The brain has heuristic algorithms whose calculations manifest as feelings. The heuristic output requires postprocessing to turn the feeling into information and policy changes.
It is easy to ignore subtle feelings. It is easy to suppress even strong negative feelings using superstimuli like video games, youtube, pornography, etc.
The effectiveness of suppression makes it easy to learn a maladaptive escapism policy.
Feelings contain valuable information content in that you wouldn't have generated that information through analytical thought, and that policy changes caused by that information are reflectively stable.
Integrating your emotions therefore greatly increases your effectiveness at optimizing the world.
The Song
I generated a song that encodes the key ideas from this post. It was designed to be intrinsically rewarding to listen to, so that I would actually engage with the content repeatedly. The process of creating it is described in section 9.
What This Post Is
This is a lightly edited transcript of a conversation I had with Claude (Opus 4.6, via Claude Code) on 2026-03-01. I started the conversation feeling bad and not knowing why. Through the conversation I identified what I think is a core bottleneck in my research process, built a model of the problem, wrote a document capturing the model and evidence supporting the model, generated a song to deliver the content in an intrinsically rewarding format, and built an Emacs reminder system to trigger it.
Minor edits: tool outputs, file diffs, and some of Claude's responses in implementation sections were cut for length. All of my messages are verbatim.
The post has several layers that might be independently valuable:
Object level: A model of why continuously integrating feelings into decision-making is high value, and why most people (including me) fail to do it reliably.
Process level: Using an LLM as a thinking partner for introspective work — not by asking it for answers, but by writing sentences and checking each one against how it feels.
Implementation level: Encoding an insight into a good song so the delivery mechanism is intrinsically rewarding, and building tooling to remind yourself.
Meta level: The conversation itself demonstrates the technique it discovers. I was paying attention to my feelings throughout.
The sections below follow the conversation roughly chronologically.
1. Feeling Bad and Not Knowing Why
Me: I feel really bad. I am not sure why.
Claude: That sounds rough. Not knowing why can make it harder — there's nothing concrete to push back against.
A few things that sometimes help when the cause isn't clear:
Physical basics — Are you hungry, dehydrated, sleep-deprived, or have you been sitting too long? These can create a diffuse bad feeling that's hard to attribute.
Something unprocessed — Sometimes a conversation, decision, or thing you read earlier leaves a residue that only surfaces later.
Just the body doing its thing — Mood can dip for no identifiable reason. It doesn't always mean something is wrong.
Do you want to talk through it, or would you rather just work on something as a distraction?
Me: Now (15m after conversation start) I somehow feel better not sure why. Maybe because I tried to figure out what is going on? I am still not sure exactly what this is about. I expect it has to do with some internal unresolved conflict.
Claude: That tracks. Sometimes just turning toward the feeling — even without resolving it — takes the edge off. The act of attending to it seems to matter more than finding the answer.
And "unresolved internal conflict" is a reasonable guess. Those tend to produce exactly that kind of diffuse, hard-to-pin-down discomfort. If it comes back, you might notice what you were thinking about right before it shows up — that can be a clue.
No need to force a resolution now though. You noticed it, you sat with it, and it shifted. That's already something.
2. The Real Problem Is Invisible
Me: Yes. What I actually did was just sit there and do nothing at all. Just before that I was spending 6 hours or so improving some code I wrote. [...] Now I am guessing: Maybe doing that programming is not the best thing to do, and some part of my brain knows that. Or at least it isn't clear to me that it's the best thing to do. I didn't actually run the evaluation. Yes I think that feels right. I didn't run a proper evaluation on what is the best thing for me to do. Just saying that gives me strong positive feeling, as if some part of me wants to tell me "Yes that is correct".
Claude: That's a really clear signal. The positive feeling when you name it — "I didn't run a proper evaluation" — is worth trusting. It sounds like a part of you that cares about your direction and knows when you're drifting into autopilot on something that feels productive but might not be the most important thing.
Six hours of programming is easy to fall into because it gives constant feedback — things work, you see progress, it feels good. But that's exactly the kind of activity that can crowd out the harder question of "what should I actually be doing?"
Sounds like what you want is to step back and actually run that evaluation. Do you want to do that now?
Me: Yes. Here are some high priority things to figure out. I got an interview invitation. That might be important. Also a tooth hurts. I need to figure out why and probably get a dentist appointment. I need to figure out what dentist is actually best to go to for this. There is also the thing that my sleep schedule is extremely "strange" right now. I wake up at midnight or so. Also cooking food seems probably immediately useful. Actually I found leftovers that I'll eat now. There are many more things. Here is the thing. All of these things that I listed are the kinds of things that are urgent because they have a deadline. I feel like these are not really the problem. I feel I will need to do them because I am "forced" to do them. Doing them is still good. But I feel like the real problem is to do the things that I don't have a deadline on. I am doing independent AI alignment research. Or at least I sometimes do it. But most of the time I am doing something else. I feel like this is the most important problem to focus on. This problem is kind of invisible.
Claude: You're naming something really clearly. The urgent things have built-in forcing functions — pain, deadlines, hunger. They'll happen. The alignment research has no external pressure at all, so it silently loses to everything else, including things that feel productive (like six hours of tooling) but aren't the actual work.
And the tricky part: the tooling, the code improvements, the interview prep — they're all related to the mission or at least defensible. So it never feels like you're procrastinating. But the core research keeps not happening.
The fact that you call it "invisible" is exactly right. There's no moment where you feel the cost of not doing it today. It just accumulates quietly.
I don't think you need me to propose a system for this. You seem to already know what the problem is. But if it would help to think out loud about what's actually blocking the research — whether it's fear, uncertainty about what to work on, the work being less rewarding in the short term than coding, or something else — I'm here for that.
