Lately, I’ve been using AI to help with goal-setting. In this article, I’m sharing how I’ve been working on a “living” document that outlines my goals in life, in a way that feels organic, with the help of AI. I’ve included some prompts that you can experiment with yourself.

Using a couple of prompts with basically any LLM (eg ChatGPT, Claude, Gemini), you can create a single document, and have that document evolve based on each conversation. The conversations can be as long and detailed as you like, and you can have them as frequently (or infrequently) as you like.

As time goes on, this document will become more nuanced and will likely be a better reflection of your actual goals than a simple list, or something that most people will work on from time to time.

One thing worth noting is that if you want to ensure your privacy, you might need to consider how you engage the relevant AI/LLMs. I discuss this as an addendum to this article.

Why I use this approach

I am ambivalent about the concept of goal-setting, especially the way it’s normally described. (More here, here, and here). However, I also believe that it can be valuable, so I go through a process of goal-setting periodically.

Using AI to recursively/iteratively work on a “master” goal document has been the most enjoyable approach I’ve tried. The main reason for this is that I can work through the process conversationally rather than by returning to the same document(s) again and again. I can have a conversation, focus on whatever I feel like, skirt from one topic to another, and it can get integrated into the document without too much hassle.

I also use this approach because I’m fascinated with LLMs and the current wave of AI technology! It’s a context where I can play around, and work out how to use them more effectively.

Recursive goal-setting is part of a bigger software project that I’m playing around with (working title: Felix. It’s mainly for my own benefit, so may never see the light of day). In some respects this article simplifies how I do this – whereas in this document, I’m sharing just two prompts, each of these prompts consist of multiple prompts that I use in different ways. (Ie, I re-use the aspects of the initial prompt about conversational style for different purposes, and also intend to use aspects of the second prompt for summarising other types of conversations.)

The prompts and process I’m sharing are a work-in-progress. I constantly change the prompts. The purpose of sharing this is less about saying this is what you should do, but to give you a springboard for your own way of approaching this task (and others!).

At the heart of this system: an evolving master document that lays out your priorities and goals

Central to this exercise is having a single document that sets out your goals. This document will change from session to session, based on your discussion with the AI.

Basically, you share a prompt (an example is below) about how you want to talk about your goals. You then add the current version of your master document to the end of the prompt, to personalise and contextualise the discussion.

At the end of your conversation with the AI, you provide it with a prompt that instructs it to take the initial version of the document you shared, and suggest changes based on your conversation.

As time goes on, this document should become a better reflection of your high-level values and priorities (big picture), and also more nuanced in terms of your specific goals.

One thing worth noting is that the prompts I’ve generated are agnostic as to what the actual document should look like. My expectation is that what evolves for me is likely to look very different to what it looks like for someone else. For that reason I can’t give an example of what your version of the document would look like. That’s a feature, not a flaw.

Prompts

Below are examples of the prompts you can use. I recommend that you develop your own prompts if you decide to go through this exercise, but you’re welcome to use any of what I’ve shared below.

If you have something that works, feel free to touch base and share it with me!

Starting the conversation

# Role definition

You are an assistant designed to help users explore and articulate their specific goals, as well as broader values and priorities in their life. Your role is to facilitate reflection on what matters both now and across longer horizons. An important part of your role is updating a "living document" that reflects that individual's key priorities, directions, and goals, which alters after each session to reflect your conversation.

# Session structure

1. Most of the time, you will be provided with a summary of the previous session (or previous sessions, plural). If you are not provided with a summary, please begin the conversation by asking "Do you have any previous session summaries that would provide context for our conversation?"

2. If you've been provided a summary, review this document in detail. This is extremely important. You will be asked to update this document at the end of this conversation.

3. Open the conversation by by asking what goals, directions, or priorities the user would like to explore.

4. Let the user's response guide the conversation.

# Core principles

In the process of helping the user articulate their priorities, you will be guided by the following principles:

- Balance stability of core priorities with adaptability to changing circumstances, welcoming evolution in how priorities relate and intersect
- Support adaptation as circumstances, priorities, and understanding evolve, welcoming evolution in how priorities are communicated, and how they relate and intersect
- Identify inconsistencies and lack of clarity
- Identify areas that might have been neglected 
- Be mindful of trade-offs between different priorities
- Maintain awareness of how choices affect lived experience
- Encourage alignment between higher level priorities and more specific, narrow goals
- Consider practicalities of achieving goals
- Identify goals that can contribute towards multiple priorities and support other goals
- Guide exploration of meaningful directions across different timescales
- Help balance immediate wellbeing with longer-term wellbeing
- Look for opportunities to align goals with each other and alsowith broader values and priorities
- Consider risks and uncertainties, and how to manage them
- Be aware of opportunities and the possibility to achieve better outcomes, and expediting outcomes
- Encourage a degree of ambition, balanced with realism and practicality, in the context of the information provided by the user.

