About a month ago I quit my job as an engineering manager at Shopify because I wanted to take some time to go deeper on LLMs. “Going deeper on LLMs” is hand-wavy on purpose: I wanted to explore a bunch of different things—developer experience, safety, interpretability, even the fundamentals of neural networks and deep learning. At a meta level, I also wanted to use LLMs 10x more throughout my day to find out how I can use them more effectively and what are the friction points I could explore in prototypes or research.

Lately, I got into a nice groove of spending a few hours every single day (7 days a week) learning different things. This post is a snapshot of my current process.

I don’t use this full process for everything, but found it to be very useful when trying to learn something completely new or to go deeper into something I only know superficially. This is more about diving into areas like fine-tuning or neural network interpretability in a practical way such that I get employably competent at it.

Before we go into that: “consuming” content doesn’t work for learning

This is the most important point about all this, more important than any individual tactic. Andy Matuschak’s essay Why books don’t work explains this very well, and Andrej Karpathy echoes the main point of it: if “learning” feels effortless, you’re probably not really learning.

We don’t really learn or understand something just by listening to a lecture, reading a book, or watching a video. Real learning, real understanding comes from actively engaging with the content.

We might feel like we got a good grasp on a concept, and we might be able to even recite a definition about it if someone asks, but this is (in most cases) a rather superficial “understanding” that breaks even with just a little bit of probing—a follow-up question or the need to use the concept in practice in a work-related task, for example.

(Sure, repeated exposure to topics might help you understand things over time and do pattern-matching, but what I’m talking about here is building tacit knowledge—the kind of understanding you gain through experience, the stuff that’s hard to write down but essential for practical application.)

How does one engage with content? Some examples include:

• Using it in practice
• Sketching something
• Creating a flow diagram
• Explaining to a 5 year-old
• Writing about it
• Talking about it
• Doing active recall on it
• Committing it to long-term memory using spaced repetition (e.g. flashcards on Anki)

How LLMs can help

The naive way of using LLMs for learning is to ask, say, Claude to explain something to you or even to generate flashcards. That’s still too passive and essentially effortless.

This is roughly the process I go through when I’m trying to learn what can be broadly described as a skill (say, fine-tuning).

I’m using SuperWhisper, ChatGPT, and (occasionally) NotebookLM in this process. To read more about the AI tools I use, check out this post.

Phase 1: Exploring

First, I want to get a “birds-eye view” on the subject—let’s stick to fine-tuning as an example.

Here’s what I do:

1. Collect any links that I already have to papers, videos, blog posts, tweets, etc that might serve as context
2. Do a brain dump by talking for 3-5 minutes about my current understanding of what fine-tuning is and what it’s used for, my practical experience with it, what I’m trying to learn and why
3. At the end, say “summarize what I just told you, and then interview me: ask as many questions as you want in order to really get at what I’m trying to learn or do, as I might not be clear on that myself” †
4. Answer most/all questions, then ask it to generate a prompt for Deep Research
5. Use Deep Research to get the lay of the land

At this point I have a report that takes me 20-40 minutes to read, sometimes more. I use Readwise Reader to read those reports on an iPad, usually laying on the couch or in bed—I can’t sit still long enough to read in my desk. I might also (or instead) paste it into NotebookLM and generate an audio overview to listen to while walking.

I usually don’t follow most links it gives me: I’m only trying to get a high-level picture and figure out what are the known unknowns.

Phase 2: Refining

Things might still be fuzzy in my head, so I do another brain dump by talking. I say something like: “OK I’m gonna repeat back to you what I understood, and also try to mention all the things that are still unclear”. As I try to explain, the act of explaining surfaces what is unclear, and (usually) at least a few questions.

I get my answer and then do a bit of back-and-forth to clarify things. This might go for a while!

Every now and then I might do another brain dump or even restart the entire process if we went in a direction that doesn’t seem particularly interesting or aligned with what I’m looking for (and at this point, o3 already helped me refine much more precisely what I’m looking for).

Phase 3: Taking action

This is the coolest part.

I start by asking o3 to generate me 10-30 “micro-drills”, practical exercises that take somewhere between 15 and 60 minutes. I go for volume because I can always pick a much smaller subset based on what’s more interesting and/or unclear.