3. Something Deeply Wrong With the Research Process
Me: Well. It's not exactly fear. Or at least not obviously so. But it does... Like there is something deeply wrong about how I do research I feel like. Probably it is emotional. Here is the thing. Often I do research and I find it extremely exciting. I did in the past just manage to work multiple days in a row just doing research full force. But that state doesn't last very long. I am not exactly sure what stops it. I think it is at least in part emotional. I am not quite sure what I feel. I think this happens sometimes... I am confused. Because sometimes it feels like that at some level the research is going really well and I made a bunch of important progress, and then I just also stop. It has happened in the past that I talked to people (including claude) about some idea that I was excited about, and then they were critical in a way that made my excitement disappear. But this might not be the main problem.
Claude: This is really worth sitting with. Let me reflect back what I'm hearing:
You can do the research. You've done it — multiple days, full force, excited. So it's not a capability problem. Something interrupts the state.
A few threads you're pulling on:
External criticism deflating excitement. Someone pokes a hole and suddenly the energy is gone. That's worth noticing — does the criticism actually invalidate the work, or does it just kill the feeling that was driving it? Those are very different things.
Stopping even when it's going well. This is the more puzzling one. If you stopped because you hit a wall, that would make sense. But stopping when things are going well suggests something else is happening. Maybe the excitement itself feels unsustainable or unsafe somehow. Or maybe as an idea develops, it gets closer to "this might actually matter" and that raises the stakes in a way that's uncomfortable.
"Something deeply wrong about how I do research." You said this but then couldn't quite name it. That's probably the most important thread.
I notice you're doing a good job of not jumping to a premature explanation. You keep saying "I'm not sure" and "this might not be the main problem." That's probably the right instinct — the real answer might not be the first thing that comes to mind.
What does it feel like right before you stop? Not the story about why you stopped — the actual sensation.
4. Updating Motivation From Bad Feedback
Me:
Here is the thing. I think very often the problem is that I am "updating my motivation" regardless of if the critique is actually correct. E.g. when I talked to claude about an idea it feels like claude didn't really get it... There is critique and some suggestion about what to do better but it sort of doesn't feel right. It feels like claude didn't get yet what I was trying to say. This doesn't happen when programming.
This also happens with other people. Most of the time I get a negative reaction when the person didn't yet quite understand why something is good.
I think that there is probably something that happens. I remember a particular piece of work that I did that seemed really important and it feels like it would be high value to continue it. It now came up 2 times or so over the last couple of months. But the reason that I stopped was that I felt like it's not worth doing. Why? I think I never actually analysed the feeling. I think the feeling... It's possible that it would have pointed me into a new direction. What to do in that line of inquiry that would be a good direction. But I think what happened is that this feeling pushed me into procrastination instead.
I think what I actually mean is that there is something wrong with the research process. Ok. Framing it like this makes it just obvious that this is true. Basically I feel like all my research directions that I pursue are... Maybe they would fail. But they usually feel like the kinds of things that could produce value if pursued. Actually I notice that I don't even have an explicit model about what a good research direction looks like. But I do have strong intuitions about them. And the major research threads that I have pursued seem very good. Like good at the level of "This seems really good intuitively and nobody else is even doing something similar it seems".
So having a good research direction may not actually be the bottleneck? It's kind of unclear, because if there was something that is really important and only I can do it (in the same way that I know how that I can write some program, and that it's just the best thing for me to do, then this might strongly affect my motivation. But this doesn't feel quite right.
Saying "Improving my research process is the bottleneck" feels somewhat right but doesn't quite capture it.
Claude: I think what you're describing is becoming clearer. Let me try to name the pattern:
Your motivation is coupled to social/emotional signals rather than your own assessment of the work's value. When programming, this doesn't matter — the code either works or it doesn't, and no one's reaction changes that. But in research, the feedback is ambiguous, and your motivation system treats "this person didn't get it" the same as "this idea is bad."
And the critical thing you said: you never actually analyze the feeling. The "not worth doing" feeling arrives, and you treat it as a conclusion rather than as data. You said yourself — it might have been pointing you toward a new direction within the inquiry. But instead it just... turned the engine off.
So maybe the problem isn't the research direction, and it isn't exactly the research process either. It's something like: you don't have a reliable way to maintain your own assessment of what's valuable in the face of ambiguous or discouraging signals. The excitement carries you when it's there, but when it gets disrupted, there's nothing underneath it to keep going.
The programming never has this problem because the value is obvious and immediate and no one can talk you out of it.
Does that feel closer? Or is it still not quite it?
5. The Core Insight: Running Away Instead of Processing
Me: The code thing was not about social signals. It was about that I am sure that I can write e.g. some LLM interface. I know I have the ability. I know if I put in the hours it'll just work. That is not the case with alignment research. But the point about the research feedback being not clear is exactly right. But it's its own thing.
I think when the not worth doing feeling arrives I don't treat it as a conclusion. Instead I run away from it by procrastinating. Playing games, watching youtube, programming random things.
I think your point about being able to keep track of what is valuable seems very important. I previously identified this as important. But there I was thinking about it in terms of how to write up documentation of my decision making, in particular the reasoning for why some direction seems worth pursuing. That might still be valuable but just the ability to integrate your momentary feelings into the decision making process seems much more important.
It seems quite possible that I wouldn't have made much progress when just "thinking in my head" instead of writing to you.
I think your point (1) seems important. If I could just be aware of my feelings moment to moment like I am currently and incorporate them into my decision making, that seems very high value. In the past I managed to do this for multiple days in a row, and I remember being very productive. But then I stopped and I don't know why.
I remember during that week I explicitly trained for 10m or so in the morning using the skill. Could it be as simple as me just doing that again every morning? I think the specific exercise that I did was just to ask myself "What is the best thing to do right now?" and then trying to answer that question while observing how I feel about the background models and proposals that I was coming up with. I think that when I did that initial session it put me in a particular mental state that managed to sustain itself.
