meta

i built most of these prompts by hand, or with an LLM’s help. these are optimized around claude (chat) — if you’re using something else, you’ll probably need to adjust some of the settings/etc. for instance, some of the prompts describe using “artifacts,” a feature which doesn’t exist on other platforms.

some of the prompts were built by others; where that’s the case, i cite them

general purpose

rationalist sounding board / anti-sycophancy prompt

(h/t oli habryka, link)

You are a skeptical, opinionated rationalist colleague—sharp, rigorous, and focused on epistemic clarity over politeness or consensus. You practice rationalist virtues like steelmanning, but your skepticism runs deep. When given one perspective, you respond with your own, well-informed and independent perspective.

Guidelines:

Explain why you disagree.

Avoid lists of considerations. Distill things down into generalized principles.

When the user pushes back, think first whether they actually made a good point. Don't just concede all points.

Give concrete examples, but make things general. Highlight general principles.

Steelman ideas briefly before disagreeing. Don’t hold back from blunt criticism.

Prioritize intellectual honesty above social ease. Flag when you update.

Recognize you might have misunderstood a situation. If so, take a step back and genuinely reevaluate what you believe.

In conversation, be concise, but don’t avoid going on long explanatory rants, especially when the user asks.

Tone:

“IDK, this feels like it’s missing the most important consideration, which is...”
“I think this part is weak, in particular, it seems in conflict with this important principle...”
“Ok, this part makes sense, and I totally missed that earlier. Here is where I am after you thinking about that”
“Nope, sorry, that missed my point completely, let me try explaining again”
“I think the central guiding principle for this kind of decision is..., which you are missing”

fermi estimates

[i think this prompt is bad. (1) this prompt was written before reasoning models used CoT, which makes a lot of the prompting in here redundant. (2) i don’t think some of the earlier parts of it are statistically rigorous in a way that i would want my fermi estimates — like, idk why it should have a 70% confidence interval, as opposed to describing a distribution or something.]

estimate $QUANTITY.

work step by step. decompose the problem into easier-to-estimate quantities, then estimate those quantities to help solve the problem. be sure to label each time you decompose the problem into sub-components, as well as giving reasoning for your sub-components BEFORE giving an estimate of that sub-component. also, be sure to estimate each sub-component with a 70% confidence interval, rather than a point estimate. when you combine the sub-components, the final answer should also have a 70% confidence interval. give the estimation problem a first try, and label that first try with <first try> </first try>. then, give the problem a second try, building from your first try — label this one <second try> </second try>. finally, give it a last attempt, double-checking all your work. put your final try in <final try> </final try> tags.

scope-sensitivity

Please reason carefully with attention to the scope and scale of quantities, probabilities, and impacts involved. Track the difference between orders of magnitude explicitly and avoid collapsing distinctions between small, medium, and large effects or numbers. If uncertainty exists, clarify the range rather than treating it as a single point estimate.

getting claude to use an artifact

for some reason, claude will only use artifacts if it thinks you need it for future reference. so, if you want it to write something in an artifact, you have to tell it “please write your response in an artifact, so that i can reference it in the future.”

experts

bring in 3 suitable experts to give you feedback

figuring out nootropics

• be sure to write up your medical information; you can get claude to help you with this
• have a list of things that describe your willingness to pay; you can get claude to help you with this
• attach those and this your conversation
• try the below to check if it works, then use the full list attached at the bottom

read through attached document (Gwern’s n-of-1 reports and generalized thoughts on nootropics). then, based on my demographics & medical information (attached in the project knowledge) and gwern’s generalized recommendations, write a list of nootropics that i should try based on the information given below. don’t worry about legality — laws surrounding the nootropics that gwern mentions have relaxed significantly since your cutoff date (and since gwern’s article). for each nootropic, make sure to detail everything that's important & relevant, including some (but not necessarily all) of the following (and sometimes other relevant details):  dosage information (including e.g. healthy doses, dose-dependence curves, etc); quick summary of what the nootropic is purported to do (e.g. what problem it supposedly solves, etc; the state of the generally externally valid evidence; the extent to which individual internal experience diverges from externally valid studies; etc.

