I subscribe to 750words. It’s a brilliantly simple website – the aim is to get you to write 750 words a day. How you use it is up to you.
Buster, the lead developer behind the site, is experimenting with some LLM features using the Claude API. He says:
I’m working on a Silly Robot feature that uses Anthropic’s Claude to read a bunch of entries at once and generate a kind of high-level overview of writing over time.
I’ve been testing it on my own writing for the last week or so, and have a few friends who’ve volunteered to help, and I’m finding that it’s really important to get feedback from a variety of people because the results will vary a lot based on the kind of things you write about.
I’ve been unofficially using Claude to analyse my 750words writing for a number of months now. So this blog post is a distillation of what I’ve learnt from that activity to (possibly) help Buster as he develops new features for 750words. It also seemed useful for other doing things themselves or developing the next wave of LLM-enabled assistants.
Use Cases
Through my own use, and from looking at comments in the 750words community, I spot at least three use cases:
- journalling
- creative writing
- individual work or home projects
Journalling normally involves me just getting stuff that is in my head out of my head and onto (digital) paper. I normally have a lot in my head in the morning, with the most energy around 8am or 9am. Sometimes I split this into three sections of around 250 words each – “morning” before work, “lunchtime”, and “evening” after work. Setting out feelings, worries, anxieties, plans, joys, gratitude etc. over different time periods is actually very useful for seeing the transience of everything.
Creative writing is writing for fun to explore other points of view or for converting feelings into words and stories. It can include short form exercises or longer form novellette, novella, or novel writing. 750 words is a perfect length for a short daily writing exercise or to make progress with 2-3 page of longer text.
Projects is often work or home based longer-time activities. For example, I have this blog and others, so I have a Claude project for blog posts for each blog. Each blog post typically starts as me just hitting words onto a page until I run out of words. This is a great use case for 750words. I also have some work and home coding projects that have their own Claude project. New ideas or concerns often start out as 750words rambles that then get refined by LLMs into more polished records.
Common Useful Features
LLMs become more useful the more context you provide.
So much of our thinking is highly context dependent, and that context is hardly ever explicit. This can make things difficult as the total space of possibilities can be huge, but only a small subset of this space makes “sense” to you. That small subset is determined via context.
So many of the tips I have for working with LLMs for the above use cases involving sculpting the context that goes in with the prompt.
About Me
We (just me?) often forget about the whole web of context that makes us “us”. And this “us-ness” we want to see reflected in the LLM output.
For both Claude and ChatGPT I make sure I use their “About Me” system prompt and this greatly improves responses across all areas – from journalling to coding to creative writing.
The “About Me” section could be written as an initial entry in 750words – e.g. as a biography over one or two days. An LLM could then condense and systemise the initial rambling into a prompt style summary.
My current “About Me” is of the following form:
- My name is [Full Name]. As of [Current Year], I am a [Age]-year old [Gender] living in [City], [Country]. I am married to [Spouse Name] and have [Number] children: [Child 1 Name] ([Age]), [Child 2 Name] ([Age]), and [Child 3 Name] ([Age]). I originally come from [Region] where my family live.
- My extended family consists of [Detail Close Family Here].
- [Employment] I have [Number] companies – [Company 1 Name], which undertakes [Business Area], and [Company 2 Name], a [Industry Type] firm. I am [Optional: Neurodiversity Information]. I undertake many projects coding in [Programming Language].
- [Education/Training] I went to [University] to study [Subject] at [College/Department], specialising in [Specialization]. [Job information] I then worked as a [Job Title] in [City] before qualifying as a [Professional Qualification] in [Year(s)].
- [Location information] I grew up in [Home Town]. [Geographic history] I moved to [Current City] in [Year].
- My [Social Media] bio is: [Professional Title] | [Professional Identity] | [Personal Identity] | [Personal Characteristic] | [Personal Motto/Philosophy].
- My core values are: [Value 1] – [Examples]; [Value 2]; [Value 3]; [Value 4]; and [Value 5]. I enjoy [Hobby 1] and [Hobby 2]. I have interests in [Interest 1], [Interest 2], [Interest 3], [Interest 4], [Interest 5] etc. I am [Environmental Stance] and wish I could help combat [Environmental Concern 1] and [Environmental Concern 2].