When doing this I would talk out loud. I would say "X is the best thing to do" and see if that felt right or wrong.
Claude: That's a really concrete finding. You've done this before, it worked, and you know the specific exercise. Ten minutes in the morning, talking out loud, asking "what is the best thing to do right now?" and checking each answer against how it feels.
And the key insight: it's not just about making a good plan. The exercise puts you into a mental state — the same state you're in right now — where you're actually listening to your feelings as information. And that state then sustains itself through the day.
So the question is just: why did you stop doing the morning exercise? And can you start again tomorrow?
I do want to flag one thing. You said "could it be as simple as me just doing that again every morning?" and I notice that's the kind of insight that feels very clear right now but is exactly the kind of thing that tends to evaporate by tomorrow. You might wake up and think "I should do that exercise" but then not quite remember why it mattered, or feel like you should jump straight into something "productive."
Would it help to write down what you figured out tonight — not as documentation for its own sake, but as something you can read tomorrow morning right before doing the exercise? A short note from present-you to future-you that says: this is what's going on, this is why the exercise matters, just do it.
6. Why "Just Do It" Is Wrong
Me: I think it's exactly right, that this is the thing that evaporates by tomorrow. I think the just do it framing... It might be exactly wrong.
I can think of 3 properties that make it easier to do something reliably:
Something is intrinsically fun.
You have built up a habit and just do it automatically without needing to decide to do it.
You come to realize that something is really the best thing for you to do.
So it seems that the correct way to approach this is not to try to coerce myself. If we could build a model about why this is good to do the formal practice, and remember that model somehow, I think I would just do it automatically.
If there is a way in which we can make the activity itself more rewarding, we should do that.
It's actually strange. When I pay attention to a feeling and "understand" what that feeling is trying to point at, that actually feels really good. So it seems if you succeed at understanding your emotions that is already intrinsically rewarding. And this is a strongly positive feeling.
Perhaps it's easy to forget about that though. Because when I fail to resolve a strongly negative emotion it just feels very bad continuously. But watching youtube makes it go away. So escapism is actually a more reliable solution to feeling bad.
Claude: That last part is really honest and important. Escapism is more reliable — it always works. The emotional processing only feels good when it succeeds. And when it fails, you're stuck with the bad feeling plus the effort of having tried. So the expected value calculation your brain is running moment-to-moment actually favors YouTube. That's not a character flaw, it's just your brain being rational about short-term outcomes.
But you also said: when you do succeed, it feels really good. And the morning exercise is specifically designed to be easy to succeed at — you're not trying to resolve a crisis, you're just checking in with yourself about priorities while calm. So the success rate is probably much higher than when a strong negative emotion hits mid-research.
So maybe the model is something like:
The morning practice is not about discipline. It's about entering a state where emotional processing is easy and rewarding, before the hard feelings arrive. Once you're in that state, you handle the difficult moments when they come up during the day, because you're already tuned in. Without it, the difficult feelings catch you off guard and you default to the reliable escape.
The reason you stopped might be as simple as: one morning you didn't feel like it, you skipped it, and then you were no longer in the state that makes it self-sustaining. And from outside that state, the memory of how good it felt isn't vivid enough to pull you back in.
Does that model feel right? And if so — what would help you remember it tomorrow? Not as a command to obey, but as something you actually believe.
7. Proactive vs. Reactive Emotional Processing
Me: That is a very good point. I think doing it in the morning is much easier. It seems plausible that the strongly negative emotional feelings only arise after an extended period of ignoring more subtle emotions. The morning exercise I described is different. I would just ask from scratch what the best thing for me to do is. There usually wouldn't be some negative feeling that needs resolving. It seems likely that if you continuously pay attention to your emotions and process them as they arise, that the task of integrating them is much much easier. You are doing something right now and the feeling will be about what you are doing right now. The inference problem of "why do I feel this way" is just much easier.
So this is very interesting. There is Gendlin Focusing and it seems to be a practice that you start in response to some emotion arising. But what I am doing is different. It's not reactive. It's proactive. I just pay attention to how I feel continuously (>50% of the time I think) and immediately integrate the feeling into what I am doing. Perhaps focusing tries to solve a problem that doesn't really arise if you continually integrate your feelings.
Maybe the simplest thing to write down would just be observations that I have already made about the effectiveness of this technique. Not some abstract argument about why it is good, but just the concrete instances where it was already useful. The idea is to load the evidence into my brain that makes the... The conclusion correct. We could have the list of concrete utility generating instances followed by a statement like "Formally practicing the technique for 10m right now if I haven't done so today already is the highest value thing to do."
If I do read the observations and the assertion, and the observations are s.t. the assertion seems obviously true, then I predict I will execute the implied policy.
Claude: That's a very precise design. You're not trying to convince future-you with arguments — you're loading the evidence so that future-you reaches the conclusion independently. That's much more robust than "just do it."
8. Writing the Document
[At this point we drafted a document capturing the model, evidence, and assertion. I wrote it while paying close attention to whether each sentence felt right, editing until it did. The document is included in full at the end of this post.]
9. The Song Idea
Me: I can think of 3 properties that make it easier to do something reliably. Something is intrinsically fun. You have built up a habit. You come to realize it's really the best thing to do. [...] There is a central problem that I am having. I don't know where to write something down s.t. I will not forget. The main problem is that when I open my computer there is just random stuff open and I am immediately distracted by it. I also have a lot of sticky notes on my walls. They kind of work, but I start to ignore them at some point. It seems potentially high value to find some way to display information on my computer automatically.
[We discussed various delivery mechanisms — Emacs buffers, notifications, etc. Then:]
Me: I have in the past used suno to generate songs about specific topics like how useful it is to take a walk. So if I had a nice song that would just deliver this content, then that seems... Like when I think about the walk song I want to listen to it right now.
