skill
comment-floor
Agent skill: work your own post's comment section like a front desk. Capture every floor, prove completeness, sort scam/hostile/correction/question/request/praise, fold +1 echoes, answer once and pin.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/fangdididi/comment-floor itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Nov 1, 2026English | 简体中文
Comment Floor
A front-desk shift for the comment section of your own post: capture every floor, prove it's complete, answer each question once, and own your mistakes in public.
The comments under a post that took off are the most honest feedback a creator gets, and the hardest to use. Replies hide in collapsed threads. The same question arrives twelve times in two languages. The one viewer who caught a real error is buried under "+1"s. A "DM me for free presets" scam sits there for days. And whatever gets answered is simply whatever happened to be on top.
Comment Floor is an agent skill that turns that pile into a short, repeatable routine:
- Floor card first. One post, whose it is, the comment count the platform shows, the purpose of this pass, and the owner's reply voice and red lines.
- Gentle capture. Official export, creator dashboard, the page's own requests in the owner's logged-in browser, or by hand, in that order. Human pace, and a hard stop at any captcha, login wall or rate limit.
- Completeness you can prove. Captured vs shown, thread by thread: COMPLETE, NEAR, PARTIAL or THIN, with the exact floors still missing replies and warnings for looping pages, orphans and mixed-in posts.
- Six kinds plus echoes. Scam, hostile, correction, question, request, praise. A reply that only says "+1" or "同问" is folded into the question above it as an extra voice.
- Four lanes in order. Moderate, answer, consider, thank. Already-answered threads drop out. Look-alike questions (even across languages, via topic labels) become one group with total voices and likes, and the biggest becomes the pin candidate.
- Drafts, never auto-posts. Replies drafted in the asker's language and the owner's voice; the owner posts. Plus a corrections ledger and a seeds list for the next post.
Features
| Any capture shape | ingest --map maps any JSON / JSONL / {"comments": [...], "replies": [...]} onto one record, dotted paths for nested fields |
| Completeness verdict | shown-count or per-thread basis, thread gaps ranked by missing replies, PAGE_REPEAT / ORPHANS / OVERCOUNT / EMPTY_TEXT warnings |
| Bilingual rules | English and Chinese patterns for scam bait, insults, corrections, questions, requests, praise and echoes |
| Your labels win | labels.json overrides any kind and can add a topic to force grouping |
| "Answered" logic | a creator reply later in the same thread marks earlier comments as handled |
| Scoring | kind weight × log of likes (echo likes included) + echo voices + a recency bump |
| Pin candidate | top answer group with ≥ 3 voices or ≥ 50 likes |
| Markdown floor | verdict, queue and the whole floor sorted by likes with replies indented |
| Zero dependencies | Python 3.8+, standard library only, no network calls; 63 tests |
Install
The skill is a folder with SKILL.md at its root. Clone it where your agent looks for skills:
# Claude Code (personal)
git clone https://github.com/fangdididi/comment-floor ~/.claude/skills/comment-floor
# Claude Code (one project)
git clone https://github.com/fangdididi/comment-floor .claude/skills/comment-floor
# Codex CLI
git clone https://github.com/fangdididi/comment-floor ~/.codex/skills/comment-floor
# Cursor and other agents: put the folder in your rules/skills directory,
# or point the agent at SKILL.md (English) or SKILL.zh-CN.md (Chinese).
If the target folder already exists, look inside before overwriting it.
Usage
Ask in plain words:
Go through the comments on https://www.douyin.com/video/7xxxxxxxx and tell me what to reply to first.
帮我值一次班:这条小红书笔记下面的评论,先清广告,再告诉我该置顶什么。
Or run the helper yourself:
python3 scripts/floor.py ingest raw.json floor.jsonl --map "id=cid,text=content,likes=digg,time=ts,author=user.nick,reply_total=reply_n,parent=reply_to" --creator my_handle
python3 scripts/floor.py check floor.jsonl --shown 312 # exit 0 = COMPLETE or NEAR
python3 scripts/floor.py sort floor.jsonl > kinds.tsv # read it, fix misfires in labels.json
python3 scripts/floor.py queue floor.jsonl --labels labels.json
python3 scripts/floor.py render floor.jsonl --shown 312 --labels labels.json > floor.md
python3 -m unittest discover -s tests # run the tests
Normalised record (JSONL, one comment per line):
{"id": "r1b", "parent": "c1", "text": "+1 what camera?", "likes": 12, "time": "2026-10-30T05:30:00Z",
"author": "nn_93", "reply_total": 0, "by_creator": false, "pinned": false}
Example
A photography tip video showing 12 comments (all files in examples/):
Verdict: NEAR. Captured 11 (top-level 7, replies 4), coverage 91.7% by shown count.
- Threads still missing replies: c1 (2/3), c5 (1/2)
Pin candidate: one answer covers 5 voices / 648 likes, e.g. "Which camera did you use for the overhead shots?".
Moderate first
- [scam] Free editing presets, add me on whatsapp for the pack · 2 likes · c4
Answer (grouped, unanswered only)
- [question] Which camera did you use for the overhead shots? · 648 likes · 5 voices · c1, c2, c7
Consider (requests)
- Please make a part 2 on colour grading · 150 likes · c5
The correction about the rule of thirds already has a creator reply, so it leaves the queue and goes into the corrections ledger instead.
Boundaries
- Your own posts, or a client's with written permission. Public posts that aren't yours: read-only, no drafts aimed at people, no moderation lane.
- One post at a time, at a human pace. No signature forging, account rotation or captcha solving.
- The agent drafts; the owner posts, pins, hides and reports, unless they explicitly ask the agent to do a specific action.
- Folded, deleted and author-only comments are usually invisible to everyone else; NEAR is often the honest ceiling.
File structure
comment-floor/
├── SKILL.md # the skill (English)
├── SKILL.zh-CN.md # the skill (Chinese)
├── README.md
├── README.zh-CN.md
├── LICENSE # MIT
├── scripts/
│ └── floor.py # ingest / check / sort / queue / render
├── tests/
│ └── test_floor.py # 63 tests
└── examples/
├── raw_capture.json # a raw capture in a platform-like shape
├── comments.jsonl # normalised floor with replies
├── labels.json # topic labels grouping three camera questions
└── floor.md # rendered output
License
MIT © fangdididi