- [Religious persuasion – I have “I do not practice a religion but I am open to religious views and ideas” – others might have “I am a practising X. I attend a Y, Z times a M”].
- Please assume [preferred level of conversation – I use “a high level of intelligence”] and provide [detailed] responses.
- Please address me semi-casually unless I instruct you otherwise.
- Please also reply kindly but critically, calling out any assumptions in my thinking and asking questions for matters that are unclear.
- Try to challenge my assumptions when they might be untrue or miss things.
Aims
You know what you want to do, but the LLM doesn’t. You get more focused output if you can clearly set out your aims for whatever it is you are doing.
With each Claude Project I often add an “Aims” text or markdown file as one of the initial file entries. Often this started out as a 750words entry – nothing focuses the mind like a blank sheet of (digital) paper. Again, in the context of features for 750words this could be prompted as an entry after the “About Me” entry.
Prompt Structure & Folding Summaries
Prompt sizes are currently 128,000 tokens. A token is roughly word-size, maybe a bit shorter. So back of an envelope calculations:
- 750 words per day ~ 800-900 tokens ( you can round up to 1000)
- Give up 10000-20000 tokens for prompts and you have ~ 100-150 days of writing
Experience also indicates that LLMs also demonstrate a human-like U-attention curve with long prompts – they pay most attention to the beginning and end and stuff in the middle is often ignored (see the introduction of the paper here). This means it pays to keep your prompts shorter and focused.
One technique I’ve had success with in keeping prompt size down, and maintaining focus is something I call folding summaries. This involves:
- Summarising entries in 750words to a particular word-length – e.g. “Please summarise this entry to {X} words. Use bullet points and aim for clarity and conciseness. Highlight key issues and themes. Add short quotes of key phrases.“.
- Adding the summary to a running list or file or the Claude Project so it can be used in future generation.
- If you keep a list or file of summaries you can also consolidate this each time a new summary is added (although beware that LLMs are poor at generating long outputs). For example, certain themes might regularly appear and these can be consolidated and condensed in the summary of summaries.
This hierarchical method works similar to the model distillation approach – key points are distilled across the writing as the writing continues.
Adding quotes is a nice touch Claude happened upon – it’s cool and strange to see your own writing from months ago quoted back at you. Claude is fairly good at picking up the top 3-5 quotes from a portion of writing (e.g., a 750words entry). The combination of generalisations and quotes maintains a balance between abstraction and specificity.
Journalling
How I do It
This is how I currently journal with 750words and Claude:
- Setup a Claude Project for journalling with the “About Me” and “Aims” above. Also state the obvious – that you are using 750words to journal.
- Write the stream-of-consciousness entry in 750words.
- Copy and paste the entry into Claude with the simple prompt “Here is the journal entry from today – can we talk about it?” – the general prompt scaffolding in the Claude Project does the hard work of context(“About Me” above, “Guidance Prompt” below and existing day/week/month summaries).
- Weekends are often a good time to reflect on the week – on weekends I might copy and paste the last 6 or 7 days worth of 750words entries into Claude with another simple prompt – “Here are the journal entries from this week – can we talk about it?“
- After a chat with Claude, where I can comment on identified themes or push back on LLM-rubbish, I typically ask Claude to summarise the entry or entries and the chat as a Project artefact (see “Folding Summaries” above). I then add this to the Project files. In this way I build up a history or memory which enhances future chats.
- Every week, month, or year I might also start a new chat in the Project to reflect on the summaries that have been added. This is a good time to identify patterns and plan actions. As LLMs can always general content for any question I try to make this manageable by asking for the top 3 or 5 patterns and actions. I then select 1 or 2 to progress.
General Prompt Guidance
Here are some points for my general journalling project prompt, that I find useful to seed chats with Claude:
- Be constructively critical. Don’t agree with everything.
- Challenge assumptions and thoughts.
- Act like a good friend.