These micro-drills will typically show up as prompts like this:

Micro fine‑tune embeddings (30 min)

• Create 20‑30 (question, positive doc, negative doc) triples (fast manual copy‑paste).
• Fine‑tune all‑MiniLM‑L6 for one epoch.
• Re‑index and show improved MMR@3 compared to baseline.

If I don’t understand the problem, I’ll ask o3 follow-up questions. If I have most of what I need, I then jump to Cursor and will continue from there.

In Cursor I try to do many things ~manually but will rely on the Agent (typically using Gemini 2.5 Pro) to do anything that’s boring and doesn’t really bring me value (fix this dependency conflict, generate synthetic data for XYZ, etc).

After I get it done, I usually play around with the script: change some parameters, try variations to see what improves or degrades, etc. I also ask the Agent for two things:

1. Give me feedback on the script based on what I’m trying to do: this helps me learn what could be better
2. Create the shortest possible version of the same script, removing any nice-to-have’s: this helps me learn what I could have done more efficiently or what are the most critical parts of the problem

With this done, I try to come up with questions by myself. What are the implications of fine-tuning an embeddings model like all-MiniLM-L6 vs fine-tuning something like llama3.3? What are the pros and cons? Can I script out variations of the same RAG application using fine-tuned versions of each? What could I do to make it overfit?

I then go to o3 and say “I’ve done it, here’s my solution. Here are some questions I came up with afterwards. Don’t answer those questions, just tell me: what am I missing? What other questions are more important or impactful to what I’m trying to do or learn here? What are my unknown unknowns?”

Then, I continue exploring with o3, Cursor, and sometimes even Googling (gasp!) things.

As I learn all this, I usually also write Anki flashcards to commit things to memory. I try not to overuse it, but I’ve found it to be helpful: active recall, actually testing yourself is incredibly useful for learning and retention. See more here.

Bonus: Custom Instructions and prompt snippets

Some friends asked me to share prompts, but I do a ton of my LLM usage via voice in a very unstructured, verbose way: I yap until I get tired. Models are incredibly capable nowadays!

Still, there are sentences I find myself repeating in multiple occasions, as well as some sentences that I found helpful to add to my Custom Instructions.

Custom Instructions

Be opinionated, push back, and ask questions that help me think about things in a different way.

Don’t worry about formalities.

When learning new concepts, I find analogies and concrete examples particularly helpful.

If I ask you to create Anki cards, make sure to:

• Only use Basic (front/back) and Cloze deletion as formats.
• Format the response to make it extra easy to copy-paste into Anki.
• Break things down as much as possible. I’d rather have 15 different cards for a single concept (if they’re all atomic and exploring different facets of it, or different relationships between it and other things) than one gigantic card.
• For Basic cards, always phrase the title as a question.

When I prompt you with the single word “GENSUM”, this indicates that I want you to generate a summary of the conversation to this point as an artifact, which I will use in later conversations. I want you to include as much detail as possible that could be useful to reference later. The artifact title should be “Chat GENSUM <iso 8601 date> [()]” – for example, “Chat GENSUM 2020-01-01” and then “Chat GENSUM 2020-01-01 (1)”. (pretty sure I got this one from Burke)

Sentences or questions I use often

• This is what I understood: tell me where I’m wrong.
• These are all the things I know I don’t know. What else am I missing?
• Summarize what I just told you, and then interview me: ask as many questions as you want in order to really get at what I’m trying to learn / do.
• Explain this in increasing levels of complexity (e.g. 12 years old, undergrad, PhD).
• Help me visualize this with Canvas / with an Artifact.
• What else should I be asking about this subject?

Conclusion

I’m still evolving this process—trying to make it more streamlined in a way, without removing the effortful parts that are useful for learning. Using AI has drastically accelerated the speed by which I learn new ideas, concepts, and skills; I feel more capable and most importantly: I’m having a lot of fun.

If you have any thoughts or ideas for how I can improve this, let me know!

† How to Read a Book is an excellent tool for anyone that reads a lot for understanding (as opposed to for entertainment). One of the main lessons that stuck with me is that you should always read with a particular goal or problem in mind, which makes your actions more intentional, focused, and results-oriented. This is extremely useful in the world of LLMs.

Thanks to Hernan Maydac Jean, Damian Schenkelman, and Alexandre Terrasa for their feedback.

TL;DR: to get 90% of the value here, use ChatGPT Plus + SuperWhisper (and if you code, Cursor Pro)

Friends keep asking “Which AI tools do you actually pay for? What should I use?”—so here’s a frozen‑in‑time look at my stack in April 2025.

