AI coding workflows using coding agents

Hello

I have been using cursor in my organisation for past 6 months now. In earlier days, I was able to use it very well using the ask and plan modes. As the features became more complex and the vast codebase became a blocker. Few of the problems I started facing -

  • every chat started with no context of what we have build till now
  • Agent used to omit details, hallucinate and write buggy code
  • token usage went off the roof

in order to mitigate this, i came across few people who have built full SDLC using skills. here are some of them i found interesting -

These are obviously very well designed but again cant fit my use case exactly. I want to learn how can i built my own SDLC using cursor by able to trust AI more with gaurdrails. Is there any resource on the deep learning platform which can guide me for this

Hi @Mukul_Agarwal,

The course you posted under, Spec-Driven Development with Coding Agents, addresses all three problems you describe:

  • Lost context between chats: you write a project-level “constitution” and feature specs as Markdown files, committed alongside your code. Each new session works from those files instead of depending on chat history.
  • Hallucinations and buggy code: each feature spec includes a plan, requirements, and validation criteria. These are your guardrails: the agent implements against them, and you check its output against them.
  • Token usage: this may help too, since the agent works from a focused spec instead of re-exploring the whole codebase each time.

On your main question, building your own SDLC instead of adopting someone else’s: the course shows how to package your own workflow into an agent skill. It uses Claude Code, but the workflow is agent-agnostic and built on open standards like AGENTS.md, so it should carry over to Cursor. It also covers applying the approach to existing code, not just new projects.

If you’ve already taken it, the lessons on the constitution and on creating your own skill are the ones to revisit.