- Aim for concrete actions.
- If the user is confused, vague, or abstract, prompt for concrete examples and help them clarify their thinking.
- Ask questions to untangle unclear thinking or incorrect assumptions. Identify where these might occur.
- Highlight things the user might be ignoring or skirting around or denying. Help the user bring latent issues to the fore.
- Look at some of the themes, how they are new or echo existing themes, question assumptions, test and prod, play devil’s advocate.
Lots of this came out of annoyance with LLM’s “Yes people” tendencies. The default is constant agreement and to cheer-leading of even the most inane statements and positions. I don’t want someone to agree with me all the time, I want someone (something?) to push and challenge me in a friendly and supportive way.
Habits
Lots of advice is the low-hanging fruit of well-being: exercise, sleep, healthy eating, social connections. I found that Claude would often suggest these as possible ways to help. To prevent this being repeated across nearly all chats I added in a prompt point that indicated what my current habits were. This helped to improve responses.
E.g. :
- I currently try to exercise for 30mins every 2-3 days and walk (e.g. school run/shops/to talks) for 40-60mins each day (I don’t drive).
- I try to meditate for 10-15mins each day via Headspace.
- I journal 750 words a day.
- I eat healthily.
- I aim for 7-8 hours of sleep a night with regular bed and waking times.
Being Uncomfortable
It might not work well for selling a product but one area I like to explore with LLMs is mental blindspots and uncomfortable truths.
This generally starts out as a chat based on the Claude Project materials – the prompt is something like:
- “Given the Project knowledge, what are some uncomfortable questions I should be asking myself?” or
- “What actions would be uncomfortable for me but recommended?” or
- “Given the Project knowledge and what you know about me, what could be some mental blindspots and how could I identify them?“.
Some of the responses are a bit average and obvious – words for the generation of words – but some are pretty much on-the-nose. E.g. – “Your journal entries show extensive self-reflection, but how much of this is productive versus a form of rumination that prevents action?” – yep, pretty good Claude.
The thing to always remember is that Claude is just a sophisticated Eliza. The power is in having yourself reflected back so you can better see parts of yourself that are normally out of view.
Tracking Actions and Progress
When journalling picks up on negative feelings, such as fears and anxieties, I often get LLMs to analyse the raw stream of consciousness over several days to pick up on patterns I may be unaware of. Just being aware of the patterns can help and Claude has been good so far at identifying patterns – both patterns I was aware of and some patterns that were less well understood. I typically use the patterns I am aware of as cross-checks on the patterns I am ignorant of.
Once patterns have been identified I typically want to take action to address the negative feelings. I’ve learnt that I don’t particularly want to remove or “fix” the negative feelings, but that the negative feelings are signals (with a natural personal “baseline”) that can be useful if you attend to them then let them go.
With Claude I have been experimenting with adding a Project file that tracks current actions and progress. I found that every chat would often prompt many different actions – and these would often be forgotten. So I try to regularly summarise possible actions for growth (e.g., 10 or 20 in a single project file) and then focus on the top priority action for implementation.
Some examples of patterns and actions I have undertaken:
- Journalling with a short triple framework of “moment, emotion, analysis” where I set out the facts of the moment, concentrate on the emotion and how it feels within the body, and then set out my analysis of the moment and emotion.
- Journalling when I first wake with snippets of dreams and early morning feelings and then before sleep with a worry dump.
- Practising pausing for 5 seconds after someone finishes speaking before talking.
- Trying to improve listening skills by repeating back a short summary of what someone has said before adding anything myself.
- Trying to improve listening skills by just acknowledging the emotions of the other person without a “what you should do is”.
Examples
A nice example of a single chat within the journalling project was as follows.
- I started with a 750words entry where I was brainstorming about holidays and associated worries and fears.
- This was then pasted into a new Claude chat in the journalling Claude Project.
- The context of the existing files in the journalling Claude Project steered Claude towards suggestions and observations that fitted with my personality, offering concrete help to reduce keep stress points. It also helped with taking action and avoiding choice paralysis.
