Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Reply to comments (批注) in Word .docx/.doc files: extract comment context, draft replies, write threaded replies back, and validate OOXML.
.claude/skills/foryourhealth111-pixel-docx-comment-reply/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -13% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -26% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -2% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -27% | 0% |
这个 skill 解决的问题：把 Word 文档里的批注（comments）按“原文锚点上下文”整理出来，生成待回复清单，然后把回复以 threaded replies 的方式写回到新的 .docx 文件里（不改原文件）。
适用场景：专利/论文/合同/内部评审等需要“逐条回复批注”的文档。
在当前工作目录的 outputs/ 下生成：
*_批注定位与上下文_*.md：人可读的批注+锚点上下文报告*_comment_context_*.json：机器可读上下文（用于并行写回复/自动化）*_replies_todo_*.json：待回复模板（键=comment_id，值=空字符串）*_批注已回复_*.docx：写回批注回复后的最终交付文件powershellpython scripts/extract_comment_context.py --input "path\\to\\file.docx"
如果输入是 .doc，脚本会尝试用 LibreOffice soffice 转成 .docx 后继续。
outputs\\*_批注定位与上下文_*.md，逐条写回复。outputs\\*_replies_todo_*.json（保持 JSON 结构不变）。回复口径（强约束）
powershellpython scripts/apply_comment_replies.py ` --unpacked "outputs\\<xxx>_unpacked_<timestamp>" ` --replies "outputs\\<xxx>_replies_todo_<timestamp>.json" ` --author "YourName" ` --initials "YN"
脚本默认会在保存时做 schema + redlining 校验；如需单独验证：
powershellpython ..\\docx\\ooxml\\scripts\\validate.py "outputs\\<unpacked_dir>" --original "outputs\\<out>.docx"
当批注数量较多（例如 ≥20 条）：
comment_context.json$vibe）apply_comment_replies.py| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 29,233 | 6,189 | -79% | 1 | 1 | 0% | 5,921 | 1,074 | -82% | 0 | 0 | — |
case-02 | fail→fail | 29,024 | 5,851 | -80% | 1 | 1 | 0% | 5,515 | 1,085 | -80% | 0 | 0 | — |
case-03 | fail→fail | 18,884 | 5,436 | -71% | 1 | 1 | 0% | 3,664 | 1,092 | -70% | 0 | 0 | — |
case-04 | fail→fail | 10,564 | 12,545 | +19% | 1 | 1 | 0% | 1,750 | 3,368 | +92% | 0 | 0 | — |
case-05 | fail→fail | 26,277 | 23,517 | -11% | 1 | 1 | 0% | 5,744 | 5,683 | -1% | 0 | 0 | — |
case-06 | fail→fail | 15,984 | 10,842 | -32% | 1 | 1 | 0% | 3,061 | 2,652 | -13% | 0 | 0 | — |
case-07 | pass→fail | 12,687 | 6,297 | -50% | 1 | 1 | 0% | 1,904 | 1,151 | -40% | 0 | 0 | — |
case-08 | pass→pass | 9,808 | 6,452 | -34% | 1 | 1 | 0% | 1,846 | 1,810 | -2% | 0 | 0 | — |
case-09 | fail→pass | 9,972 | 2,364 | -76% | 1 | 1 | 0% | 1,732 | 1,114 | -36% | 0 | 0 | — |
case-10 | pass→pass | 10,747 | 5,421 | -50% | 1 | 1 | 0% | 1,870 | 1,639 | -12% | 0 | 0 | — |
case-11 | fail→pass | 19,132 | 10,805 | -44% | 1 | 1 | 0% | 3,110 | 2,697 | -13% | 0 | 0 | — |
case-12 | fail→pass | 9,480 | 4,216 | -56% | 1 | 1 | 0% | 1,825 | 1,359 | -26% | 0 | 0 | — |
case-13 | fail→pass | 8,739 | 4,872 | -44% | 1 | 1 | 0% | 1,574 | 1,549 | -2% | 0 | 0 | — |
case-14 | fail→pass | 32,804 | 2,354 | -93% | 1 | 1 | 0% | 1,561 | 1,134 | -27% | 0 | 0 | — |
case-15 | fail→pass | 11,490 | 3,368 | -71% | 1 | 1 | 0% | 2,133 | 1,131 | -47% | 0 | 0 | — |
case-16 | fail→pass | 11,372 | 2,492 | -78% | 1 | 1 | 0% | 2,312 | 1,093 | -53% | 0 | 0 | — |
case-17 | pass→pass | 18,180 | 14,412 | -21% | 1 | 1 | 0% | 3,082 | 3,193 | +4% | 0 | 0 | — |
case-18 | fail→pass | 11,381 | 6,362 | -44% | 1 | 1 | 0% | 1,906 | 1,927 | +1% | 0 | 0 | — |
case-19 | fail→pass | 8,063 | 3,311 | -59% | 1 | 1 | 0% | 1,516 | 1,245 | -18% | 0 | 0 | — |
case-20 | fail→pass | 12,746 | 3,880 | -70% | 1 | 1 | 0% | 2,022 | 1,245 | -38% | 0 | 0 | — |
case-21 | pass→pass | 11,617 | 2,883 | -75% | 1 | 1 | 0% | 1,968 | 993 | -50% | 0 | 0 | — |
case-22 | fail→pass | 14,642 | 6,383 | -56% | 1 | 1 | 0% | 2,441 | 1,797 | -26% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 20 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.