Since AI is being blamed for “taking our jobs”, I’m trying to use AI to take away my need for a job.
But, that means I’m paying for AI (Claude Code, in my case) out of my pocket, and I want to get the most work out of my token budget.
I do most of my work with Opus 5 (High), but some of the individual, repeated tasks inside a session may work just as well with a cheaper model, or lower effort. For example, if you have Claude Code handle commits, that work can be done by something cheaper than Opus 5 (High).
I’m also building a code analysis tool that runs from the command line. Instead of asking Claude to review the code, and see if it meets my coding standards (and hope it doesn’t get non-deterministic and hallucinate), I run it through a deterministic tool that outputs a nicely-formatted report of issues to investigate. Then, I feed that report back to Claude and askit to review them and decide which are worthy of doing.
That kind of gets to the heart of using AI for programming. It’s non-deterministic, and programs are all about being deterministic (getting the same output from the same input, every time).
So, I try to wrap the AI’s work with as many checks as I need, in order to produce something with good quality. And by using these tools that run outside of Claude, I’m not using a bunch of tokens trying to fix that randomness.
I’m also using agents, with lower-cost models/efforts, to do some of the repeated work that I initiate inside a Claude session.
I’ve been cleaning up a lot of my old code and doing things like standardizing the labels I use for GitHub Issues. After adding the standard labels, I want to tag any existing issues with the correct new label(s). Again, that’s something I don’t need Opus or Fable to do. So that agent uses Haiku (low).
By offloading this easier work to cheaper models, that leaves me with more tokens to use for Opus 5 (High), my preferred coding model/effort, and the occasional Fable run for large, architectural work.
I plan to regularly review my process and look for ways to improve it. This is prompt I’ll run every week, to get ideas for more ways to move work out of expensive models.
“Give me a summary of work I’ve asked you to do over the last week that used models/efforts higher than needed and whose work could have been moved to a rule/skill/agent/non-model tool at a cheaper cost (fewer, or no, tokens).”
If you have ideas on how to get more out of your LLM subscriptions, please share them below.
Leave a Comment