Ask HN: What works for cutting AI token costs?

My LLM token bill is getting painful.

Besides switching to cheaper models, what have you personally used to reduce cost in real applications?

[flagged]

When I work with AI to build something I like to summarize every conversation into separate folder and basically keep a journal for AI. This is how it looks like:

"journal" folder:

- 2026-05-12 Add comments support for the cells.md

- 2026-06-15 Rework onboarding screen.md

Each document contains show description of the problem and explanation how did we solve it. Then If I need to go back in time and improve one of the features I can use one or multiple of those files as a reference so we don't need to "rescan" entire code base and re-analyze it. And the most important - I'm not loosing some decisions we made which from the code alone can be considered bugs but they are features actually

[dead]