name	claude-code-experience
description	Compute how many hours the user has worked with Claude Code on this machine, from local transcripts (~/.claude/projects) and prompt history (~/.claude/history.jsonl). Use when the user asks "how many hours have I worked with Claude Code", "what's my Claude Code experience", "my Claude Code usage/stats/hours", or wants a monthly breakdown of their Claude Code activity.

Claude Code Experience

Estimate total hands-on hours with Claude Code from local data. No network access needed.

Workflow

Step 1 — Run the analysis

Run the bundled script (can take a minute on large transcript dirs):

python3 scripts/analyze_usage.py

It prints a complete report; present the results to the user as a short summary plus the monthly table. Lead with the calibrated lifetime estimate — that is the headline number.

Step 2 — Offer a shareable image via ChatGPT

After presenting the numbers, give the user a copy-paste-ready block for ChatGPT that turns the analysis into an image. Emit a single fenced block containing exactly this prompt with the script's raw output inlined in place of <output>:

turn this analysis into an image. Only the key facts.
<output>

How the numbers are computed (for explaining to the user)

Active time = sum of gaps between consecutive events, each gap capped at an idle timeout (5/15/30 min) — the WakaTime method. All events are merged into one global timeline first, so parallel sessions are not double-counted.
Transcripts (~/.claude/projects/**/*.jsonl) have event-level granularity but are auto-cleaned after ~30 days, so they only measure the recent window.
Prompt history (~/.claude/history.jsonl) survives cleanup and covers the long term, but only records user prompts, so it undercounts time Claude spends working between prompts.
Calibration: over the window where both sources exist, the transcript-based hours divided by the history-based hours gives an undercount factor; the lifetime history number scaled by that factor is the calibrated estimate.

Caveats to mention

Machine-local only: other machines, claude.ai/code web sessions, and anything before the start of history.jsonl are not counted. The script flags when firstStartTime in ~/.claude.json predates the history file.
"Wall clock with sessions open" is much larger than active time when the user leaves sessions running — don't present it as hours worked.
Honors CLAUDE_CONFIG_DIR if set; ~/.claude.json stats (launches, projects) are read best-effort.

Claude Code Experience

Estimate total hands-on hours with Claude Code from local data. No network access needed.

Workflow

Step 1 — Run the analysis

Run the bundled script (can take a minute on large transcript dirs):

python3 scripts/analyze_usage.py

It prints a complete report; present the results to the user as a short summary plus the monthly table. Lead with the calibrated lifetime estimate — that is the headline number.

Step 2 — Offer a shareable image via ChatGPT

turn this analysis into an image. Only the key facts. <output>

How the numbers are computed (for explaining to the user)

Active time = sum of gaps between consecutive events, each gap capped at an idle timeout (5/15/30 min) — the WakaTime method. All events are merged into one global timeline first, so parallel sessions are not double-counted.

Transcripts (~/.claude/projects/**/*.jsonl) have event-level granularity but are auto-cleaned after ~30 days, so they only measure the recent window.

Prompt history (~/.claude/history.jsonl) survives cleanup and covers the long term, but only records user prompts, so it undercounts time Claude spends working between prompts.

Calibration: over the window where both sources exist, the transcript-based hours divided by the history-based hours gives an undercount factor; the lifetime history number scaled by that factor is the calibrated estimate.

Caveats to mention

Machine-local only: other machines, claude.ai/code web sessions, and anything before the start of history.jsonl are not counted. The script flags when firstStartTime in ~/.claude.json predates the history file.

"Wall clock with sessions open" is much larger than active time when the user leaves sessions running — don't present it as hours worked.

Honors CLAUDE_CONFIG_DIR if set; ~/.claude.json stats (launches, projects) are read best-effort.

claude-code-experience