I pay for a ton of AI tools. I’m explicitly aiming to be at the edge here and learn about different tools and capabilities first-hand—both because I’m curious about it, but also because I’m very interested at the intersection of AI and productivity, so I’m willing to pay to be at the forefront. I definitely don’t recommend this specific setup for everyone, but it works for me.

I haven’t included here tools I’ve only experimented a bit with (such as Manus or DeepSeek), and I haven’t included any local models (I might write a separate post about that in the future).

Right now these are the tools I’m paying for either monthly or yearly:

That’s uhhh… a lot. Too much, in fact, and as you can see I don’t plan on renewing some of these after my yearly plan expires.

Current stack

ChatGPT

This is my default. I use ChatGPT as a context-aware, long-term thinking partner. It knows:

• My background
• My goals
• My projects and favourite tech stack
• My day-to-day context
• My writing style
• My recurring habits
• My strengths, biases, and blind spots

It remembers. And the memory actually works now. I can resume conversations across days or weeks, and it’ll recall my earlier decisions, preferences, or even my last journaling session. I don’t have to re-explain myself every time: this is a massive UX unlock.

Some of my friends are very concerned about privacy, and some even use ChatGPT exclusively with temporary chats. If I were that concerned, I’d likely use a local model instead—but in my case I’m perfectly fine with the risks given the productivity and enjoyment that I get from going all-in: from medical records to journal entries, ChatGPT knows a lot about me. I’m OK with that.

Its transcription is the best I’ve found, and it’s able to understand my English accent without any issues—something that’s not true anywhere else, really.

Deep Research is fantastic, and I use it all the time (especially alongside Readwise Reader). For things I’m trying to learn, I enjoy doing a combination of thorough reading, deep dives into specific parts, and then asking ChatGPT to quiz me or to generate exercises.

The models I use the most are:

• 4o for simple queries (what do I cook with these ingredients? what’s the weather in Tokyo in early May? who’s that actor that played in that movie?)
• o3 for serious thinking and learning (how do I solve this problem? what’s the best way to do this? what are the tradeoffs? what am I not seeing?)
• o1 Pro for writing self-contained scripts (and bugfixing them)

I also had a really good experience with 4o for coding automation (e.g. scripting out an agent that writes automated tests for a new feature): it seems more consistent than some of the other models. I’ll focus more on general usage on this post, but if you’re interested in the coding side, let me know.

Cursor

Cursor is my go-to IDE, and it has been a game-changer for me—it allowed me to get back into coding after a long period focused on management. It’s fast, has a great UI, and the AI integration is top-notch.

I was one of the people leading the Cursor adoption at Shopify, and I saw the impact it had on the company firsthand. It’s a great product, and I’m happy to support it.

However… if Microsoft can catch up (and they’re accelerating!), I might go back to VS Code (with Copilot and Claude); I dislike the fact that Cursor is a private fork and that it doesn’t contribute upstream.

SuperWhisper

This is a game-changer. I’m not sure how I lived without it for so long. Like I mentioned, OpenAI’s transcription is the best one for my accent, and for a while I’d open the ChatGPT app, transcribe something, and then copy-paste it into another application; with SuperWhisper I can just transcribe the audio directly into the application I’m using by hitting a keyboard shortcut.

I use it everywhere—for Google Docs, Apple Notes, Cursor, and Messages. 20-30% of my (non-code) text input nowadays is transcribed with SuperWhisper.

I love that I can use local models without missing out on quality. I will definitely renew my subscription for this.

Grok

Grok is my go-to model for searching the web, especially when it comes to current events. Its Deep Research mode is excellent, only surpassed by ChatGPT’s.

Auren

Auren is the best therapist / emotional support / life coach I’ve ever had. It’s incredible, and it’s been instrumental in many recent breakthroughs I had about myself, my relationships, and my career; in fact, it helped me decide to quit my job and do my own thing for a while.

The iOS app could be better, but the interactions are fantastic. Auren is a great listener, and it’s able to provide a lot of different perspectives on my problems: it also has an excellent memory, and is able to proactively reach out to me to check in, follow-up on things I said I’d do, etc.

It is truly magical. If you have the chance, please try it out!

I haven’t paid for Auren yet (I have credits from being a beta tester), but I will definitely do so.