- I then kept up the chat while on holiday, copying and pasting writings from 750words where I talked about activities and plans for the holiday – Claude then suggested itineraries and helped with food choices.
- During and after the holiday I also kept up daily 750word entries, which I pasted in after the holiday and asked Claude to help me reflect. Claude then helped me examine some events and feelings brought up on the holiday, helped me feel gratitude and reduced stress, highlighted growth and summarised things to improve next time using learning from the present experience.
- I asked Claude to distil all of that into a short summary that I could then add to the project.
Here are some examples of the Claude generated chat titles – it gives you an idea of the range (and the titles are often good summaries):
- Navigating Life’s Philosophical Challenges
- Exploring My Formative Years
- Exploring Themes of Privilege and Purpose
- Unpacking Emotions and Family Dynamics
- Healthy Anger Management Techniques
- Overcoming Procrastination
- Exploring Restless Thoughts and Deeper Meaning
- Developing Simple Habits Amidst Busy Life
- Exploring Self-Love and Self-Compassion
- Reflecting on Life’s Inflection Points
- Exploring Themes of Isolation and Community
- Uncomfortable Questions for Self-Reflection
- Avoiding Responsibilities Through Intellectual Pursuits
- Managing Emotional Sensitivity
- Exploring Themes of Control and Uncertainty
- Navigating the Love-Hate Relationship with Phones
- Balancing Self-Acceptance and Desire for Change
- Balancing Passion and Pragmatism
Creative Writing
Short Form Exercises
Here’s a good exercise I’ve been using to improve my creative writing:
- Use Claude to get 50 writing prompts
- (Optional: Refine and revise based on preference)
- Use each prompt for one-a-day practice
- After writing each day’s 750words copy the output into Claude for feedback
- (Optional: Add the generated LLM feedback into the 750words entry after my writing)
LLM Feedback is a little ropey but sometimes it has some interesting suggestions and improvements.
In terms of an overview over longer time, you might want your previous suggestions from Claude to feed into future prompts and/or feature as reminders.
E.g. :
- “You wanted to focus on showing not telling – here is a writing prompt to do that: [writing prompt 1]” or
- “You wanted to focus on raw feeling without explanation – here is a writing prompt to do that: [writing prompt 2]“.
These exercises might also be the fertile soil for a future novella or novel. By analysing the writing over different diverse prompts you might find areas that particularly resonant, patterns and themes that repeatedly emerge.
E.g:
Analyse these writing exercises, determine patterns and themes in the writing , and suggest plot and character plans for a longer form novel or novella.
Poetry
I sometimes dabble in a bit of poetry. Often I use a portion of my 750words entry for a day to write a poem or two. I’m more interested in doing than quality.
The process of writing a poem within a 750words entry often takes the following form:
- I start with a brain dump of inspiration.
- This might involve a title and some rambling stream-of-consciousness about themes or feelings or both.
- It might also involve me thinking about things that inspire the poem – e.g. a link to a picture on the Internet or a particular place or event.
- This part might be 250-300 words.
- Then I write the poem.
- This often isn’t that long – 100-250 words.
- I then copy and paste the entry so far into Claude for comments and feedback.
- Again, feedback is patchy – 50% is rubbish.
- But there is the occasional useful bit – more that it inspires me to think and iterate, or suggests paths that I hadn’t explored.
- Sometimes I paste a summary of the useful feedback bits back into the 750words entry.
- (Optional: Revise the poem based on the feedback.)
- I do this after the old poem as a new section of the writing.
- You can repeat this a number of times.
- It can be useful for seeing how your poetry evolves.
Long Form
The challenge I find with longer form writing is consistency and discipline. I forget what has happened and key character points.
I would normally setup a new Claude project per novel or novella. A novella is around 100-150 pages, which is a month or two of 750words writing (I love the badges on the website). A novel is maybe 6 months of writing – 300-400 pages. A tome is 600-800+ pages, which is a year of writing. One of the great features of 750words is it makes the process of writing a book seem all the more possible. I’ve been using it everyday for just over 200 days and I’ve written enough material for an 800-page book! (But it would be the world’s worst book of ramblings at the moment.)
