~/.claude — session-report
> claude usage
loading…
—

findings

36.5%
b로 바꿔줘 — a 4-character prompt consumed 36.5% of all tokens (470K) because it inherited a massive prior-turn context chain of 9 API calls
87.1%
Cache hit rate is 87.1%, below the 90% optimal threshold — cache_create tokens (158K) are relatively high compared to cache_read (1.07M), suggesting frequent context churn
84.9%
Top 3 prompts account for 84.9% of total token spend across 23 of 28 API calls — session was dominated by iterative skill-documentation edits in a single long context
0
Zero cache breaks over 100k uncached tokens — context boundaries are clean and no runaway cache invalidation occurred
33 min
All 5 sessions completed within a 33-minute window in a single project (gameservice-fe-agent-2); no subagents were spawned

summary

tokens by projectshare of total

session timeline by dayclick a day · ←/→ to navigate

—
00:0006:0012:0018:0024:00

most expensive promptsclick to expand context

cache breaks>100k uncached · click for context

projects

subagent types

skills & slash commands

recommendations

Use /compact before short follow-up edits. "b로 바꿔줘" (36.5% of tokens) was a tiny edit that paid for 9 prior turns of context. Running /compact after large document edits compresses history before making small tweaks, cutting per-prompt cost significantly.
Stabilize CLAUDE.md & skills files to improve cache reuse. Cache hit rate is 87.1% — each time skill files or rules are modified mid-session the cache invalidates. Batch skill documentation changes into a single commit so subsequent sessions reuse the warmed cache.
Consider subagents for large document analysis. The session processed large HTML and MD files (fe-ai-workflow.html, WDG00 AI 활용.md) in the main context. Spawning an Explore or general-purpose subagent to read/analyze big documents protects the main context window and reduces token accumulation across turns.
Break iterative doc-editing sessions into shorter conversations. The longest session (183bbac6) ran 23 API calls and 1.14M tokens over ~9 minutes. Starting a new session after major milestones (e.g., after adding skills) resets context cost to zero and keeps individual sessions cheap.