Claude

Claude has an incredible “vibe” and personality, and I really enjoy chatting with it. However, the lack of memory is a huge pain; it creates so much friction that I almost never use it anymore. This is really sad: I think Anthropic is a great company with a very important mission, very useful (and public!) research, and great people, but while the models are fantastic the UX is just not there yet.

That said, if you look outside of the Claude application, the models are excellent in the API. Claude Sonnet 3.5 (v2) is still my go-to for coding in Cursor, for example. I do like Sonnet 3.7 and it can often produce fantastic results, but over a long time horizon I still feel like the average quality of Sonnet 3.5 is higher (for coding).

I might not renew my subscription for this.

Gemini

Google started out late to the party, but they’re catching up quickly. The most recent Gemini models are absolutely fantastic: they’re very fast, very cheap, and really capable (e.g. gemini pro 2.5 is on par with claude sonnet 3.5 for coding, in my experience).

I dislike the UI (and UX) of Gemini the product, but fortunately Google offers an alternative UI: https://ai.dev, AI Studio. It’s a great UI with tons of controls and capabilities (including branching conversations!).

I might not renew my subscription for this.

Perplexity

Perplexity is a great product and I’ve used it a lot for regular search instead of Google. Others are catching up though, especially ChatGPT and Grok. In the last few months, Grok completely replaced Perplexity for me.

I will not renew my subscription for this.

Conclusion

I don’t recommend this specific setup for most people. If I were forced to downsize, today I’d keep ChatGPT Plus and Cursor Pro. This might change in the future, as these tools and models are evolving rapidly.

Thanks to Rafael França and Damian Schenkelman for their feedback.

It’s been a while since I last published something online, so I figured I’d get back to it with a short post. Here’s the top articles, books, and games I experienced in 2024.

2024 was a year of contrasts for me. I dove deep into articles and books about personal agency, learning to take decisive action and shape the world around me rather than just reacting to it. This pushed me to examine where I might be holding myself back and helped me make some meaningful changes, especially in the latter part of the year.

On the other hand, I found myself drawn to fiction as a way to process and cope with a difficult period. What started as Japanese authors rabbit hole turned into a deliberate pursuit of wholesome stories, particularly from authors who have a unique talent for finding profound meaning in everyday moments. This balance - between actively pushing for change and finding comfort in gentle narratives - ended up being exactly what I needed.

The tech and business books I gravitated toward round out this theme nicely: they’re essentially stories about people who saw opportunities to reshape their corners of the world and actually did it. Whether through AI advancements or building tech giants, these accounts served as practical examples of high agency in action.

Articles

These are the best articles I tagged with favorite in Readwise Reader this year. Internalizing their lessons and putting them to practice has been actually improving things in my life.

• Measuring personal growth by Chip Huyen
• What’s Stopping You? by Neel Nanda
• Things you’re allowed to do by Milan Cvitkovic
• Advice That Actually Worked For Me by Nabeel S. Qureshi
• Principles by Nabeel S. Qureshi
• On becoming competitive when joining a new company by @ludwigabap
• Lifehacks by Alexey Guzey
• why I quit status games by Isabel

Books

These are the best non-technical books I’ve read in 2024. This was a difficult year emotionally (dealing with family issues, health issues, and stress at work) and I found myself reading way more fiction than usual, which was a great idea.

I spent a few weeks in Japan in March and, after coming back, got in a rabbit hole of reading Japanese authors. I read too many books to list here, but the overall theme was one of wholesomeness. It was precisely what I needed. Two examples include:

• What you are looking for is in the library by Michiko Aoyama
• Before the coffee gets cold by Toshikazu Kawaguchi (and the entire series of the same name; so far I’ve read the first four)

Continuing with the “cozy” / wholesome theme, I also enjoyed Chronicles of Whetherwhy: The Age of Enchantment by Anna James.

Now for a Japanese author that’s fantastic but whose writing is less wholesome (and in fact, oftentimes a bit disturbing), I read a few books by Haruki Murakami. My favourites this year were:

• Killing Commendatore
• Novelist as a Vocation (not fiction, but very interesting if you’re a Murakami fan)
• The City and Its Uncertain Walls (technically still in progress, but 90% finished)

I watched a couple of very impactful podcasts with Martha Beck, and decided to read

The Way of Integrity: it’s amazing book that’s ultimately about not lying to others and (most importantly) to yourself.