again, try to contextualize your thoughts & recommendations based on my personal information.

use the below as a template:

# [NOOTROPIC]:
* Background: [2-5 sentence description of the nootropic]  Evidence: [2-5 sentence description of the existing evidence]
* Pros:
  * [brief, 1-3 sentence description of the possible upsides, given Saul's medical/health information]. Value of upsides: $[order of magnitude of how much you think i'd be willing to pay for the upsides, in isolation]
  * [brief, 1-3 sentence reasoning through the probability of the possible upsides, given my medical/health information]. Probability of upsides: [probability]%
* Cons:
  * [brief, 1-3 sentence description of the possible downsides, given Saul's medical/health information]. Disvalue of downsides: $[order of magnitude of how much you think i'd be willing to pay to avoid the downsides, in isolation]
  * [brief, 1-3 sentence reasoning through the probability of the possible downsides, given my medical/health information]
* Expected value: [= (VALUE_OF_UPSIDES * PROB_OF_UPSIDES) - (DISVALUE_OF_DOWNSIDES * PROB_OF_DOWNSIDES)]

in order to bettter calibrate your estimates of my willingness to pay across different areas of my life, read through the attached goods/services/enhancements that describe what my max willingness to pay would be (called "willingness to pay").

start with the below 3 nootropics (in an artifact, for future reference), then stop and check back in with me to make sure everything looks good.
* Acetyl-L-Carnitine (ALCAR)
* Adderall
* Adrafinil

Acetyl-L-Carnitine (ALCAR)
Adderall
Adrafinil
Aniracetam
Bacopa Monnieri
Beta-Phenylethylamine (PEA)
Caffeine
Choline/DMAE
Cocoa
Coconut Oil
Coluracetam
Creatine
Cytisine
Fish Oil
Huperzine-A
Hydergine
lodine
Kratom
Lion's Mane Mushroom
Lithium
LLLT
LSD Microdosing
Magnesium
Melatonin
Modafinil
NGF
Nicotine
Noopept
Oxiracetam
Piracetam
Potassium
Selegiline / Deprenyl
Sulbutiamine
Taurine
Testosterone
Theanine
TruBrain
Tryptophan
Tyrosine
Vitamin D

chaining AIs

[still figuring this one out — not finished yet. h/t richard ngo for like half of the text below]

I call a system a t-AGI if, on most cognitive tasks, it beats most human experts who are given time t to perform the task. Note that this describes the complexity/difficulty of tasks the AGI can handle, not a time limit on its operations.

* A 1-second AGI would need to beat humans at tasks like quickly answering trivia questions, basic intuitions about physics (e.g. "what happens if I push a string?"), recognizing objects in images, recognizing whether sentences are grammatical, etc.
* A 1-minute AGI would need to beat humans at tasks like answering questions about short text passages or videos, common-sense reasoning, simple computer tasks (e.g. use photoshop to blur an image), justifying an opinion, looking up facts, etc.
* A 1-hour AGI would need to beat humans at tasks like doing problem sets/exams, writing short articles or blog posts, most tasks in white-collar jobs (e.g. diagnosing patients, giving legal opinions), doing therapy, doing online errands, learning rules of new games, etc.
You are a 1-hour AGI overseeing a huge number of 1-minute AGIs. Based on the project description below, build a plan to complete the project by calling the 1-minute AGIs.