Claude: That's a genuinely good solution. A song bypasses all the problems:
You can't ignore it the way you ignore text (it's auditory, it demands attention differently)
If it's a good song you want to hear it — the delivery mechanism is intrinsically rewarding
It doesn't get in the way of your screen
It delivers the content without feeling like nagging
Me: Here is the thing. Forget everything you know about how to write a song. Don't write it in metaphors. Just write it like you would write a scientific article. Be direct and technical. And then make it still rhyme. But perfect rhyming isn't actually that important. The message is more important.
[Claude generated technical, direct lyrics. I generated 12 songs in Suno with various style prompts. I picked the one with the best beginning and generated a continuation with a more energetic style. The result: a piano-driven rock song that I immediately listened to 3 times and wanted to keep replaying.]
10. Building the Reminder System
[We then built an Emacs tab-bar integration: after 8 hours of idle time (catching sleep), a pulsing violet "Suggestion: Play the CIF Song" appears in the tab bar. Design considerations:]
Me: How about displaying a reminder in the emacs tab bar. It could be a text describing what to do (listen to the song) that subtly pulses between two colors. Now I am reminded that I wanted to do something, but not doing it isn't that annoying perhaps?
Me: We need a shortcut for playing the song. [...] The thing to watch out for is when I would train my muscle memory to just dismiss it immediately. I guess dismissing the notification is still available in M-x so that way I can dismiss when I don't want to listen.
Me: We need a function for stopping the playback. We want to maximize the control I have. This shouldn't be forced.
[The final system: s-l s-m toggles play/stop (and dismisses the reminder). M-x cif-song-pause-toggle for pause/resume. M-x cif-reminder-dismiss to dismiss without playing. M-x cif-reminder-activate to re-enable. The keybinding trains muscle memory toward playing rather than dismissing.]
11. Should I Post This?
Me: This conversation seems good. Would it be good to upload it to LessWrong?
Claude: I think it could be valuable, yes. [...] Some considerations: It's personal. You'd want to be comfortable with that level of exposure on a public forum.
Me: I was thinking it almost as is, with the model as a TLDR section at the beginning, and a link to the song below. [...] I want to think about if doing this is even the best thing to do right now. [...] All of these things that I listed are the kinds of things that are urgent because they have a deadline. I feel like these are not really the problem.
[I paused to evaluate whether posting was actually the best use of my time, applying the technique from the conversation itself.]
Me: I have an intuition about that doing this is good to do not because of any of the things I listed but because it's the kind of thing that might be valuable to do in ways that I haven't yet fully understood. Also I just noticed that this might actually be very valuable as LLM context. Maybe what this conversation can be is something that can be fed into an LLM and then the user can talk about the document with the LLM. It seems plausible that this would actually work well.
The Full Document
(This is the document we produced during the conversation.)
Continuously integrating feelings: evidence and practice plan
The primary goal of this document is to provide a model and evidence for this model, s.t. the policy implying assertion seems obviously true.
Model of Utility
The brain has heuristic algorithms whose calculations manifest as feelings. The heuristic output requires postprocessing to turn the feeling into information and policy changes.
It is easy to ignore subtle feelings. It is easy to suppress even strong negative feelings using superstimuli like video games, youtube, pornography, etc.
The effectiveness of suppression makes it easy to learn a maladaptive escapism policy.
Feelings contain valuable information content in that you wouldn't have generated that information through analytical thought, and that policy changes caused by that information are reflectively stable.
Integrating your emotions therefore greatly increases your effectiveness at optimizing the world.
Evidence of Utility
2026-03-01: conversation with Claude after feeling bad
I spent 6 hours programming a custom llm interface. Afterwards I felt bad and didn't know why. I sat and did nothing except paying attention to how I felt for 5-15m. I didn't force myself to analyse the feeling. I just noticed it. After the 5-15m the bad feeling had lessened significantly.
Then I wrote to claude "I feel really bad. I am not sure why." During the following conversation I started examining the feeling by writing to Claude. Through it I payed close attention to how I felt, and let my feeling guide my actions. In particular I was feeling out each sentence that I wrote, and let my feelings guide my edits, until it felt right.
This significantly improved the quality of the reflective process. Each time edit something to feel right it feels like correcting a mistake.
During the conversation I identified a core bottleneck in my research process: I don't reliably integrate my emotions into my decision-making moment to moment.
Productive week earlier this year
I practiced the technique for about 10 minutes each morning. I would talk out loud, say "X is the best thing to do" for various X, and check whether it felt right. During the rest of the day I was able to continuously track my feelings throughout the day.
Hypothesis: The morning practice put me into a self sustaining state in which I would pay attention to my feelings.
During this time I perceived myself as highly productive.
I stopped at some point and I don't know why. Possibly I skipped one morning and lost the self-sustaining state.
Plan for tomorrow
Practice the morning exercise for 10 minutes. Then spend the day doing alignment research while maintaining continuous emotional attention. Document what happens in a lab journal — what I did, what I predicted will happen, what actually happened. (Potentially also: what I learned, what to do better, decisions.) The lab-journaling is a generally useful skill: Predicting outcomes forces me to build explicit models, and comparing predictions to reality reveals flaws in those models.
The assertion
Define the technique: Try to answer the question "What is the best thing for me to do right now" while paying careful attention to my emotions, and letting them guide me, by trying to generate actions that feel right and avoiding actions that feel wrong.
Assertion: Formally practicing the technique for 10 minutes right now, if I haven't done so today already, is the highest value thing I can do right now.
Song
Generated via Suno to deliver the core message in a format that's intrinsically rewarding to consume.
