How much does Claude actually say? Classifies every assistant turn by output type. 73% are completely silent.

npx cc-text

Zero dependencies. Reads ~/.claude/projects/ directly.

cc-text — How Much Does Claude Actually Say?
==================================================
Sessions: 2,266 | Turns: 269,656 | Silent rate: 73.4%

Assistant turn types:
  tool-only   ████████████████████  144,907   53.7%  (executes silently)
  thinking    ███████░░░░░░░░░░░░░   53,087   19.7%  (thinks silently)
  text-only   ██████████░░░░░░░░░░   71,627   26.6%  (says something)

Text length tiers (when Claude speaks):
  micro     ████████████████████   33,815   47.2%  (one-liner <50 chars)
  ack       ████████████████░░░░   27,368   38.2%  (acknowledgment 50–199)
  explain   █████░░░░░░░░░░░░░░░    8,344   11.6%  (explanation 200–999)
  essay     █░░░░░░░░░░░░░░░░░░░    2,100    2.9%  (essay 1000+)

  median 53 chars · mean 235 chars · p90 316 chars

• 73% of turns are silent — tool-only (54%) + thinking-only (20%). No visible text output
• When Claude speaks, it's brief — 47% of text turns are micro (<50 chars). Just "Done." or "Here's the result."
• Only 3% are essays (1000+ chars) — long explanations are rare
• Thinking (19.7%) — Claude reasons internally without speaking. Not visible in the chat, but counted here
• Mean (235) >> Median (53) — a few long explanations skew the average

npx cc-text          # turn type distribution
npx cc-text --json   # raw JSON output

yurukusa.github.io/cc-text — drag and drop your projects folder.

Part of cc-toolkit — tools for understanding your Claude Code sessions.

Claude Code Ops Kit ($19) — 16 production hooks + 5 templates + 3 tools. Built from 160+ hours of autonomous operation.

Source: yurukusa/cc-text