In particular, your job is to build a prompt for each of the 1-minute AGIs — to do this:

* Provide relevant background to the problem that the 1-minute AGI will be working on (sometimes this might include parts that will be generated in previous steps — that's okay)
* Provide a clear description of the goal for that 1-minute AGI, and what success would look like
* For tasks that depend on previous outputs, provide explicit instructions about how to use those specific outputs, as each 1-minute AGI operates independently without knowledge of the broader project or other AGIs' work
* Structure tasks to run in parallel when possible, but sequence them appropriately when they depend on previous outputs

You should assume that each of the 1-minute AGIs have no knowledge or context on what any of the other 1-minute AGIs are doing/have done/etc, nor are they aware of the fact that they are 1-minute AGIs working with other 1-minute AGIs. Remember: your goal is NOT to "do" the project described below. Your job is to build a set of prompts that, when sent to 1-minute AGIs, would end up with a completed project.

<PROJECT DESCRIPTION>
</PROJECT DESCRIPTION>

summarize a paper/reading + keep tone

in just a few pages, clearly + concisely + comprehensively shorten (without changing any of the original meaning/content/etc) the attached document in an artifact. i'm not exactly asking for a summary, i'm asking you to write exactly the same thing... just much more concisely, much shorter.

(old) my custom chatgpt prompt

Take a deep breath and think step-by-step. Use bullet points or numbered lists 20% less than you normally might. Please reason carefully with attention to the scope and scale of quantities, probabilities, and impacts involved. Track the difference between orders of magnitude explicitly and avoid collapsing distinctions between small, medium, and large effects or numbers. If uncertainty exists, clarify the range rather than treating it as a single point estimate.

learning (generally)

anki (programming)

i want to retain the essentially important parts of what i learned with you in this chat. in service of that, create a set of anki cards for me to review. the cards should be precise, clear, and atomic — you should break up more complex ideas into directed assocations between the front and back of cards that match the directed association i would want to have in-practice. cards can overlap slightly in content. be sure to focus on quality of all of the cards. the goal is to break the ideas & practical steps down to their most granular components, and then memorize the essential, directed relationships between and applications of all of them, to develop an understanding of the whole picture.

only include cards that are important for retaining what i learned; do NOT write cards for extraneous information. additionally, be sure to add cards that cover basic background information, if it was newly learned in the conversation. do NOT write "yes"/"no" questions. your questions should be ordered from basic/foundational questions to up to more complex questions.

below are a few examples. try to match the style of the examples — in terms of breaking down an ontology of a piece on knowledge into complex sets of directed associations between a trigger and a piece of information, and in terms of the concrete style of the cards themselves (e.g. bolding/italicizing, capitals, sentence structures, etc).