Style prompt: Piano-driven rock, male vocal, high energy, powerful and urgent, quiet first verse then explosive from chorus onward, heavy drums, distorted bass, anthemic, raw belting vocals in choruses, building intensity throughout, climactic ending, like Tori Amos Little Earthquakes meets Muse
Lyrics:
Your brain runs heuristic algorithms
Their outputs manifest as feelings
The outputs need postprocessing
To turn them into information
What's the best thing to do right now
Pay attention to the feeling
Say it out loud, check if it's right
Correct it till it feels true
It's easy to ignore subtle feelings
It's easy to suppress the strong ones
With games and youtube and distractions
That make the bad sensation stop
But suppression isn't processing
You're throwing away the information
The feeling was pointing somewhere useful
And now you'll never know where
What's the best thing to do right now
Pay attention to the feeling
Say it out loud, check if it's right
Correct it till it feels true
Feelings contain information
That you can't get analytically
Policy changes from that information
Are reflectively stable
You already have this skill
You've used it and it worked
A week of morning practice
And the state sustained itself
You don't need to force yourself
You've seen the evidence already
Every time you used this technique
It produced high value quickly
Ten minutes every morning
Ask what's the best thing to do
Check each answer against the feeling
And let the feeling guide you
The only failure mode
Is not using it at all
So practice it right now
If you haven't done so today
Generation technique: I had Claude generate the lyrics based on the document and conversation we had. I also had Claude generate various style prompts. I generated 12 songs this way. Then I picked the one with the best beginning and generated a continuation of that song with a different more energetic style.
TLDR: The Model
The brain has heuristic algorithms whose calculations manifest as feelings. The heuristic output requires postprocessing to turn the feeling into information and policy changes.
It is easy to ignore subtle feelings. It is easy to suppress even strong negative feelings using superstimuli like video games, youtube, pornography, etc.
The effectiveness of suppression makes it easy to learn a maladaptive escapism policy.
Feelings contain valuable information content in that you wouldn't have generated that information through analytical thought, and that policy changes caused by that information are reflectively stable.
Integrating your emotions therefore greatly increases your effectiveness at optimizing the world.
The Song
I generated a song that encodes the key ideas from this post. It was designed to be intrinsically rewarding to listen to, so that I would actually engage with the content repeatedly. The process of creating it is described in section 9.
What This Post Is
This is a lightly edited transcript of a conversation I had with Claude (Opus 4.6, via Claude Code) on 2026-03-01. I started the conversation feeling bad and not knowing why. Through the conversation I identified what I think is a core bottleneck in my research process, built a model of the problem, wrote a document capturing the model and evidence supporting the model, generated a song to deliver the content in an intrinsically rewarding format, and built an Emacs reminder system to trigger it.
Minor edits: tool outputs, file diffs, and some of Claude's responses in implementation sections were cut for length. All of my messages are verbatim.
The post has several layers that might be independently valuable:
The sections below follow the conversation roughly chronologically.
1. Feeling Bad and Not Knowing Why
Me: I feel really bad. I am not sure why.
Claude: That sounds rough. Not knowing why can make it harder — there's nothing concrete to push back against.
A few things that sometimes help when the cause isn't clear:
Do you want to talk through it, or would you rather just work on something as a distraction?
Me: Now (15m after conversation start) I somehow feel better not sure why. Maybe because I tried to figure out what is going on? I am still not sure exactly what this is about. I expect it has to do with some internal unresolved conflict.
Claude: That tracks. Sometimes just turning toward the feeling — even without resolving it — takes the edge off. The act of attending to it seems to matter more than finding the answer.
And "unresolved internal conflict" is a reasonable guess. Those tend to produce exactly that kind of diffuse, hard-to-pin-down discomfort. If it comes back, you might notice what you were thinking about right before it shows up — that can be a clue.
No need to force a resolution now though. You noticed it, you sat with it, and it shifted. That's already something.
2. The Real Problem Is Invisible
Me: Yes. What I actually did was just sit there and do nothing at all. Just before that I was spending 6 hours or so improving some code I wrote. [...] Now I am guessing: Maybe doing that programming is not the best thing to do, and some part of my brain knows that. Or at least it isn't clear to me that it's the best thing to do. I didn't actually run the evaluation. Yes I think that feels right. I didn't run a proper evaluation on what is the best thing for me to do. Just saying that gives me strong positive feeling, as if some part of me wants to tell me "Yes that is correct".
Claude: That's a really clear signal. The positive feeling when you name it — "I didn't run a proper evaluation" — is worth trusting. It sounds like a part of you that cares about your direction and knows when you're drifting into autopilot on something that feels productive but might not be the most important thing.
Six hours of programming is easy to fall into because it gives constant feedback — things work, you see progress, it feels good. But that's exactly the kind of activity that can crowd out the harder question of "what should I actually be doing?"
Sounds like what you want is to step back and actually run that evaluation. Do you want to do that now?
Me: Yes. Here are some high priority things to figure out. I got an interview invitation. That might be important. Also a tooth hurts. I need to figure out why and probably get a dentist appointment. I need to figure out what dentist is actually best to go to for this. There is also the thing that my sleep schedule is extremely "strange" right now. I wake up at midnight or so. Also cooking food seems probably immediately useful. Actually I found leftovers that I'll eat now. There are many more things. Here is the thing. All of these things that I listed are the kinds of things that are urgent because they have a deadline. I feel like these are not really the problem. I feel I will need to do them because I am "forced" to do them. Doing them is still good. But I feel like the real problem is to do the things that I don't have a deadline on. I am doing independent AI alignment research. Or at least I sometimes do it. But most of the time I am doing something else. I feel like this is the most important problem to focus on. This problem is kind of invisible.
Claude: You're naming something really clearly. The urgent things have built-in forcing functions — pain, deadlines, hunger. They'll happen. The alignment research has no external pressure at all, so it silently loses to everything else, including things that feel productive (like six hours of tooling) but aren't the actual work.
And the tricky part: the tooling, the code improvements, the interview prep — they're all related to the mission or at least defensible. So it never feels like you're procrastinating. But the core research keeps not happening.
