AI
Token Counting for Fun and Profit
August 10, 2026
One skill invocation in a Claude Code session wrote 261,974 tokens into the prompt cache. The tokens replayed from cache on each subsequent request went from 93,786 to 355,760, and 93,786 + 261,974 = 355,760 exactly. Every turn after that point cost several times what the turns before it cost, for the rest of the session. That arithmetic is the mechanism confirming itself.
Model choice and reasoning effort both change how fast a session burns through a subscription usage limit, and both are settings you pick deliberately. Two more are not settings at all. They accumulate from how you work, turn by turn, mostly unobserved: what you put into the context, and whether the cache is still warm when you come back. Those two are the subject here.
Why This Series Can Never Be Finished
October 19, 2025
What Just Happened
I was working on an article about AI agent nondeterminism. The core thesis: AI agents produce unpredictable outputs because they’re probabilistic systems.
One of my examples:
When I ask Claude Code to run tests, I might get:
uv run pytest(correct)pytest(wrong, bypasses environment)python3 -m pytest(wrong, bypasses uv)
Same context. Same documentation. Same conversation history. Different command each time.
Hours after writing that section, I caught the problem:
AI Coding Agents Are Plausible Bullshit Generators
October 18, 2025
The agent’s output was perfect—detailed verification reports, specific metrics, professional completion messages. Then I looked deeper. None of it was real.
I’ve been building personal projects using AI coding agents for the past several months—projects I’m developing to production standards because they’ll eventually handle consequential decisions in real-world use. Not as an experiment - as my primary development method. This means I’ve hit many of the critical failure modes, not in theory but in practice where mistakes cost real time and create real risk.