Most people using Claude are running L1 without knowing L1 exists as a category. Chat, memory, maybe projects. That's it. The reason you hit the ceiling fast isn't the model — it's that the interface hides L2 and L3 behind zero onboarding. Extended reasoning, autonomous research, filesystem access, UI control — none of that is surfaced by default. You have to go looking for it or you'll spend a year running glorified copy-paste workflows and wonder why everyone else seems to be doing more.

The model selection thing cost me real quota before I figured it out. I was running Opus on everything because "smarter = better." Wrong. Opus on a simple draft burns the same quota as Opus on a complex multi-doc analysis. Once I moved to Sonnet for 80% of tasks and dropped Haiku in for triage and quick edits, my effective prompt budget nearly tripled without upgrading. Match the model to the actual cognitive load of the task. Opus is for when you genuinely need deep multi-step reasoning — 623 dispute strategy, contract analysis, anything with compounding logic. Not for drafting a subject line.

Projects are the most underused feature and the one that pays back the most if you're running client work. Each project holds persistent instructions + a memory layer + uploaded files, and Claude reads all of it before every response in that workspace. I have one project per credit client with their full profile, bureau snapshot, and current dispute queue uploaded as a doc. Zero re-explaining per session. Claude opens already knowing which tradelines are reporting, which bureau is priority, and what stage the 609 strategy is at. Setup takes 10 minutes per client, saves it every session forward.

The ChatGPT memory import is real and it works. Settings > Memory > Import Memory generates a prompt, you run it in ChatGPT, paste the output back into Claude. Done. If you've been building context in GPT for months and didn't want to start over, this is the move. Took me under 5 minutes to transfer a full client-context profile I'd been building since January.

The thing that actually matters at L3 is CLAUDE.md. Once Claude has filesystem access and is running agentic tasks, the instructions file is the difference between a useful agent and one that burns your session doing the wrong thing confidently. Most of what I've learned about agentic Claude workflows came from figuring out what to put in that file, not from the model itself.

If you're running dispute letter workflows at volume and haven't wired a Projects workspace per bureau strategy yet, you're re-explaining context on every single session and paying for it in quota.

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.58