Newb having trouble understanding how to best use task-master

[ashepp]

Hi there, Thanks for building this I think it has huge potential and I've made some good progress in working with Cursor 3.7 Sonnet with it. However, I'm having a few problems that I'd love to get help with. For full transparency, I'm a hobbyist coder and working my way through these tools so apologies if some of these questions are naive.

1. I'm finding that after the initial integration with my project it seems that the cursor agent is "working around" task-master or forgets how to use it. I don't know if my cursor rules got changed or what expectations are but I keep having to remind it on syntax and how to use task-master. Sometimes the agent just changes the json directly

2. How should task-master integrate with github? I'm working on PC and mac on a repo and it seems things like the .env were in the ignore file.

3. How to update or if things go wrong get back to a working state? I think right now I'm in a broken state so don't know how to get it back on track?

4. How do tests work for each individual task or subtask? Are they generated automatically?

5. Should I be editing the tasks directly or always through task-master if I want to be more specific?

6. Is it possible to add some generic instructions for cursor? Like "Always use CUDA when setting up a local LLM model". I've been having issues getting deepseek working locally for a project and it doesn't seem to follow the tasks that I created.

Appreciate any tips and advice.
Keep up the great work!

[eyaltoledano]

Thanks for trying it out! Here's some thoughts:

1. Pretty sure this is a bug because the cursor rules are trying to use the dev.js scipt directly in many cases. This is a known bug Adjust cursor rules to use global CLI command instead of dev.js #47
2. Taskmaster generally won't appear in your github stuff. The gitignore removes the tasks, tasks.json, env (if it's not already removed). If you do want the tasks to be checked in, which I would recommend, simply rmeove it from the gitignore. I'm going to be improving this by Sync task list into project README.md #40 so that the output of task-master list shows up and stays synced in the project's README (and shows up in GitHub). Beyond that, I'd say Feature: GitHub Integration #11 is the next answer if you're asking me from the perspective of team collaboration. If you're solo -- a local tasks.json file works very, very well. GitHub integration in earnest will unlock "Mutliplayer mode" for taskmaster by avoiding conflicts between tasks across branches.
3. It depends on what you mean by broken state --
   a. If the state is broken because you've been freestyling and now you're integrating taskmaster, then my recommendation is to rollback to the last clean/working state, and ensure your PRD defines the features you just discarded (that were broken) as well as all features to get the project feature-complete. And then do a taskmaster parse-prd to generate a new tasks.json file using that PRD, and task-mater generate to get all your individual task_xxx.txt files. You can do a task-master analyze-complexity to trigger a Perplexity call against each of your tasks, which will create a report you can access with task-master complexity-report. It will be ranking all your tasks by hardest to easiest, with dedicated expand prompts for each task. You can tell Cursor Agent to run complexity-report and then sequentially run the expand commands for each of the tasks one by one.

That ensures you have an accurate task file, fully spec'd and researched, with as many subtasks as is logically reasonable relative to the complexity of the task and its spec. Simple tasks might not even need subtasks.

Now you can do task-master list --with-subtasks to identify the entire project task and dependency chain, and taskmaster next which will pick the next task to work on based on all task statuses, priorities and dependencies.

ALL of the above can be done either through the CLI manually, through the CLI via Cursor Agent with natural language, or through the MCP soon (which is still cursor agent natural language, just less spotty).

TLDR: Above is how to solve your issue, and MCP support will make it much easier to do so via natural language where Cursor Agent will pick up on your intent/frustration and attempt to remediate using the above and other functions.

4. You're on the money. This is an upcoming feature I'm working on right now. Each task should have a test suite with as many tests as there are subtasks. I just realized there is no GitHub issue for it. That goes back to the issue of GitHub integration - it would be better if the tasks.json and GitHub were in sync.

Task -> https://github.com/eyaltoledano/claude-task-master/blob/main/tasks/task_024.txt

5. I don't recommend writing directly to the tasks.json file manually. Use the CLI commands. You can add-task if you want just one more task. You can add-subtask if you want just one more subtask. You can remove either as well. You can set-status to tasks and sub tasks, passing in one or many id's at a time.

The update and update-task (coming next release) commands are extremely useful if you learn something mid-implementation. You can pass taskmaster update-task --id=22 --prompt="we need to mock the functions in the jest tests, not try to import them directly from the core modules. otherwise it adds a huge amount of unnecessary boilerplate" and it will hit Claude to update the task with this insight. If you use update instead of update-tasks it will do this to ALL incomplete tasks forward of task 22 that rely or mention task 22.

6. Yes! I'm working on this too. It's called the learn command. You should be able to adjust the Cursor rules quickly with some new pattern/insight or whatever you want to enforce as you go. learn command to automatically create or improve Cursor rules #30

You're on the ball. We're on the same page :)