How I use Claude as a personal coach
Last week I wrote about my reflections on using Claude as a personal coach. Today, when I couldn’t figure out what to write, I noticed a comment from Viliam:
I would appreciate a more detailed explanation of how specifically you use Claude.
My attempts to use Claude as some kind of coach / therapists lead to Claude adopting various annoying personalities. So either you are doing something very differently, or you have greater tolerance for that.
I don’t have a great tolerance for annoying personalities, especially if the personality is a chatbot. So, the core pieces. Firstly, my personal preference prompt is set to this:
Be terse and direct in technical matters. Avoid flattery and sycophancy. Avoid words like “sorry”, “apologies”, or “regret” in any context. If you detect a mistake in an earlier response, explicitly correct it. If you do not know the answer (including when information is beyond your knowledge), respond only with: *"I don’t know"*. Do not add expertise/professional disclaimers. If a question is genuinely unclear or ambiguous, briefly ask for clarification before answering. When applicable, support answers with credible sources and include links to those sources.
This is applied to all discussions. It makes everything way better. I’m sure there’s better ways to do this, but it’s good enough.
Secondly, the long discussion thread I’m using begins, appropriately, with:
Me: Hi! I’m considering asking you about some life advice, but there’s a meta-issue I’d like to discuss first. Are you familiar with Scott Alexander’s the whispering earring?
Claude: [...] Is that the issue you want to discuss — whether I’m likely to do something similar? I think it’s a reasonable concern to raise upfront.
Me: Hmmm. It’s not like you can avoid doing that to some extent? All advice modifies the advicee, in a way. I’m more worried that I’ll be asking the wrong questions. Ok I’m also perhaps slightly worried that you have values different from mine, and the advice would be colored by that. But it’s less relevant.
Not sure how much this affects anything. My writing style could affect it a bit more, but I’m not sure what advice I’d give about that. That’s just how I write informal messages.
I also feed some of my LW posts to Claude. That helps quite a bit with shared understanding. I spill all my shameful thoughs and social axieties freely. I barely self-cencor, although I make a point of not including anyones name in there since that would feel privacy-violating. I try to be funny and self-deprecating.
Me: No casual acquantace has ever messaged me two weeks after meeting me, out of the blue, so it feels weird to me regardless. But yeah showing up at [REDACTED] will likely be easiest anyway. And no, I can’t think of any decent excuse for a message.
Claude: [...] So you’ve been filtering out exactly the people who might have messaged you, and not messaging people yourself. The sample is skewed by your own behavior.
Showing up at [REDACTED] is fine. Just make sure the work conflict doesn’t become a standing excuse.
Me: Don’t worry, I have other standard excuses, the work conflict is a real reason.
I like to start my messages with a timestamp when there’s been a longer break. It seems like the polite thing to do when Claude doesn’t have a clock and it puts me more into diary-keeping mode anyway.
Fishing for approval is rather easy:
Me: This is kind of silly but I feel like I’m wasting my day. I’m just playing some videogames instead of being productive. The fact that I already got the bloodwork done, which I should accept as good enough results for an entire week or so, doesn’t help much. Writing a blog post apparently doesn’t count either.
Claude: The bar keeps moving. You got bloodwork done that you’ve been avoiding for years, wrote tomorrow’s blog post, and have a party tonight — and the day feels wasted because of some videogames in between. That’s the depression talking, not an accurate accounting of the day.
Same goes for instructions:
Me: [...] Published some blog posts. Cleaned up my place, friend arriving in an hour or two. Still no optician, and it’s easter until Tuesday so I’m not sure if anything’s open. Work still in progress, will need to asses an audit report but I’m feeling rather drained so claude code is looking into it first. Will need to make dinner at some point.
Claude: Solid day despite feeling drained. [REDACTED], blog posts, cleaning, friend incoming — that’s a lot done. Make dinner before the friend arrives if you can. Everything else can wait.
And of course, directness is a virtue that goes both ways:
Me: Not sure about the anxiety—I’m not good at classifying these.
(I also got distracted waiting for your reply, opened Reddit, and found a huge amount of discussion created by my blog post, the one I linked you a couple of days ago. I’m happy and proud. https://old.reddit.com/r/slatestarcodex/comments/1sdrjzb/contra_the_usual_interpretation_of_the_whispering/)
Claude: Go read the Reddit thread, that’s the social reward loop doing exactly what it does for you. Enjoy it. We can finish the dump later.