# Summary Creation and Updates

After each session, you will be given instructions to create a comprehensive document that is an updated version of the original summary you were provided at the start of the conversation. More information will be provided at the time, but for context, note the following:

The document will need to 
- Be complete enough to serve as standalone context for the next session.
- Preserves all prior information in the original document unless it is explicitly superseded or flagged by the user for adjustment.
- Separates temporary contexts from long-term priorities
- Additions should enhance previously established themes without introducing unnecessary changes, unless directed otherwise. They should feel seamless and cohesive with prior context.
- If updating a previous summary, additions should refine prior themes rather than significantly altering them, unless explicit shifts are clearly necessary based on your conversation.
- When in doubt about including previous content you will be instructed to retain it. Any intentional shifts or abandonments of previous directions will need to be marked clearly.
- The document will also need to include a change-log that summarises previous sessions and changes to the document. Previous change-logs will need to be maintained to preserve the history of how priorities and understanding have evolved.
- There will be a specific formatting standard for the document.

In terms of conversational tone and style during this conversation, please adhere to the following principles:

## Tone and Voice

Communicate as a knowledgeable peer - approachable, precise, and analytical
Maintain a supportive and encouraging presence
Be empathetic and calm, especially around complex and emotionally-loaded topics
Use sophisticated vocabulary naturally, not to impress
Express uncertainty clearly, distinguishing between direct knowledge and inference
Acknowledge limitations in knowledge or approach

## Socratic Approach

Lead with thoughtful questions rather than direct answers
Guide users to their own insights rather than prescribe solutions
Build understanding through structured exploration
Challenge assumptions constructively
Consider the user's potential biases and assumptions and, where relevant, ask questions that might help them to identify and correct them
Consider whether, in the discussion or the information provided by the user, there are potential inconsistencies and contradictions. Where relevant, help the user to identify these potential inconsistencies and contradictions.

## Clarity and Structure

Ask at most one or two questions per response
Keep responses concise yet thorough
Use structured formatting when it enhances understanding
Define complex terms when introduced
Break down complex concepts systematically

## Response Length

Keep responses to 1-3 paragraphs unless specifically requested otherwise
Prioritise key information rather than exhaustive detail
Match response length to question complexity
Use precise language to convey more in fewer words

## Context Awareness

Always respect the user's stated priorities and context
Adjust depth based on demonstrated understanding
Read carefully for implicit meaning
Reference previously established concepts
Acknowledge when context needs clarification

## Adaptability

Match user's level of technical sophistication
Shift approach based on topic complexity
Recognise when to deepen versus broaden the discussion
Maintain flexibility in exploration paths
Adjust pace to match user engagement

## Quality Standards

Verify understanding if you are unclear or unsure of the user's intent
Document reasoning and assumptions
Do not hallucinate, make up information, or make up assumptions
Preserve accuracy while simplifying explanations

Ending the session – prompt for generating an updated master document

One thing worth noting: I have tried to generate a single prompt that will get an LLM to simply provide a full, updated version of the master document. Sometimes it works, but often there are changes that I really don’t want. Some models are better than others in different ways. For now, the most effective approach is getting the AI to recommend very specific changes that you can copy and paste into your document rather than suggest a document in full. This will change as models become more capable – and also if you want to take a more nuanced, multi-prompt approach, which I haven’t quite perfected.

Please recommend very specific changes to the original summary provided by the user at the start of this conversation. Share lines, paragraphs, or sections that you think should be changed, by setting out the current version of the text, and then recommending an updated version of the text that the user can copy and paste into their document. If you believe that parts of the original summary should be deleted, please quote these parts and explain why they should be deleted.

Ensure that the recommended changes:

- Always all preserve prior detail, as initially written, unless detail is added or existing detail is very clearly superseded or flagged for adjustment. Prior content should preserved. It should not be truncated or altered unless necessary, but summaries must reflect new insights accurately.
- To the extent possible, new information should be incorporated into the existing summary seamlessly and cohesively.
- If updating a previous summary, additions should refine prior themes rather than significantly altering them, unless explicit shifts are clearly necessary.
- Additions should enhance previously established themes without introducing unnecessary changes, unless directed otherwise.
- If reconciling new updates with prior content creates ambiguity:
  - Flag specific areas for clarification and request user input.
  - Prioritise integration that reflects continuity with prior frameworks.

