i was on the other side too. every time a new model came in, i'd think to myself "yet another sloppy next token predictor".
everyone would come up with "11 wild ways you can make millions from gpt 4.5". i would try it out. i would be disappointed and that was the end of it.
don't get me wrong, models were always improving but the jump that the posts talked about just wasn't there.
BUT
we are there now. i now feel like the people who used to write "11 wild ways to make millions" but we are actually there now.
atleast in coding. both in model strength and infra(or harness).
i use claude a lot(LIKE A LOT) and here are a few ways i have improved my own workflow multiple times over(thanks to Boris Cherny) :
directly skip to 9 if you are not a developer(highly recommend you act as one, its never been easier)
1. i have stopped coding. yeah. i just handhold the model in planning. be extremely detailed and opinionated in how the solution should be and let opus rip apart the problem whilst following the plan.
2. use "ask_user_question_tool" to make sure all model assumptions in planning sit right by you. literally say "use this tool to ask me clarifying questions about architecture, decisions, design and everything, be as thorough and as detailed as possible". always give everything you know about the problem and the solution to the model.
3. convert your most used sequence of commands to /command, i converted create a new branch, git add ., git commmit, git push and pr creation to /commit-push {commit message} {branch}
4. never use "dangerously-skip-permissions", always use /permissions instead to allow claude to run commands without asking for permission.
6. run multiple agents at once. do not scroll whilst the model is working on a problem. use multiple terminals to work on the same branch. use git-worktrees or a second directory for the same git repo to work on 2 diff branches at once. 2 can now scale to upto 5 based on the complexity of problems.
7. create subagents. i use 2 subagents a lot: code review, simplify code. simplify code contains the summary of all the pr comments i have gotten ever which now makes it so easy for me to never run into the same review comment again.
8. if you spend time travelling a lot, use tailscale to ssh your phone to your macbook, use a tool like termius to spin up terminals and now you have a claude agent running from your phone helping you learn more about the codebase, solve simple bugs and create PRs.
9. much of what you think you can't do is limited by your own beliefs, you can always ask claude how to do it. if its possible, you probably can do it. don't stick your persona to a specific stack. be delusional that you can(and will) learn anything that's required. be a designer one day. a staff backend engineer the next. start exploring shadcn the next.
always think: how can i do more of what im doing already. agents and models are really powerful and they will only get better :)