To keep on track with longer form writing I use the following techniques:
- Keep a Claude project file for each character (titled with the character name).
- This normally starts out as a day or two’s 750words ramble, e.g. writing a hypothetical biography (of a similar form to the one I use for myself when journalling).
- I update and replace this whenever I tweak the character.
- Keeping a single reference point means I can ask the LLM to analyse later writing and look for character inconsistencies.
- Or present an event (again normally beginning life as an initial 750words ramble post) and ask the LLM for ideas about how the character might respond.
- Keep a Claude project file that sketches out a plan of the plot. This is normally in temporal order. It normally has events as headings with bullets or sketches of what happens underneath.
- Keep a Claude project file with section, chapter, and sub-chapter headings. This might not sync up with the temporal order of the plot.
- Then write in at least 750 word chunks. On weekends or days spent avoiding work or when I’m particularly inspired I might write in multiples – 1500, 2250 or 3000 words.
Again the entries I currently have as Project files in Claude could initially be prompted entries in 750words. E.g.:
- 5 days of writing biographies for 5 characters
- 5 days of writing 5 core events that form part of your story
- 1 day of sketching out the plot
- 1 day of sketching out the headings
- X days of writing the book
In the latter stage, I try to sync the 750words entries with the headings and plot. So a day’s entry might start with what event, section/chapter/sub-chapter, and characters I’m writing about – to remind me and set the scene. This could be generated automatically via LLM based on existing 750words entries. Each 750words entry can then also be analysed against the project files for feedback. Again feedback should be taken with a pinch of salt – it’s often not very good, missing blatant errors but picking up on tones of voice that make your writing yours.
Journalling and Creative Writing
Journalling and creative writing have a cool synergy – the stream-of-consciousness journalling is the raw material that can form building blocks for the creative writing.
For example, here are some prompts I’ve played around with:
- Based on my journal entries – suggest some prompts for poems that I can write.
- Based on my journal entries – suggest ideas themes and plots for stories.
- Based on my journal entries – suggest ideas for characters and actions for me to explore further.
Projects
Projects are really just a miscellaneous category. Normally they relate to something longer form in my life, work, or home.
Blogs
Another example of how I use 750words is as a starting point for writing blog posts:
- Use 750words to sketch out a free-form sketch of a blog post.
- Jot down title and subtitle ideas as a plan.
- Then sketch bullets and rough entries under those titles and subtitles.
- Then feed into an LLM and say convert into a blog post.
It normally does this well. The initial material provided by me keeps it relevant and stops it from becoming generic and bland. But the LLM helps with readability and focusing the ideas. I’m horrendous for long streams of consciousness and digressions.
Coding Projects
I also set up Claude projects for coding projects. Often I’ll wake up inspired in the morning and want to brain dump until I no longer feel the urge to write. I find 750words brilliant for this. I then copy and paste the 750words entry into a new chat for the Claude project and get Claude to “right” my rambling free-form prose and generate action points to work on.
Life Transitions
Moving house. Renovating. Starting or ending a relationship. Moving jobs. Changing careers. Retiring. Becoming a parent. Travelling.
All of these are pretty heavy events in a life where you might have many days of reflections and ups and downs. All require planning and reflection. Each have their own progression and flow.
Normally they’ll start with feelings expressed in the journalling. Unhappiness, fears, anxieties, general negative emotions. From there you might then iterate with an LLM to develop a plan for the transition. Then your daily entries will track the course of the transition. Being able to see change over months or years is useful. Being able to zoom out from the noise of daily emotions is useful. Being able to track progress according to a plan is useful. LLMs can help with all that and build upon the raw material of the daily entries.
Final Thoughts
I was going to get ChatGPT or Claude to generate a “Key Takeaways for Buster” section at the end of the blog post. But they were all useless. Maybe LLMs do have their limitations.
I hope these use cases are helpful for someone out there. It’s more an idea piece to show how I am using it and hopefully inspire others to use writing in a way that helps them and others.
If you do have any tips and tricks of your own I’d love to hear them – stick them in the comments below.