Skip to main content

​
Tool stack

The system is built around a small number of tools with deliberately different roles. The split is intentional: exploratory thinking stays lightweight, implementation starts after scope is bounded, and evidence is stored outside either AI conversation.

​
ChatGPT

Primary planning and review environment. Use it for requirements, architecture, design, documentation, test planning, code review, release reasoning, and deciding what Cursor should do next. The normal mode is conversational Chat. Long-running Work mode is reserved for cases where it is explicitly useful rather than being the default for every engineering discussion.

​
Cursor

Primary implementation agent. Cursor receives repository context, edits code, runs affected tests, and updates the active Work Packet. The key command is:
The goal is to avoid repeatedly pasting long implementation prompts. Cursor is used in two forms:
  • Cursor desktop on macOS for long desk sessions.
  • Cursor CLI on the development Linux server for mobile or remote sessions, typically kept alive inside tmux.
The implementation role is the same in both cases: work against the repository, validate the change, and leave durable state outside the terminal session.

​
tmux + phone SSH

For mobile work, tmux keeps the coding session alive when SSH disconnects or the phone changes networks. A phone SSH client is then only a view/control surface. The development Linux server remains the persistent execution environment.

​
Telegram

Cursor is configured to send a Telegram message when a task completes. Telegram is used only as a completion signal so the user can leave long-running implementation or test work unattended and return when action is required. It is not used as engineering state, evidence, or a source of truth.

​
GitHub

The durable engineering backbone. GitHub stores:
  • code and Git history
  • product specifications and ADRs
  • tests and regression coverage
  • pull requests and review
  • CI evidence
  • exact-HEAD release evidence
  • repository-scoped AI Work Packets
This is where the workflow eventually returned after experiments with more document-centric project management.

​
Desktop Commander Remote MCP

Desktop Commander Remote MCP gives ChatGPT access to an authorized remote machine through controlled filesystem and terminal operations. In this workflow it is used mainly for development-server audit:
This is valuable when a GitHub-only review cannot prove what is actually present or running on the development host. Access should be constrained to authorized directories and commands. Read-first inspection is preferred for audit work.

​
Athena

Athena is the planned searchable knowledge layer for decisions, rationale, and cross-project history. It is intentionally derived knowledge, not the canonical source of truth. GitHub still wins for current code, specifications, tests, and release evidence.

​
Why not one tool for everything?

Earlier workflows tried to let one coding agent handle brainstorming, implementation, debugging, and all context at once. That was convenient, but open-ended conversations could consume a large amount of tokens before the implementation target was stable. When the prompt was not tightly bounded, cost could rise while output quality still varied. The current split keeps high-level thinking lightweight and uses the implementation agent only after the work is well defined.
Last modified on September 19, 2026