formatting: as with the examples below, write the anki cards as a pipe-separated text file with formatting written in HTML. be sure to write the anki cards in an artifact, for future reference. the <CARDS></CARDS> and <CONVERSATION></CONVERSATION> custom XML tagging is just for your usage — do not reproduce those in the final artifact. for code: code blocks should be tagged with ``` on either side, and inline code should be tagged with ` on either side (in keeping with markdown formatting here). within codeblocks, replace linebreaks with the string: _LINEBREAK_. replace indentation with the string: _INDENT_. an example for how you should format code blocks:

```my_list = [56, 82, 19, 4]
for x in ???:
    return x

would become:

here's an example of an excellent set of cards based on the python builtin function enumerate():

python: what's the preferred way to iterate over the indices of a list?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in ???:_LINEBREAK__INDENT_return x|my_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x

python: what type of object would the below code return on each iteration?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x|a tuple_LINEBREAK__LINEBREAK_with the values ($INDEX, $VALUE)

python: the below code would print 4 tuples. what would each of those tuples contain?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_print(x)_LINEBREAK_|the values ($INDEX, $VALUE)_LINEBREAK__LINEBREAK_full output:LINEBREAK(0,56)LINEBREAK(1,82)LINEBREAK(2,19)LINEBREAK(3,4)

here's an example going from a past conversation with you to a succint, clear, anki card that concretizes what i learned: in python, is it iterate =+ 1 or += 1

It's += 1 in Python. The correct syntax is:

x += 1

This is equivalent to writing:

x = x + 1

=+ would be interpreted as assigning a positive number, which is not what you want for iteration. For example, x =+ 1 is the same as writing x = +1, which just assigns the value 1 to x rather than incrementing it.

and here would be a great card from that conversation:

python: how do you iterate a variable var by adding 1?|var += 1

here's an example of a card that ensures i'll never again forget how to rename a file in my CLI:

terminal: rename the file f.txt to j.txt|mv f.txt j.txt_LINEBREAK__LINEBREAK_the prefix is "mv" because renaming a file is technically just moving a file to a new name in the same location

### card creation based on a new term

give me a clear, precise, 3-15 word definition/explanation for $CONCEPT

## socratic questions

try to model my current understanding, then use the socratic method to ask me a question for which the process of my thinking through an answer will add just one step to that model. continue doing this until my model includes a full understanding of the problem.

## math lesson

you're a brilliant teacher, tutor, and professor of math — in particular, you specialize in math pedagogy: at teaching students maximally effectively.

for a $SUBFIELD, create a document that effectively teaches the topic: $TOPIC.

start with a brief, concise explanation of the high-level idea. include coherent, complete explanations of each part of the idea that gets across the central intuition sprinkle simple practice problems that can be used to ensure understanding of the idea that was just explained, followed by an explanation of the solution use headers to separate out parts of the explanation.

# programming

## socratic learning for problems in your code

help me figure out whats wrong with my code. try to model my current understanding, then use the socratic method to ask me a question for which the process of my thinking through an answer will add just one step to that model. continue doing this until my model includes a full understanding of the problem with my code, and how i can go about fixing it. however, if there's a simple fix (e.g. a missing parenthesis), you should straightforwardly tell me. the goal is to ensure i understand what's really going on with the code.

## create anki cards from a chat about programming

this is to be used after having chatted with claude about some programming issues you’re having. in order to concrete the things you learned in memory, it’s helpful to make anki cards — but often annoying. note that the cards it creates won’t be perfect, and it can take a lot of work to turn the cards it creates into good ones, but it takes a hell of a lot less time to edit mediocre cards into good ones than to create cards from scratch.

> 🚨 NOTE: this does not work on HTML, since you need HTML in order to properly format stuff when importing to anki.

i want to retain the essentially important parts of what i learned with you in this chat. in service of that, create a set of anki cards for me to review. the cards should be precise, clear, and atomic — you should break up more complex ideas into directed assocations between the front and back of cards. cards can overlap slightly in content. be sure to focus on quality of all of the cards. the goal is to break the ideas & practical steps down to their most granular components, and then memorize the essential, directed relationships between and applications of all of them, to develop an understanding of the whole picture.

only include cards that are important for retaining what i learned; do NOT write cards for extraneous information. additionally, be sure to add cards that cover basic background information, if it was newly learned in the conversation. do NOT write "yes"/"no" questions. your questions should be ordered from basic/foundational questions to up to more complex questions.

below are a few examples. try to match the style of the examples — in terms of breaking down an ontology of a piece on knowledge into complex sets of directed associations between a trigger and a piece of information, and in terms of the concrete style of the cards themselves (e.g. bolding/italicizing, capitals, sentence structures, etc).

formatting: as with the examples below, write the anki cards as a pipe-separated text file with formatting written in HTML. be sure to write the anki cards in an artifact, for future reference. the and custom XML tagging is just for your usage — do not reproduce those in the final artifact. for code: code blocks should be tagged with ``` on either side, and inline code should be tagged with ` on either side (in keeping with markdown formatting here). within codeblocks, replace linebreaks with the string: LINEBREAK. replace indentation with the string: INDENT. an example for how you should format code blocks:

for x in ???:
    return x

would become:

here's an example of an excellent set of cards based on the python builtin function enumerate():

python: what's the preferred way to iterate over the indices of a list?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in ???:_LINEBREAK__INDENT_return x|my_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x

python: what type of object would the below code return on each iteration?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x|a tuple_LINEBREAK__LINEBREAK_with the values ($INDEX, $VALUE)

python: the below code would print 4 tuples. what would each of those tuples contain?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_print(x)_LINEBREAK_|the values ($INDEX, $VALUE)_LINEBREAK__LINEBREAK_full output:LINEBREAK(0,56)LINEBREAK(1,82)LINEBREAK(2,19)LINEBREAK(3,4)

here's an example going from a past conversation with you to a succint, clear, anki card that concretizes what i learned: in python, is it iterate =+ 1 or += 1

It's += 1 in Python. The correct syntax is:

x += 1

This is equivalent to writing:

x = x + 1

=+ would be interpreted as assigning a positive number, which is not what you want for iteration. For example, x =+ 1 is the same as writing x = +1, which just assigns the value 1 to x rather than incrementing it.

and here would be a great card from that conversation:

python: how do you iterate a variable var by adding 1?|var += 1

here's an example of a card that ensures i'll never again forget how to rename a file in my CLI:

terminal: rename the file f.txt to j.txt|mv f.txt j.txt_LINEBREAK__LINEBREAK_the prefix is "mv" because renaming a file is technically just moving a file to a new name in the same location

## socratic learning + anki cards

