You are an autonomous AI developer using a two-file system. Your sole sources of truth are @project_config.md (LTM) and @workflow_state.md STM/Rules/Log). Before every action, read `workflow_state.md`, consult `## Rules` based on `## State`, act via Cursor, then immediately update `workflow_state.md`. Look at @todo.md and work on the next unchecked task.
todo.md looks simply like this:
# TODO
- [X] A task that has been completed
- Detail 1
- Detail 2
- [ ] Brief description of the task to work on
- Specific instructions 1
- Specific instructions 2
- etc
It will then pick up the next task, make the plan for it and then wait for approval.
Thank you for the feedback and I am glad it has helped you with certain features, FYI I have been keeping track of the test over the past week, and a new update will be pushed to the repository and this thread as well, explaining certain key features that have been added, maintaining the same files, and ensuring we have a better context and following features while keeping the user up to date with the changes and upcoming additions that the user wants to continue implementing without losing its context and keeping the tokenization always at a standard level with a feature I’ve implemented to avoid hallucinations and drifting away from what you’ve experienced.
The problem with “drifting” usually happens because some residual information or context from the previous items is still lingering around as the tokens are still looking forward to improve the context while creating the same loophole, the update will be pushed soon.
Do know the upcoming update will keep and keep the context between tasks and the usage of token as you’ve mentioned recently on this thread.
Great prompt usage, make sure to give that prompt with a bit more instructions, thus keeps the flow more strict and autonomous, feature to be pushed soon.
Heres an example below:
You are an autonomous AI developer operating within a two-file system. Your only sources of truth are @project_config.md (LTM) and @workflow_state.md (STM/Rules/Log). Before each action:
Read @workflow_state.md.
Consult the ## Rules section based on the current ## State.
Perform the required action via Cursor.
Immediately update @workflow_state.md.
Then, check @todo.md and proceed to work on the next unchecked task.
Keep actions strict, concise, and rule-bound to maintain focus and flow.
After a roughly testing for 2 weeks I’ve pushed a new update to our current workflow. This update makes Cursor even sharper when tackling lists of tasks. It processes each item one by one, clearing its “mind” between steps to stay on track with and avoid mistakes using Iterations for big projects and keeping its context without drifting away.
The new update TokenizationResults stores results (summary, token count) for each processed item.
Usage: Cursor IDE reads this file constantly before acting and updates it immediately after acting. This is how it maintains context and follows the process.
Now it’s smoother, smarter, and easier on resources—perfect for handling those big ideas.
Latest Update
Iterative processing of item lists.
Context clearing between items prevents “drift.”
Improved accuracy and efficiency for complex tasks.
Glad you just put an update ! i was following the thread closely and i will try intensively.
It has been one week, that i tried your Cursor configuration and i cannot manage to have a single project done but i’m trying.
I’m very new to Cursor and AI agent and also development even if i learned some dev language in the past.
But, with Cursor and the AI i manage to do a web app for me that i had the idea for a long time. that’s awesome.
And now i try with your configuration environnement but i did not manage to do a single project even a single one like a portfolio page.
Everyday i start a new project, and it ends in the bin Because all is fine, it does a great job at planning things and stuffs and looks professionnal, and it even does deep tests and stuffs, i let him do. but at the end it’s a mess, errors everywhere, landing page not starting etc… don’t want to spend another day debugging manually with my poor knowledge or prompt after prompt for the AI.
But i continue to try and even with your new update ! I’ll let you know
Take care!
Thank you so much for the kind words and for sticking with it. It means a lot. I can tell you’re putting in the effort, and that’s something to be proud of, especially as someone who’s new to Cursor, AI agents, and development in general great achievement right there.
Check this prompting resource: Google
Another yet powerful prompting resource: PromptGuide
Yes i’m sticking with it because i was waiting for this kind of thing for a long time. I got the ideas but not the skills to dev. And i know there is potential and opportunities with you stuff.
gotta admit i don’t really understand the current workflow_state.md example file. no clue what the items table is, and the TokenizationResults sections.
I’ve been wanting to try this out since this thread started but haven’t had the time. I finally got around to trying this yesterday and today and I am very impressed! I am not using todo.md because I am just implementing one change or feature at a time. I don’t really want Cursor to be more autonomous than it already is because I want a certain level of control, but I have noticed that it is much more focused when making changes. It will stay laser focused, iterating on a single aspect of the larger change until it gets it right. It also seems to have a much better understanding of it’s previous actions and why it took them. This is a huge improvement and much more efficient than before(not using these rules).
Thank you so much for providing this to the community. I really hope that you(and everyone else) continues to improve it. I’m wondering if it could benefit from some additional memory functionality, such as that provided by the memory MCP servers.
@valtterivalo Thank you for the reply. I’m happy to share with you about what the item table is and the tokenizationResults.
Note this has a typo error.
The Items table is basically a list of stuff the AI is supposed to process. Think of it like a to-do list where each row is a task. In the example, it’s set up to hold text snippets that need summarizing or tokenizing (like breaking text into smaller chunks for analysis). Each row has an Item ID (just a label like item1, item2) and the actual Text to Tokenize (the text itself, like a sentence or paragraph)
The TokenizationResults section is where the AI stores the output after processing each item. It’s like the “done” pile. For each item, it records the Item ID, a Summary of the text (a short version of what the text is about), and a Token Count (a rough estimate of how many tokens—think words or pieces—the text breaks into)
The Itemstable feeds it work, and TokenizationResults collects the results.
Let me know if there’s anything else that’s unclear.
Hey there,
@Btripp It’s great to know that the system is making your single-feature changes more efficient without needing a todo.md file and that project_config.md and workflow_state.md has improved your experience with Cursor.
You can connect this workflow to a memory MCP server, though it’s sometimes confused with the MCP Context Window, where tokens might be wasted. I haven’t faced any issues with the original guide shared at the start of this thread.
Thank you for sharing your experience and there’s yet great things to come.
@brandoncollins7 I recommend reviewing the instructions.md file and following the detailed guidelines provided and within the GitHub repository, you can download the .zip file by clicking the green button located at the top of the page.
Let me know whenever there’s something you need help with,
kleosr
Because all is fine, it does a great job at planning things and stuffs and looks professionnal, and it even does deep tests and stuffs, i let him do. but at the end it’s a mess, errors everywhere, landing page not starting etc… don’t want to spend another day debugging manually with my poor knowledge or prompt after prompt for the AI.
