md² gives you more control and allows for better multi-tasking when working with agents.
These are some of the problems I ran into and am trying to solve:
- a new worktree means a new vscode window and a new chatlog. Codex makes these chatlogs global for all projects on the machine, claude per folder. both have problems:
- with codex: which chatlog was running on this worktree again?
- claude: a chat done in a worktree is not visible in the main project.
so basically: who is doing what where
Especially if a feature requires multiple prompts and is spread between multiple agents, this became tricky.
- I wanted to track token usage per feature, action, release, project... to have a better insight on how best to optimize and minimize token usage.
- same idea: some tasks like git, graphify, full tests runs, releases and such can be automated with scripts and don't need to be done by an agent which just wastes tokens. They do need close cooperation though: ex when an agent is done, run graphify with a command line and verify result with regex instead of letting the agent do it.
- also: very often I use the same prompts (install, complete, implement, fix merge errors, verify,...). Putting everything in skills just makes the prompts larger, again wasting tokens and I still found myself typing whole sentences to run the same prompt all the time.
also:
- I didn't want a cloud solution, cause of many reasons:
- wanted to have all data local so the agent can easily get to it
- didn't want extra costs
- software dev for me is something local.
working:
- feature lifetime (create, modify, delete, release, run agents,...)
- claude cli & codex cli integration through
- create worktrees, assign worktrees to features, merge back into project
- git management
- remote control (also for codex on windows which I sooo love)
-> last 3 releases have been done almost completely with the tool itself
what still needs more work:
- command actions (things like graphify) still need testing
- remarkable integration (still like writing features on paper)
md² gives you more control and allows for better multi-tasking when working with agents.
These are some of the problems I ran into and am trying to solve:
- a new worktree means a new vscode window and a new chatlog. Codex makes these chatlogs global for all projects on the machine, claude per folder. both have problems: - with codex: which chatlog was running on this worktree again? - claude: a chat done in a worktree is not visible in the main project. so basically: who is doing what where Especially if a feature requires multiple prompts and is spread between multiple agents, this became tricky. - I wanted to track token usage per feature, action, release, project... to have a better insight on how best to optimize and minimize token usage. - same idea: some tasks like git, graphify, full tests runs, releases and such can be automated with scripts and don't need to be done by an agent which just wastes tokens. They do need close cooperation though: ex when an agent is done, run graphify with a command line and verify result with regex instead of letting the agent do it. - also: very often I use the same prompts (install, complete, implement, fix merge errors, verify,...). Putting everything in skills just makes the prompts larger, again wasting tokens and I still found myself typing whole sentences to run the same prompt all the time.
also: - I didn't want a cloud solution, cause of many reasons: - wanted to have all data local so the agent can easily get to it - didn't want extra costs - software dev for me is something local.
working: - feature lifetime (create, modify, delete, release, run agents,...) - claude cli & codex cli integration through - create worktrees, assign worktrees to features, merge back into project - git management - remote control (also for codex on windows which I sooo love)
-> last 3 releases have been done almost completely with the tool itself
what still needs more work: - command actions (things like graphify) still need testing - remarkable integration (still like writing features on paper)
anyway, love to hear what you guys think. useful?