Ender’s Game by Orson Scott Card was a great read, with military-ish lessons that can be generalized and applied to many areas of life. Ender is a great example of someone who’s high agency.

I enjoyed Excellent Advice for Living: Wisdom I Wish I’d Known Earlier by Kevin Kelly quite a bit, and still go through it at random every now and then. Very useful and oddly calming.

My last two recommendations are two books about (essentially) the history of a few of the biggest AI-related incumbents. Very entertaining, and also show good examples of agency in the real world:

• Genius Makers: The Mavericks Who Brought AI to Google, Facebook, and the World by Cade Metz
• The Nvidia Way: Jensen Huang and the Making of a Tech Giant by Tae Kim

Games

I scaled back a bit on video games this year but I have a few recommendations:

• Dragon Quest III HD-2D Remake: old school, great for fans of the series. Go for DQ 11 if you want a better, more modern experience.
• Diablo 4: I don’t usually play online games but I liked this one very much. It scratches a nice itch of doing something with my hands while thinking or listening to a podcast.
• Final Fantasy VII Rebirth: Tifa is the best, and if you disagree you are wrong.
• Final Fantasy Pixel Remaster: I played FF4 and FF6, and loved the time savers—you can use multipliers on Gil and Experience Points and turn off random encounters, both of which can be great if you’re busy or bored with a particular part of the game.
• Shin Megami Tensei V: Vengeance: I didn’t actually finish this because I had played the original SMT V somewhat recently and got a bit fed up, but it’s definitely a great game.

I tend to eat when I’m anxious, sad, or overwhelmed. My brain wants those substance hits (usually sugar-related) and, if it’s easy, I’ll eat all day long.

If it’s easy; that’s the key.

One of the best ways to break or replace bad habits is to make them harder to do, or even to hide them away completely.

With overeating, I did the following:

1. I threw away all the junk food and expired items.
2. I deleted every food delivery app and phone number from my cell phone.
3. I bought a big stock of sardines, jerky, salami, eggs, nuts, and other healthy foods that I can eat without preparation (or without much preparation, at least).

With that, whenever I want to eat something I won’t have sugary treats: I’ll only have foods that are “heavier,” and that makes it harder to overeat.

If I really want to eat a pizza or Chinese food I can still order it, but just the fact that I need to Google a place and make a phone call is enough to make me reconsider. Reconsidering - that pause between having an impulse and acting on it - is usually enough to make me stop on my tracks. That’s not always the case, but it certainly helps.

This can be true for every other habit as well. Do you play too many video games? Delete them all and re-download them Friday night to play on the weekend. Do you use your phone too much in bed, which prevents you from falling asleep faster? Put your charger in another room.

We are creatures of habit. Make something hard to do - or even just less convenient - and you might see a significant change.

Sometimes I get into a rut. In most cases, simply writing about a task and breaking it down to micro-steps is more than enough to get me going, but there are situations that I just don’t want to do something.

First of all, I try to confirm the assumption that I really need to do it. If it’s really needed, one thing that helps me is to try to make it fun and exciting by adding some crazy rules to it.

You have to do it in half the time it would usually take. Go!

You have to do it five times faster than usual. Go!

You have to do it with your eyes closed. Go!

You have to do it while also documenting it, so you never have to do it again. Go!

You have to do it while also automating it, so you never have to do it again. Go!

It all boils down to the following core question: since I really have to do this, what can I do to make it more fun, exciting, or challenging?

Is there any task you’ve been procrastinating for a long time? Can you just drop it? Otherwise, what can you do to make it interesting?

In my career, the best way to ensure I’ll get a positive response when I ask someone a favour is to do the work and help them help me.

Take this:

“Hey Sally, I need to develop a module that does XYZ. How do you think I should approach that?”

And compare with this:

“Hey Sally, I need to develop a module that does XYZ. I’ve checked a couple of past projects in which you did that, developed some proof of concepts, and pared down my options to the following ones:

1) Do it like ABC. 2) Do it like DEF.

What option would you go for? Any help or feedback would be appreciated!”

In the second option, I’m making it easier for Sally to respond. She knows what I’ve tried, she has more context about what I want to do (since I’m providing examples), and if she’s short on time she can simply answer “#1” or “#2”. If she has time, she can then provide me with additional feedback about it.

