Skip to main content

​
Daily workflow

The default workflow is intentionally simple.
1

Think in ChatGPT

Discuss the problem, requirements, architecture, edge cases, UX, documentation, test strategy, and scope before implementation begins.
2

Create or update the AI Work Packet

When the work is ready for implementation, ChatGPT updates the repository-scoped [AI Work] Issue instead of generating another giant Cursor prompt.
3

Run /resume in Cursor

Cursor resolves the current repo, branch, HEAD, and active Work Packet, then executes only the current Next Action.
4

Implement and validate

Cursor changes the repository and runs the cheapest affected deterministic tests first. Expensive qualification is deferred until it is actually required.
5

Return to ChatGPT

After Cursor updates the Work Packet, the user can simply say “Cursor 끝났어. 계속”. ChatGPT reloads durable state and continues.
6

Audit directly when needed

If GitHub evidence is not enough, ChatGPT can inspect the authorized development server through Desktop Commander Remote MCP.

​
The user-facing commands

The hidden complexity stays in GitHub instead of being copied by the user.

​
Why ChatGPT first

The design intentionally spends more time in cheap, interactive planning before starting an implementation agent. In this workflow, ChatGPT is used for:
  • brainstorming and product decisions
  • architecture and design
  • specification and documentation
  • test strategy
  • review and debugging hypotheses
  • deciding the smallest next implementation action
Cursor is then given a bounded task with current repository context. This separation is partly technical and partly economic: open-ended exploration inside a coding agent can consume a large amount of context before the implementation target is stable.

​
Validation rhythm

A known deterministic blocker stops expensive downstream validation until it is fixed.

​
Desk and mobile operation

The same engineering loop is used in two physical modes:
  • At a desk, implementation normally runs through the Cursor desktop application on the Mac.
  • While moving, the development Linux server becomes the stable execution environment. Cursor CLI can continue inside tmux, and the user can reconnect from a phone over SSH.
  • Cursor sends a Telegram completion notification so the user does not need to watch a long-running terminal session.
  • After notification, the result can be brought back to ChatGPT for review and the next decision.
  • When copied output is not enough, ChatGPT can inspect the authorized development server directly through Desktop Commander.
See Mobile & Remote Workflow for the full operating pattern.
Last modified on September 19, 2026