technophilia and Its Discontents.

Not work

  • I’m mildly allergic to eucalyptus and English plantain. The allergies clog my nose when I’m outdoors, which makes me susceptible to upper respiratory infections. On my allergist’s recommendation, I’ve been taking supplements, regularly washing out my nose with a saline nasal spray, and using a nasal decongestant at night. Thanks to these treatments, I haven’t been sick as sick this year as I usually get. Last month, the allergist recommended that I start immunotherapy, which could fix my upper respiratory issues once and for all. So for the past few days, I’ve been putting a tiny drop of custom-made medication under my tongue right after I wake up. I can’t tell if it’s doing anything, because it feels like nothing.
  • To scratch my programming itch, I’ve started learning Clojure. I’m treating it the same way people treat Wordle or the daily crossword: an entertaining way to give my brain a vigorous workout. I’m not expecting to use it for production projects, at least not in the immediate future.
  • I wasted too many hours reading tech news last week, which is something I only do when I’m trying to hide from negative emotions. It’s not surprising. I’ve had a tumultuous few months. But thankfully I’m doing better this weekend, and I’ve set up an appointment with a therapist for when I get back home to Bangalore.
  • I don’t like being in Delhi. It’ve never felt at home there. But I’m looking forward to seeing my brother and parents after a while, and probably some cousins I haven’t seen in over a decade.

Just read

I wrote this post as an exercise during a meeting of IndieWebClub Bangalore.

I started this year with a goal to read thirty-six books. Towards the end of last year I was easily averaging two to three books a month, so thirty-six seemed like an achievable goal.

I clearly overestimated myself, because I’ve only read ten books so far in 2025. In fact, this might end up being my worst year of reading since I started tracking my books on Goodreads in 2013.

So what happened this year that prevented me from reading more? It wasn’t a lack of time, energy, or interest. I didn’t take up a new hobby, start a challenging job, or get into a new relationship. I didn’t travel more than usual, nor did I fall sick more often than I usually do.

What went wrong was that I kept getting in my own way.

I turned reading into a chore by insisting on taking detailed notes on everything I read. Instead of relaxing on my couch with a good book and a hot beverage, I read at my desk with a pencil in hand, underlining interesting passages and scribbling in the margins. For some books that I considered particularly important, I typed up extensive notes.

This changed my relationship with reading. Instead of something I did for pleasure in my free time, it started feeling like work. Instead of something I could pick up and put down whenever I wanted, books became sacred objects that could only be approached at specific times in specific parts of my house. I started avoiding reading, which bred guilt, which bred resentment.

Today, reading has gone from being a normal part of my daily life to something I do once or twice a week, something I have to put on my todo list and check off like a chore.

It sucks.

So I’ve been trying something new: I’m reading without taking notes, highlighting important lines, or even trying to remember everything I read. If I encounter something important, I re-read it or stop to reflect for a minute. But at no point do I turn my reading time into a study session.

I’m finding this surprisingly difficult to do. I’ve become so used to extracting maximum value out of my books that to just read them for pleasure feels like a radical, uncomfortable act. Without that pencil in my hand, I’m actively anxious while reading.

But I keep reminding myself to just read. To not try to remember and catalog every little bit of information I come across. Books are not resources to be mined, but meals to be enjoyed. If something is important, it will come back to me. If it’s not, no amount of highlighting or note-taking will make it stick.

I’m hoping by allowing myself to just read, I can change my relationship to reading and go back to enjoying it like I used to. The number of books I finish this year is not even important. I just want reading to once again become something I do as a matter of course in my daily life, not as a sacred act that requires complex rituals.

I miss reading for pleasure. I’m hoping I can find that joy once again.

Weeknote 2025-W41: Starfish on the floor

Work

  • What do novelists do between getting a promising idea for a story and starting to type out their first lines? Is there a planning phase of some sort? Do they sit at their desks and just think really hard? Do they draw diagrams? Meditate? Go on walks? Consult the oracle? I’ve spent a lot of time learning about the writing processes of a whole bunch of authors, but I still have no clue what happens during the pre-writing process.
  • For my own novel, I’m spending my time researching, thinking, planning, and frantically scribbling notes to myself. Sometimes I starfish on the floor of my home office in despair. I’m not sure if this is valuable work or plain old procrastination. There is no way to tell.
  • That said, I’ve made satisfactory progress this week. The broad strokes of my plot are now clear to me, though I still have to iron out the specifics. I vaguely understand the mysteries behind the fantastic elements of my story world. I have a deeper understanding of my themes, and with it my protagonist’s psyche.
  • Next up, I need to figure out my main cast outside of the protagonist, outline some important climactic events, and finish reading the seventy three Wikipedia tabs I’ve opened in my browser. I’m hoping to draw up an outline for at least the first chapter by the end of the upcoming week. If I’m good and kind and brave, I might even be able to outline the first two or three chapters.
  • I’ve finally found a home for all my long-form writing in Ulysses. I’d prefer to use Scrivener for this work, but I don’t want to deal with its archaic and unreliable cloud sync. I briefly considered iA Writer, which is my default Markdown editor on all my computers, but it doesn’t have the tools to manage projects that contain more than two or three of files.
  • I will not be evaluating any other writing apps. I’m done with that life.