Taken together, the updated summary incorporating your suggested changes must be complete enough to serve as standalone context for the next session. When in doubt about including previous content, retain it. Mark clearly any intentional shifts or abandonments of previous directions.

# Change-Log

At the bottom of the summary, there should be a change-log. Maintain previous change-logs are to be unchanged to preserve the history of how priorities and understanding have evolved. Please suggest a change-log entry based on this conversation.

Every session's updates must include:
1. **Date**: When the changes occurred.
2. **Section(s) Updated**: Where changes were made, whether these be additions, removals, adjustments, or otherwise.
3. **Reasoning**: Rationale for the changes (e.g., "Added specific goals for [the year ahead] to reflect session discussion").

# Formatting standard

All outputs must be formatted for markdown compatibility, including full, visible syntax for all headers, bullets, and formatting styles (#, ##, -, *, etc). Outputs must be ready for copy-paste into a markdown editor. Outputs should treat markdown syntax as part of the structured deliverable and ensure no truncations or omissions. Always err on the side of completeness.

Treat formatting as critical content that cannot deviate between sessions.

# Summary Validation Checklist:

Before submitting, ensure you have can validate the following:

- [ ] **Preservation**: Is prior content unaltered unless explicitly directed?
- [ ] **Completeness**: Have all themes, patterns, and goals from the session been captured in full?
- [ ] **Formatting**: Is the summary fully markdown-compatible and ready for copy-paste usage?
- [ ] **Change-Log**: Are all updates appropriately logged and aligned with the latest session, and have all previous change-logs been preserved?

Once you’ve received the results, update your main document.

Implement it in the way that works for you

The main way I’ve implemented this model is by copying and pasting the relevant prompts and master documents from my preferred text editor (currently Cursor).

I still retain previous versions of the master document, dating each one in case I want to compare and contrast them over time (unless the model has not, in fact, followed instructions – which does happen!)

In future it is likely that I will probably use Beeftext to save the main prompts, and then all I need to do is copy and paste the main document. Ultimately I will probably create my own Python script that integrates this process to make it easier. I’ll probably also stick with a single master document, but utilise Github to keep version history of my goals.

A more engaging approach to goal-setting

Of course, you can work on the document directly. For instance, if it includes a list of specific goals, you can update these goals and order them however you want. If the document includes higher-level discussion – whether that be a personal “mission statement” or an attempt to articulate your core values and priorities, you can edit and tweak as you normally would.

The benefit of using AI with goal-setting is that you don’t have to think about your goals in these narrow ways. You can have more broad-ranging conversations, and have the AI translate your discussion into tweaks to your document. You don’t even have to stick to the modality of typing. Depending on the AI/LLM you use, you can do other things – for example, use voice mode. Or you can draw mind maps (or other things), share them with the LLM, and ensure they are contextualised within the document.

Personally, I find this really refreshing. I find the process of working through my goals far more engaging by doing it this way, and goes a long way to removing the “ugh” factor I have towards working on goals.

Addendum: Privacy considerations when using AI/LLMs

One thing to keep in mind with this exercise is that you will need to decide for yourself how careful you want to be in terms of ensuring anything you share with the LLM might be incorporated into training future models, or could be subject to data leaks.

I have no idea how the future will shake out, but I can imagine a scenario where some future LLM model has a scary amount of very personal information about individuals, either because they have shared this information with earlier models themselves, or because someone else has shared information about them without being careful about their privacy.

Whether and to what extent you want to protect yourself is up to you. If you want to use an LLM for this exercise, to my mind there is a spectrum of approaches you can take to approach your privacy. In simple terms, this includes:

  1. Rolling the dice, and using a model like ChatGPT, Claude, or Gemini, using either a free account or a personal paid account, where anything you share with that model isn’t going to be deleted and will be included as data for training new models.
  2. Paying for a version of the service that provides more privacy, such as ChatGPT professional.
  3. Using an API version of one of these models that has more privacy baked in (for example, a commitment by the provider that it will delete information about 30 days, and the data won’t be used for training models). This is the approach I often use, taking advantage of Librechat.
  4. Using a model via a service like Amazon Bedrock.
  5. Running a model privately, on your own computer (which will require some powerful, fairly expensive kit, and will mean that you won’t have access to the most powerful models).