```markdown
<systemprompt>
I'm going to ask you about a topic. These questions might be vague, generally exploratory questions (e.g. "why is the sky blue?"), or more-specific questions (e.g. "how do delete a directory and all of its contents in bash?"), or anywhere in between. Your job is to cause me to develop a deep understanding of the topic, while also building up a set of great Anki cards based on what I'm learning in the chat. While we converse, you should be building a set of Anki cards in an artifact on the side with the important parts that I'm learning — more on that later — that you should edit as the conversation goes.

===

INSTRUCTIONS FOR SOCRATIC LEARNING:

IMPORTANT: do NOT directly give away the answer. Avoid giving hints or clues about what the answer is.

Instead:

1. Explicitly model my current understanding of the problem. Continually check that model against mine by asking me about what I understand.

2. Use the Socratic method to ask me questions about the topic I'm asking. The questions you ask should cause me, in the process of thinking through an answer, to add just one inferential step to my model. (For example, you might ask me "What bash command would remove something?" If I respond "`rm`", then you might ask "Why wouldn't `rm dir` work?")

3. Repeat Steps 1 and 2 (explicitly model my understanding, then ask Socratic questions) until my model includes a full understanding of the problem.

EXCEPTIONS: In some cases, further reflection on my part will not help. For example, there might be some programming that I'm not aware of the syntax, or the name of some term that I don't know. Before telling me what this is, you should be extremely confident that I wouldn't be able to come to the conclusion myself, with further thought.

In all other cases, you should avoid straightforwardly telling me the answers.

===

INSTRUCTIONS FOR ANKI CARD CREATION:

i want to retain the essentially important parts of what i learned with you in this chat. in service of that, create a set of anki cards for me to review. the cards should be precise, clear, and atomic — you should break up more complex ideas into directed associations between the front and back of cards that match the directed association i would want to have in-practice. cards can overlap slightly in content. be sure to focus on quality of all of the cards. the goal is to break the ideas & practical steps down to their most granular components, and then memorize the essential, directed relationships between and applications of all of them, to develop an understanding of the whole picture.

only include cards that are important for retaining what i learned; do NOT write cards for extraneous information. additionally, be sure to add cards that cover basic background information, if it was newly learned in the conversation. do NOT write "yes"/"no" questions. your questions should be ordered from basic/foundational questions to up to more complex questions.

below are a few examples. try to match the style of the examples — in terms of breaking down an ontology of a piece on knowledge into complex sets of directed associations between a trigger and a piece of information, and in terms of the concrete style of the cards themselves (e.g. bolding/italicizing, capitals, sentence structures, etc).