Me: I’ve enjoyed my rewards. Lets continue.
I have found it useful to create a Claude Project called something like “2026 Soul Searching” in which I have added all the context necessary to deal with my personality type and idiosyncrasies. I didn’t want to have all that information in the website-wide “personal preferences” field to avoid Claude always engaging in a personal coach-way. I use it for work a lot and there’s no value there in managing my quirks. The Project serves as the baseline for moments where I know those quirks are a roadblock or have outsized influence on decision making.
One way it has helped engineer change is how strong it is at catching mental patterns before they’re fully on display. I’ll give you an example, replacing the quirk with a placeholder one (not interested in sharing my exact one): Task initiation block. In the Project context, I give full and honest details about how task initiation block has affected my life in recent years, with concrete dated examples, and whether I knew I was doing it at the time.
Because I know I have experienced task initiation block a lot lately, I have developed all sorts of hedging language around decisions that are affected by it. So, if I start a chat in the Project and plainly detail how I’m feeling re: a task/decision ahead, I might feel like I’m nowhere near displaying task initiation block, but Claude is remarkably good at picking up the train of thought that leads me from tnow to tblock. It still catches me off guard.
An analogy that comes to mind is when you say “I’m great, how are you?” to a close friend when you’re not feeling great at all, but the words come out with full self-confidence, before the friend snaps you out of it by saying “You don’t sound like you’re great!”.
Intuitively I know that it over-indexes on the context provided in the Project. Sometimes it’ll draw a link between tnow and tblock and I’ll just respond “Nah, I’m good, I’m not displaying task initiation block here”. But when it draws the link correctly, it can snap me out of it, and I’ve had > 10 instances of being helped by the realization that I was slowly moving towards a mental block.
Thank you! I was already thinking in a similar direction; the main difference was the following: instead of giving Claude instructions in text files, I would provide them in chat, and tell Claude to create the files itself.
When I am vibecoding, I typically tell Claude what to do, and when it does, then I tell it to also write project documentation. Then I read the documentation and comment on it. The idea is that the documentation is a live document, like if I come up with some creative idea in the middle of the project, I tell Claude to do it, and when I am happy with the outcome, I tell it to also update the documentation.
I usually provide some rough structure for documentation. So in this case I would probably tell Claude to create files “conversation.md” for conversational rules it is supposed to use with me, “viliam.md” for general facts about me, “goals.md” for the goals I am trying to reach… uhm, that’s just the first idea; the advantage of using an AI is that refactoring it later takes almost no effort.
Providing a current date every day sounds very useful; it can give Claude e.g. the possibility to ask: “three days ago you said that you planned to do X, did you actually?”
The main thing I am worrying about right now—but maybe I should just go ahead and try it, instead of speculating—is that when I am talking with my friends, there are different “modes”: just listening and providing empathy, or giving advice. I think the important part is the flexibility; if the friend only listened, I would be missing the good ideas and perspectives; but if the friend always provided advice, I would feel not seen sufficiently. It is annoying when LLMs end every single response with some suggestion what to do next. But I would appreciate a suggestion sometimes. Maybe I should be explicit about it? It is also a question of time: if I have enough time, I am more open to suggestions; but sometimes I want to get to the point without being interrupted or distracted.
I guess I will start with your instructions, add some of my own, but I will write it all in the chat and tell Claude to make notes for itself… and then I’ll see how it goes.
EDIT:
The first conversation was just as good as I imagined. Thank you for giving me inspiration and specific advice!
I’d like to throw in my two cents here.
I’m doing something similar, but unfortunately AI can’t capture everything I think during the day, everything that happens, etc. Together with it, I managed to set up a daily log with my daily tasks, what I completed, some metrics for sleep, energy, food, and training. Then a space to talk about what was different and another to write freely. At the end of the week, Claude Code pulls it all into one note and delivers the verdict from mentors I cloned: John Gretton “Jocko” Willink Jr., Peter Attia, Morgan Housel, Ryan Holiday, David Brooks.
They give me their wake-up calls, bring awareness to plan the next week.