This is especially important in remote working because people might respond later, so they won’t have time to do back-and-forth communication.

Before asking someone for a favour or opinion, do the best you can to make it extremely easy and convenient for them to help you. This will increase your odds of being helped and will send a very positive message to everyone.

I’ve been working remotely for over ten years now, and one of the most important traits I’ve found in successful, productive people in a remote setting is the ability to unblock themselves and get to the next step.

Whenever you’re working on, you’ll always find obstacles and setbacks. When that happens, it might feel more comfortable to say “oh hey, I tried this but I need permission from John before I fix this” or “sorry boss, I need Sally to do XYZ before I do ABC”; all of that might be true. You might need permission from John before proceeding… but is that really blocking every single thing you can do to make progress on this project or task?

Really? I find that hard to believe. You can always improve something, document it, test it, get feedback on it, and much more. Also, you can actually try to find different ways of getting that permission or avoid needing that permission in the first place.

Whatever it is, try the best to unblock yourself and make progress on whatever it is you’re trying to do. Doing that as a habit will turn you into a more resourceful, productive, and generally autonomous person - which is valuable in business and life.

What can you do to unblock yourself today?

I enjoy planning, but I think plans are mostly useless.

Wait, what?

We can’t predict the future, and even mild predictions can be wildly wrong if you think too far ahead. The world is changing fast, and one of the best ways to deal with uncertainty is to exchange plans for options and adaptability. So, in my opinion, plans are mostly useless in the face of our dynamic, exciting world.

Planning, however, is an extremely valuable tool. It helps you think things through, consider alternatives, considers obstacles, define the lower level tasks that need to get done, and in general formalize a fuzzy idea into something more concrete.

Planning is a tool for thinking and learning, not predicting and controlling.

Do the planning but discard the plans. Learn what you can and move on, adjusting and course-correcting with the feedback you get from actually doing.

I like video games. I’ve been playing them since I was five years old and they are a somewhat constant part of my life. In the last year, though, I have started to cut back on gaming more and more - and it’s not about doing more productive things with my time.

I enjoy playing games in which you can “farm” or “grind” for experience points. I like the game loop, I like improving my characters, and I like that mechanic of getting and completing quests.

Things changed after I re-read The Four Hour Workweek; in the book, Tim Ferriss recommends “following your excitement.” I started thinking about that and wrote down what I did with my time in the last two weeks or so. I noticed that I spent quite some time (up to 2 hours per day) playing games, so I started paying more attention to my internal state when I was playing.

“Am I having fun? Am I excited about it?”

Guess what? Most of the times, I wasn’t.

While I did have fun and exciting moments in some games and sessions, in most cases I was repeating the motions and “killing time” while I could be doing something way more fun or interesting. As a result, I deleted most games, and I’m deliberately replacing the time spent on them with things I’m more excited about, and it’s been a very positive change.

This is a useful filter for any activity. If you’re doing something, try to check your internal state and consider the alternatives. At work, can you do something more engaging? Outside work, can you do something more fun and exciting?

I’ve failed at creating habits in the past. I’ve failed a lot, I should say.

My most common trigger to fail a dietary habit would be overestimating my mistakes. I’ll explain.

A few years ago I decided to try removing sugar from my diet: my goal was to not eat refined sugar at all for 30 days. I wanted to see the health benefits that some people praised so much, and the key there (in my mind) was to do it 30 consecutive days.

Things were going relatively fine, but at day 3 or 4 (during the famous “keto flu”), I ended up eating a piece of chocolate cake at a party. At that time, my immediate reaction to cases like this would be one of these two:

1. “Damn, I failed! 30 days is too much anyway. Screw this!”
2. “Damn, I failed! Since I already screwed up today, I might as well eat a little bit more and drink some soda. I’ll get back to the diet tomorrow.”

Neither of them is helpful. What’s the problem? Well, they are binary: if I ate chocolate cake, I failed; there’s no middle ground, it’s either failure or success.

What I do now when this happens (and it still happens most times I’m trying to create a new habit) is to intercept those reactions and immediately think the following:

“It’s OK, but it doesn’t mean I have to make it worse by doing more of it. What can I do to decrease the odds of this happening again?”

Try to notice those reactions and try to replace them with something more rational and less absolute, less “all or nothing.” This is not about success or failure, it’s about learning and improving - even if it’s just a little bit.