The fact that you call it "invisible" is exactly right. There's no moment where you feel the cost of not doing it today. It just accumulates quietly.
I don't think you need me to propose a system for this. You seem to already know what the problem is. But if it would help to think out loud about what's actually blocking the research — whether it's fear, uncertainty about what to work on, the work being less rewarding in the short term than coding, or something else — I'm here for that.
3. Something Deeply Wrong With the Research Process
Me: Well. It's not exactly fear. Or at least not obviously so. But it does... Like there is something deeply wrong about how I do research I feel like. Probably it is emotional. Here is the thing. Often I do research and I find it extremely exciting. I did in the past just manage to work multiple days in a row just doing research full force. But that state doesn't last very long. I am not exactly sure what stops it. I think it is at least in part emotional. I am not quite sure what I feel. I think this happens sometimes... I am confused. Because sometimes it feels like that at some level the research is going really well and I made a bunch of important progress, and then I just also stop. It has happened in the past that I talked to people (including claude) about some idea that I was excited about, and then they were critical in a way that made my excitement disappear. But this might not be the main problem.
Claude: This is really worth sitting with. Let me reflect back what I'm hearing:
You can do the research. You've done it — multiple days, full force, excited. So it's not a capability problem. Something interrupts the state.
A few threads you're pulling on:
External criticism deflating excitement. Someone pokes a hole and suddenly the energy is gone. That's worth noticing — does the criticism actually invalidate the work, or does it just kill the feeling that was driving it? Those are very different things.
Stopping even when it's going well. This is the more puzzling one. If you stopped because you hit a wall, that would make sense. But stopping when things are going well suggests something else is happening. Maybe the excitement itself feels unsustainable or unsafe somehow. Or maybe as an idea develops, it gets closer to "this might actually matter" and that raises the stakes in a way that's uncomfortable.
"Something deeply wrong about how I do research." You said this but then couldn't quite name it. That's probably the most important thread.
I notice you're doing a good job of not jumping to a premature explanation. You keep saying "I'm not sure" and "this might not be the main problem." That's probably the right instinct — the real answer might not be the first thing that comes to mind.
What does it feel like right before you stop? Not the story about why you stopped — the actual sensation.
4. Updating Motivation From Bad Feedback
Me:
This also happens with other people. Most of the time I get a negative reaction when the person didn't yet quite understand why something is good.
I think that there is probably something that happens. I remember a particular piece of work that I did that seemed really important and it feels like it would be high value to continue it. It now came up 2 times or so over the last couple of months. But the reason that I stopped was that I felt like it's not worth doing. Why? I think I never actually analysed the feeling. I think the feeling... It's possible that it would have pointed me into a new direction. What to do in that line of inquiry that would be a good direction. But I think what happened is that this feeling pushed me into procrastination instead.
I think what I actually mean is that there is something wrong with the research process. Ok. Framing it like this makes it just obvious that this is true. Basically I feel like all my research directions that I pursue are... Maybe they would fail. But they usually feel like the kinds of things that could produce value if pursued. Actually I notice that I don't even have an explicit model about what a good research direction looks like. But I do have strong intuitions about them. And the major research threads that I have pursued seem very good. Like good at the level of "This seems really good intuitively and nobody else is even doing something similar it seems".
So having a good research direction may not actually be the bottleneck? It's kind of unclear, because if there was something that is really important and only I can do it (in the same way that I know how that I can write some program, and that it's just the best thing for me to do, then this might strongly affect my motivation. But this doesn't feel quite right.
Saying "Improving my research process is the bottleneck" feels somewhat right but doesn't quite capture it.
Claude: I think what you're describing is becoming clearer. Let me try to name the pattern:
Your motivation is coupled to social/emotional signals rather than your own assessment of the work's value. When programming, this doesn't matter — the code either works or it doesn't, and no one's reaction changes that. But in research, the feedback is ambiguous, and your motivation system treats "this person didn't get it" the same as "this idea is bad."
And the critical thing you said: you never actually analyze the feeling. The "not worth doing" feeling arrives, and you treat it as a conclusion rather than as data. You said yourself — it might have been pointing you toward a new direction within the inquiry. But instead it just... turned the engine off.
So maybe the problem isn't the research direction, and it isn't exactly the research process either. It's something like: you don't have a reliable way to maintain your own assessment of what's valuable in the face of ambiguous or discouraging signals. The excitement carries you when it's there, but when it gets disrupted, there's nothing underneath it to keep going.
The programming never has this problem because the value is obvious and immediate and no one can talk you out of it.
Does that feel closer? Or is it still not quite it?
5. The Core Insight: Running Away Instead of Processing
Me: The code thing was not about social signals. It was about that I am sure that I can write e.g. some LLM interface. I know I have the ability. I know if I put in the hours it'll just work. That is not the case with alignment research. But the point about the research feedback being not clear is exactly right. But it's its own thing.
I think when the not worth doing feeling arrives I don't treat it as a conclusion. Instead I run away from it by procrastinating. Playing games, watching youtube, programming random things.
I think your point about being able to keep track of what is valuable seems very important. I previously identified this as important. But there I was thinking about it in terms of how to write up documentation of my decision making, in particular the reasoning for why some direction seems worth pursuing. That might still be valuable but just the ability to integrate your momentary feelings into the decision making process seems much more important.
It seems quite possible that I wouldn't have made much progress when just "thinking in my head" instead of writing to you.
I think your point (1) seems important. If I could just be aware of my feelings moment to moment like I am currently and incorporate them into my decision making, that seems very high value. In the past I managed to do this for multiple days in a row, and I remember being very productive. But then I stopped and I don't know why.
I remember during that week I explicitly trained for 10m or so in the morning using the skill. Could it be as simple as me just doing that again every morning? I think the specific exercise that I did was just to ask myself "What is the best thing to do right now?" and then trying to answer that question while observing how I feel about the background models and proposals that I was coming up with. I think that when I did that initial session it put me in a particular mental state that managed to sustain itself.
