Toknary
Most tokens in a coding session are not spent on the answer. They go on re-reading files, carrying a long history, and printing code you already had. The good news is that each of those is a habit you can change.
Where do the tokens actually go?
Three places: the context you send (files, history, instructions), the work the model does, and the text it writes back. The first is usually the biggest. A long session sends the same history again on every step.
What cuts token use the most?
- Read ranges, not files. Search for the symbol, then open the lines around it. See read the lines you need.
- Ask for a diff. A patch is smaller than a rewritten file. See change the lines, do not reprint the file.
- Keep one task per session. Start fresh when the goal changes. Long sessions get slower and noisier, as in keeping a long session sharp.
- Be specific. A vague request makes the agent explore. A named file and a clear outcome do not. See writing prompts a coding agent can finish.
- Keep caching on. Repeated context is cheaper to re-read. See why repeated reading should cost less.
- Do not use the biggest model for everything. Searching a folder does not need the same attention as designing a change.
Do I have to manage this by hand?
Not with Toknary. It chooses the right model for each piece of work, so you do not pick or count. You still control the habits above: how wide the question is and how long the session runs.
Try it on a real task: install Toknary or create an account.