Repository navigation
Excessive Token Usage with Claude Code (Anthropic Pro) #1492
Description
Hi team, thanks for the excellent work on this framework — it has been very useful when combined with Claude Code.
When using the framework with Claude Code on an Anthropic Pro account, usage limits are typically reached in under an hour of active work. This is noticeably faster than what many users report for typical Pro usage patterns, even with relatively heavy coding sessions.
This suggests that the generated spec-related artifacts may be larger or more verbose than necessary, driving up token usage for both generation and downstream interactions. In particular, long, highly redundant, or repeatedly regenerated specs can quickly consume a significant portion of the Pro allocation.
It might be helpful to explore strategies to reduce the overall token footprint of these artifacts, for example:
- Making spec and auxiliary artifacts more concise by removing redundancy and boilerplate.
- Introducing a “compact” or “minimal” mode that aggressively shortens descriptions, metadata, and comments.
- Caching or reusing previously generated spec fragments instead of regenerating them from scratch on each run, where feasible.
Optimizing for token efficiency would let users stay within their Claude Pro limits longer while still benefiting from the framework’s capabilities, especially during extended coding sessions.
Happy to share concrete usage examples or logs if that would help with debugging or design discussions.
Activity
dhilipkumars commentedon Feb 19, 2026
related to : #1401
Not exactly the same problem , but can be put under the same bucket to optimize token usage.
dhilipkumars commentedon Feb 20, 2026
this is interesting i have seen it being excessively verbose especially if i try to build a small feature with it, I would rather use the prompt or change the file directly for tiny changes. at the same time it was not that bad that i exhausted my tokens in an hour.
BTW we recently merged
AgentSkillsto this project. that automatically increases cacaching and lowers the token usage, if you want to give it a try with--ai-skillsoption.i did a quick test supplied same prompt on a project with
ai-skillsand without it, on Gemini CLI.and here are the results
ai-skills(Run 1)commands(Run 2)ai-skillsai-skillsai-skills(34% fewer)commands(Higher volume)ai-skills(23% fewer)ai-skills(More concise reasoning)not sure how much of this will translate into real-world usecases but thought its worth letting you know. Would you like to re-test it? @john-agentic-ai-tools
dhilipkumars commentedon Feb 23, 2026
@mnriem may be a good candidate to convert into discussion.