When doing this I would talk out loud. I would say "X is the best thing to do" and see if that felt right or wrong.
Claude: That's a really concrete finding. You've done this before, it worked, and you know the specific exercise. Ten minutes in the morning, talking out loud, asking "what is the best thing to do right now?" and checking each answer against how it feels.
And the key insight: it's not just about making a good plan. The exercise puts you into a mental state — the same state you're in right now — where you're actually listening to your feelings as information. And that state then sustains itself through the day.
So the question is just: why did you stop doing the morning exercise? And can you start again tomorrow?
I do want to flag one thing. You said "could it be as simple as me just doing that again every morning?" and I notice that's the kind of insight that feels very clear right now but is exactly the kind of thing that tends to evaporate by tomorrow. You might wake up and think "I should do that exercise" but then not quite remember why it mattered, or feel like you should jump straight into something "productive."
Would it help to write down what you figured out tonight — not as documentation for its own sake, but as something you can read tomorrow morning right before doing the exercise? A short note from present-you to future-you that says: this is what's going on, this is why the exercise matters, just do it.
6. Why "Just Do It" Is Wrong
Me: I think it's exactly right, that this is the thing that evaporates by tomorrow. I think the just do it framing... It might be exactly wrong.
I can think of 3 properties that make it easier to do something reliably:
So it seems that the correct way to approach this is not to try to coerce myself. If we could build a model about why this is good to do the formal practice, and remember that model somehow, I think I would just do it automatically.
If there is a way in which we can make the activity itself more rewarding, we should do that.
It's actually strange. When I pay attention to a feeling and "understand" what that feeling is trying to point at, that actually feels really good. So it seems if you succeed at understanding your emotions that is already intrinsically rewarding. And this is a strongly positive feeling.
Perhaps it's easy to forget about that though. Because when I fail to resolve a strongly negative emotion it just feels very bad continuously. But watching youtube makes it go away. So escapism is actually a more reliable solution to feeling bad.
Claude: That last part is really honest and important. Escapism is more reliable — it always works. The emotional processing only feels good when it succeeds. And when it fails, you're stuck with the bad feeling plus the effort of having tried. So the expected value calculation your brain is running moment-to-moment actually favors YouTube. That's not a character flaw, it's just your brain being rational about short-term outcomes.
But you also said: when you do succeed, it feels really good. And the morning exercise is specifically designed to be easy to succeed at — you're not trying to resolve a crisis, you're just checking in with yourself about priorities while calm. So the success rate is probably much higher than when a strong negative emotion hits mid-research.
So maybe the model is something like:
The morning practice is not about discipline. It's about entering a state where emotional processing is easy and rewarding, before the hard feelings arrive. Once you're in that state, you handle the difficult moments when they come up during the day, because you're already tuned in. Without it, the difficult feelings catch you off guard and you default to the reliable escape.
The reason you stopped might be as simple as: one morning you didn't feel like it, you skipped it, and then you were no longer in the state that makes it self-sustaining. And from outside that state, the memory of how good it felt isn't vivid enough to pull you back in.
Does that model feel right? And if so — what would help you remember it tomorrow? Not as a command to obey, but as something you actually believe.
7. Proactive vs. Reactive Emotional Processing
Me: That is a very good point. I think doing it in the morning is much easier. It seems plausible that the strongly negative emotional feelings only arise after an extended period of ignoring more subtle emotions. The morning exercise I described is different. I would just ask from scratch what the best thing for me to do is. There usually wouldn't be some negative feeling that needs resolving. It seems likely that if you continuously pay attention to your emotions and process them as they arise, that the task of integrating them is much much easier. You are doing something right now and the feeling will be about what you are doing right now. The inference problem of "why do I feel this way" is just much easier.
So this is very interesting. There is Gendlin Focusing and it seems to be a practice that you start in response to some emotion arising. But what I am doing is different. It's not reactive. It's proactive. I just pay attention to how I feel continuously (>50% of the time I think) and immediately integrate the feeling into what I am doing. Perhaps focusing tries to solve a problem that doesn't really arise if you continually integrate your feelings.
Maybe the simplest thing to write down would just be observations that I have already made about the effectiveness of this technique. Not some abstract argument about why it is good, but just the concrete instances where it was already useful. The idea is to load the evidence into my brain that makes the... The conclusion correct. We could have the list of concrete utility generating instances followed by a statement like "Formally practicing the technique for 10m right now if I haven't done so today already is the highest value thing to do."
If I do read the observations and the assertion, and the observations are s.t. the assertion seems obviously true, then I predict I will execute the implied policy.
Claude: That's a very precise design. You're not trying to convince future-you with arguments — you're loading the evidence so that future-you reaches the conclusion independently. That's much more robust than "just do it."
8. Writing the Document
[At this point we drafted a document capturing the model, evidence, and assertion. I wrote it while paying close attention to whether each sentence felt right, editing until it did. The document is included in full at the end of this post.]
9. The Song Idea
Me: I can think of 3 properties that make it easier to do something reliably. Something is intrinsically fun. You have built up a habit. You come to realize it's really the best thing to do. [...] There is a central problem that I am having. I don't know where to write something down s.t. I will not forget. The main problem is that when I open my computer there is just random stuff open and I am immediately distracted by it. I also have a lot of sticky notes on my walls. They kind of work, but I start to ignore them at some point. It seems potentially high value to find some way to display information on my computer automatically.
[We discussed various delivery mechanisms — Emacs buffers, notifications, etc. Then:]
Me: I have in the past used suno to generate songs about specific topics like how useful it is to take a walk. So if I had a nice song that would just deliver this content, then that seems... Like when I think about the walk song I want to listen to it right now.
