菜单
Run a command and redirect output to a temp file instead of dumping it into context. Returns the file path, line count, and a preview (first/last 10 lines). Use for test runs, build logs, or any command with large output. Part of claude-cost-helpers (watching-cost).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Delegate a task to a subagent with its own clean context window. Keeps heavy file reading out of your main session. Use when reading many files, doing codebase audits, or research tasks. Part of claude-cost-helpers (subagent-isolation).
Load the most recent session file and resume work with full context from where the last session ended. Use at the start of a new session after a save-session or split. Part of claude-cost-helpers (idle-tax).
Save session state and recommend starting fresh instead of compacting. A safer alternative to lossy compaction — compaction discards context permanently with no recovery. Part of claude-cost-helpers (compact-gamble).
Save current session state for handoff to a future session. Creates a structured file capturing what worked, what failed, and the exact next step. Use when stepping away, before cache expires, or when context is getting heavy. Part of claude-cost-helpers (idle-tax).
Save current session and split — start fresh to avoid context rot and rising costs. Use when context is heavy (300K+ tokens estimated) or response quality is degrading. Part of claude-cost-helpers (just-one-more-turn).
| name | to-file |
| description | Run a command and redirect output to a temp file instead of dumping it into context. Returns the file path, line count, and a preview (first/last 10 lines). Use for test runs, build logs, or any command with large output. Part of claude-cost-helpers (watching-cost). |
Run a command and capture its output to a temp file instead of printing it into context. Use this when the command would produce large output (test runs, build logs, directory listings, file contents) that you don't need in full right now.
/to-file <command>
Example: /to-file npm test
/to-file)/tmp/claude-output-<short-hash>.txtwc -l)head -10)tail -10)<filepath>. Read specific sections with Read <filepath> using offset and limit parameters."User: /to-file npm test
Claude runs:
HASH=$(echo "npm test" | md5 | cut -c1-8)
OUTFILE="/tmp/claude-output-${HASH}.txt"
npm test > "$OUTFILE" 2>&1
echo "File: $OUTFILE"
echo "Lines: $(wc -l < "$OUTFILE")"
echo "--- First 10 lines ---"
head -10 "$OUTFILE"
echo "--- Last 10 lines ---"
tail -10 "$OUTFILE"
Then tells the user: "Full output is at /tmp/claude-output-a1b2c3d4.txt — use Read /tmp/claude-output-a1b2c3d4.txt offset:40 limit:20 to read specific sections."
Every line of output that appears in the conversation sits in context permanently and is reprocessed on every subsequent message. A 5,000-line test run pasted into context costs roughly the same as 5,000 lines of your own code — but unlike your code, it has zero value after the first glance. Redirecting to a file keeps context clean and costs near zero. You can always read the parts that matter.
This skill is part of the Watching Cost helper (watching-cost). The companion hook (output-size-monitor.sh) automatically warns when tool output exceeds 5K tokens per call and tracks cumulative output with escalating alerts at 25K, 50K, and 100K tokens. Install the hook for automatic warnings: cd watching-cost && ./install.sh