Things you should be doing

This is written somewhat in jest. You don’t actually have to be doing any of these things. These are just things i’ve done to some extent that I think are…‘worthwhile’.

Agents

• You should be hating agentic coding
• You should be enjoying agentic coding
• You should be thinking about ways to scale up agentic coding
• You should be saying things like ‘launch 10 subagents to do X’
• You should be saying things like ‘launch a subagent to review plan Y’
• You should be saying things like ‘launch 2 subagents to review plan Y and get consensus’
• You should be thinking that you are still in your infancy of agentic coding
• You should be thinking of launching subagents like you’re learning to program…a ralph-loop is like a do-while loop https://cmdcolin.github.io/posts/2026-03-15-maxagent/…

Workflows

• You should be using git worktrees to implement changes. I previously said I wasn’t using these but now I am
• You should be using a custom claude statusbar (mine https://github.com/cmdcolin/dotfiles/tree/main/claude)

Responsibility

• You should be thinking of ways to avoid making your software take on more responsibility
• You should be demonstrating how responsibilities can be externalized to other tools
• You should be demonstrating how AI can agentically operate your tools
• You should try to minimize your dependencies, even vendoring dependencies https://htmx.org/essays/vendoring/

Docs

• You should be using agents to generate tutorials for your software
• You should be aiming to make every single screenshot in your docs automatically generated https://cmdcolin.github.io/posts/2026-07-31-autogenscreenshots/

GUI and CLI

• You should realize that using agents to write tutorials for your GUI software can improve accessibility, since they use accessible selectors
• You should realize that your point-and-click GUI software is a bottleneck potentially scalability issue, and make CLI solutions

Rewrites and refactoring

• You should be trying to break through significant performance barriers with AI
• You should be doing large scale rewrites with AI

Writing

• You should be writing english prose by hand still
• You should be using LLM to check your english prose though
• You should write the shitty first draft by hand at least
• You should use AI to look for AI writing patterns. When you identify them, make a list of AI tropes. My list https://github.com/cmdcolin/claudish#references

Learning and concepts

• You should be learning how to train models yourself (note to self: still need to do this)
• You should be making your visualization system use grammar of graphics concepts
• You should be thinking about ways to generalize your software
• You should be making side projects. Think fun. Think big. Lessons from these can often apply to your work
• You should use AI to the extent you feel is true to yourself and know when you are becoming fake with it. This is one of the hardest problems of AI

Lingo

• You should likely be using the same lingo opus likes, however cringy, back to it. Tell it to ‘spike’ a prototype ;)

Others are saying

• “You should be asking way more questions” https://www.seangoedecke.com/you-should-all-be-asking-way-more-questions/