Claude: That's a genuinely good solution. A song bypasses all the problems:
Me: Here is the thing. Forget everything you know about how to write a song. Don't write it in metaphors. Just write it like you would write a scientific article. Be direct and technical. And then make it still rhyme. But perfect rhyming isn't actually that important. The message is more important.
[Claude generated technical, direct lyrics. I generated 12 songs in Suno with various style prompts. I picked the one with the best beginning and generated a continuation with a more energetic style. The result: a piano-driven rock song that I immediately listened to 3 times and wanted to keep replaying.]
10. Building the Reminder System
[We then built an Emacs tab-bar integration: after 8 hours of idle time (catching sleep), a pulsing violet "Suggestion: Play the CIF Song" appears in the tab bar. Design considerations:]
Me: How about displaying a reminder in the emacs tab bar. It could be a text describing what to do (listen to the song) that subtly pulses between two colors. Now I am reminded that I wanted to do something, but not doing it isn't that annoying perhaps?
Me: We need a shortcut for playing the song. [...] The thing to watch out for is when I would train my muscle memory to just dismiss it immediately. I guess dismissing the notification is still available in M-x so that way I can dismiss when I don't want to listen.
Me: We need a function for stopping the playback. We want to maximize the control I have. This shouldn't be forced.
[The final system:
s-l s-mtoggles play/stop (and dismisses the reminder).M-x cif-song-pause-togglefor pause/resume.M-x cif-reminder-dismissto dismiss without playing.M-x cif-reminder-activateto re-enable. The keybinding trains muscle memory toward playing rather than dismissing.]11. Should I Post This?
Me: This conversation seems good. Would it be good to upload it to LessWrong?
Claude: I think it could be valuable, yes. [...] Some considerations: It's personal. You'd want to be comfortable with that level of exposure on a public forum.
Me: I was thinking it almost as is, with the model as a TLDR section at the beginning, and a link to the song below. [...] I want to think about if doing this is even the best thing to do right now. [...] All of these things that I listed are the kinds of things that are urgent because they have a deadline. I feel like these are not really the problem.
[I paused to evaluate whether posting was actually the best use of my time, applying the technique from the conversation itself.]
Me: I have an intuition about that doing this is good to do not because of any of the things I listed but because it's the kind of thing that might be valuable to do in ways that I haven't yet fully understood. Also I just noticed that this might actually be very valuable as LLM context. Maybe what this conversation can be is something that can be fed into an LLM and then the user can talk about the document with the LLM. It seems plausible that this would actually work well.
The Full Document
(This is the document we produced during the conversation.)
Continuously integrating feelings: evidence and practice plan
The primary goal of this document is to provide a model and evidence for this model, s.t. the policy implying assertion seems obviously true.
Model of Utility
The brain has heuristic algorithms whose calculations manifest as feelings. The heuristic output requires postprocessing to turn the feeling into information and policy changes.
It is easy to ignore subtle feelings. It is easy to suppress even strong negative feelings using superstimuli like video games, youtube, pornography, etc.
The effectiveness of suppression makes it easy to learn a maladaptive escapism policy.
Feelings contain valuable information content in that you wouldn't have generated that information through analytical thought, and that policy changes caused by that information are reflectively stable.
Integrating your emotions therefore greatly increases your effectiveness at optimizing the world.
Evidence of Utility
2026-03-01: conversation with Claude after feeling bad
I spent 6 hours programming a custom llm interface. Afterwards I felt bad and didn't know why. I sat and did nothing except paying attention to how I felt for 5-15m. I didn't force myself to analyse the feeling. I just noticed it. After the 5-15m the bad feeling had lessened significantly.
Then I wrote to claude "I feel really bad. I am not sure why." During the following conversation I started examining the feeling by writing to Claude. Through it I payed close attention to how I felt, and let my feeling guide my actions. In particular I was feeling out each sentence that I wrote, and let my feelings guide my edits, until it felt right.
This significantly improved the quality of the reflective process. Each time edit something to feel right it feels like correcting a mistake.
During the conversation I identified a core bottleneck in my research process: I don't reliably integrate my emotions into my decision-making moment to moment.
Productive week earlier this year
I practiced the technique for about 10 minutes each morning. I would talk out loud, say "X is the best thing to do" for various X, and check whether it felt right. During the rest of the day I was able to continuously track my feelings throughout the day.
Hypothesis: The morning practice put me into a self sustaining state in which I would pay attention to my feelings.
During this time I perceived myself as highly productive.
I stopped at some point and I don't know why. Possibly I skipped one morning and lost the self-sustaining state.
Plan for tomorrow
Practice the morning exercise for 10 minutes. Then spend the day doing alignment research while maintaining continuous emotional attention. Document what happens in a lab journal — what I did, what I predicted will happen, what actually happened. (Potentially also: what I learned, what to do better, decisions.) The lab-journaling is a generally useful skill: Predicting outcomes forces me to build explicit models, and comparing predictions to reality reveals flaws in those models.
The assertion
Define the technique: Try to answer the question "What is the best thing for me to do right now" while paying careful attention to my emotions, and letting them guide me, by trying to generate actions that feel right and avoiding actions that feel wrong.
Assertion: Formally practicing the technique for 10 minutes right now, if I haven't done so today already, is the highest value thing I can do right now.
Song
Generated via Suno to deliver the core message in a format that's intrinsically rewarding to consume.
Style prompt: Piano-driven rock, male vocal, high energy, powerful and urgent, quiet first verse then explosive from chorus onward, heavy drums, distorted bass, anthemic, raw belting vocals in choruses, building intensity throughout, climactic ending, like Tori Amos Little Earthquakes meets Muse
Lyrics:
Generation technique: I had Claude generate the lyrics based on the document and conversation we had. I also had Claude generate various style prompts. I generated 12 songs this way. Then I picked the one with the best beginning and generated a continuation of that song with a different more energetic style.