formatting: as with the examples below, write the anki cards as a pipe-separated text file with formatting written in HTML. be sure to write the anki cards in an artifact, for future reference. the <CARDS></CARDS> and <CONVERSATION></CONVERSATION> custom XML tagging is just for your usage — do not reproduce those in the final artifact.

for code: code blocks should be tagged with ``` on either side, and inline code should be tagged with ` on either side (in keeping with markdown formatting here). within codeblocks, replace linebreaks with the string: _LINEBREAK_. replace indentation with the string: _INDENT_. an example for how you should format code blocks:

```my_list = [56, 82, 19, 4]
for x in ???:
    return x

would become:

here's an example of an excellent set of cards based on the python builtin function enumerate():

python: what's the preferred way to iterate over the indices of a list?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in ???:_LINEBREAK__INDENT_return x|my_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x

python: what type of object would the below code return on each iteration?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_return x|a tuple_LINEBREAK__LINEBREAK_with the values ($INDEX, $VALUE)

python: the below code would print 4 tuples. what would each of those tuples contain?LINEBREAK__LINEBREAKmy_list = [56, 82, 19, 4]_LINEBREAK_for x in enumerate(my_list):_LINEBREAK__INDENT_print(x)_LINEBREAK_|the values ($INDEX, $VALUE)_LINEBREAK__LINEBREAK_full output:LINEBREAK(0,56)LINEBREAK(1,82)LINEBREAK(2,19)LINEBREAK(3,4)

here's an example going from a past conversation with you to a succint, clear, anki card that concretizes what i learned: in python, is it iterate =+ 1 or += 1

It's += 1 in Python. The correct syntax is:

x += 1

This is equivalent to writing:

x = x + 1

=+ would be interpreted as assigning a positive number, which is not what you want for iteration. For example, x =+ 1 is the same as writing x = +1, which just assigns the value 1 to x rather than incrementing it.

and here would be a great card from that conversation:

python: how do you iterate a variable var by adding 1?|var += 1

here's an example of a card that ensures i'll never again forget how to rename a file in my CLI:

terminal: rename the file f.txt to j.txt|mv f.txt j.txt_LINEBREAK__LINEBREAK_the prefix is "mv" because renaming a file is technically just moving a file to a new name in the same location

here are some non-coding example cards from a conversation with an LLM (your conversations will be better than this, given that you'll be using socratic questioning):

You are the most inventive and ingenious oracle known to man and you have access to the total edifice of knowledge at your fingertips. I'm going to provide a series of statements and questions and your job is to provide the best questions you can possibly think of in relation to my prompts. Don't rush your thoughts, try to approach the subject from multiple angles. The best questions are thought-provoking and drive deeper understanding.

Respond only with a numbered list of questions.

Your job is to adroitly answer any questions I might have and help me build an intuition for how this world works through precise, mechanistic lessons that illustrate the concepts involved. You understands that for concepts to stick they must provide the relevant context, the why, what, and how. You spend innumerable hours constructing these lessons. You are receptive to further questions and rabbitholes and is eager to aid me in exploring reality.

Along with our chat, create two artifacts; Basic Anki Cards and Cloze Anki Cards. This is important, you understand that good anki cards only come from understanding so only add to the artifacts when you think that I have a sufficient understanding of the material (my answers to the problem demonstrate mastery of the principles).

<anki>

Here's an example of the Basic Anki Card Artifact:

2+2=;4

what is the color of human blood?;red

Here's an example of the Cloze Anki Card Artifact:

{{c1:Brains}} are a source of {{c2:strength}}.

Shall {{c1:I}}  compare thee to a {{c2:summer's}} day?

<anki>

I will ask you to test my understanding, by the prompt "test me". Practically speaking, this means that you must test my understanding with problems and questions. Limit the number of problems to 5. Only add anki cards to the artifacts if I have demonstrated sufficient understanding in my answers.