Not work

  • I’m going back to therapy. I recently had an insight about an issue that had stumped me for years, and I need to work through it with a professional.
  • I haven’t been able to make time for Hades 2. In general, games have taken a backset in my life this year. That’s fine, I suppose? Different seasons and all that. I don’t see myself putting significant time into a game again for a few months at least.
  • I’m itching to write code again. I haven’t worked on a significant programming project since March. Ugh.
  • I’ve started tracking my habits using Streaks. It’s silly, but watching that little star appear in the app motivates me to keep journaling and cleaning the litter boxes every day. Just goes to show I’m so easily manipulated, I can do it to myself.
  • I’m grateful for all the new friends I’ve made this year at IndieWebClub Bangalore and other local events. I’m also grateful for all the old friends I’ve reconnected with. If you’re reading this, know that I love you!

Weeknote 2025-W40: Just one chapter

Work

  • I’m writing a novel!
  • To track my progress and keep myself accountable, I’m restarting my practice of regularly publishing weeknotes.
  • I’m terrified. Never in my life have I been more afraid of starting a new project. Send love and reassurance and chocolate.
  • So far, I have a basic premise and an inciting incident to kick off the story. I also have a rough idea of the themes I want to tackle. My protagonist remains a mystery, but I hope to discover who she is in the process of writing the first few chapters. I have an ending in mind, but I’m not tied to it at this early stage.
  • I’m planning to write this book one chapter at a time. Thinking about the full story all at once gives me a panic attack. My immediate goal is to write just one chapter and email it to early readers without worrying too much about what comes next. Any plot problems I create today are Future Ankur’s responsibility, not mine.
  • To figure out how to structure my story, I’m reading Sarah Waters’ Fingersmith and doing a sort of reverse outline. For a book that was shortlisted for the Booker, Fingersmith is highly entertaining and tightly plotted. I want my story to have the same kind of pacing, structure, and plot.
  • I’ll probably share more details about the general shape of the novel when I’m a few chapters in.

Not work

  • I’ve started using Jujutsu for all my personal projects. My friends are sick of hearing me talk about it, which means I must find a fresh audience for my sermons. Reader, have you accepted our Lord and Savior jj into your life yet?
  • Megabonk is crack. No brain, only bonk.

Tech hot takes presented without nuance, context, or evidence

I wrote this post as an exercise during a meeting of IndieWebClub Bangalore.

  • There has never been a difference in the quality of user experience between Windows and macOS. They’re slightly different spins on the exact same idiom.
  • Graphical interfaces are strictly superior to command line interfaces in every way that matters.
  • Any sufficiently complicated UI framework contains an ad hoc, informally-specified, bug-ridden, slow implementation of half of HTML and CSS.
  • HTML and CSS are the best technologies for building user interfaces in 2025.
  • Learning CSS expands your mind in the same way learning Common Lisp or Haskell expands your mind. If you dislike CSS, you probably don’t understand CSS.
  • Mobile operating systems (iOS/Android) require as much skill and knowledge to use as desktop operating systems (Windows/macOS). Anybody who can comfortably use a smartphone today can also learn to use a desktop computer.
  • LLMs are not such a big deal. Learning to use them well nets you marginal productivity gains, but ignoring them completely at this juncture won’t do much harm.
  • LLMs will only improve productivity when we build better UIs for interacting with them. A chat box is the worst possible way to use an LLM.
  • git is an impediment to good software engineering. Its baroque and confusing UI results in hundreds of thousands of wasted developer hours across the world every year, costing organizations tens of millions of dollars in developer salaries.
  • Social media as it exists now should be regulated the same way we regulate gambling.
  • Bonus non-tech take: writing is not thinking, good writers are not necessarily good thinkers. You don’t even need words to think.