[REPLACE: SKILL_NAME] Skill
在社交平台上对 [REPLACE: KEYWORDS] 做提及/关键词扫描——趋势、情感、热门帖子
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
${var} — Optional. Pass alternative keywords (comma-separated) to override the default. If empty, monitors [REPLACE: KEYWORDS].
Today is ${today}. Monitor social mentions of [REPLACE: KEYWORDS] and produce a summary.
Steps
- Resolve keywords - use
${var}if set, else[REPLACE: KEYWORDS]. Split on commas, trim each, lower-case. Each token becomes its own search query, written into its own calls literally (thebelow;is its position in the list).
- Search X — for each keyword, use the X / xAI search path (project's standard pattern):
Key check: read the "Keys set for this run" line in your run notes, never a shell test (those are refused and cost a turn). With XAI_API_KEY on that line, build the body with a literal jq -n filter and pipe it straight into ./secretcurl (set the Bash tool timeout to at least 180000; the live X search takes 30-120s):
jq -n '{model:"grok-4.7", input:[{role:"user",content:"Search X for posts from the last 24 hours about: <KEYWORD>. Language: [REPLACE: LANGUAGE]. For each post include @handle, the full text, date posted, exact like, repost and reply counts, and the direct link (https://x.com/handle/status/ID). Return as a numbered list."}], tools:[{type:"x_search"}]}' \
| ./secretcurl -s -o /tmp/social-x-<N>.json -w 'xai http=%{http_code}\n' --max-time 150 -X POST "https://api.x.ai/v1/responses" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer {XAI_API_KEY}" \
--data-binary @-
On http=200, read the text with jq -r '.output[] | select(.type == "message") | .content[] | select(.type == "output_text") | .text' /tmp/social-x-.json. Without the key, or on a non-2xx code, an empty reply or a timeout, fall back to WebFetch (or a Nitter mirror) and log the true reason.
Restrict to language [REPLACE: LANGUAGE] (e.g. en, fr, any). Drop posts with fewer than [REPLACE: MIN_LIKES] likes — that filter is what protects the channel from low-signal noise.
- Search Reddit — for each keyword:
curl -s -o /tmp/social-reddit-<N>.json -w 'reddit http=%{http_code}\n' -H "User-Agent: aeon/1.0" -G "https://www.reddit.com/search.json" --data-urlencode "q=<KEYWORD>" -d t=day -d restrict_sr=0
Reddit's JSON endpoint is keyless. On a non-2xx code (often 429), make a second call with WebFetch on https://www.reddit.com/search.json?q=&t=day.
- Score and pick top 5 per platform — score on engagement (likes, comments, score) × recency (last 24h gets full marks). Drop reposts and obvious bot accounts (handles like
*_bot, account age < 7 days with > 100 posts).
- Tag sentiment — for the top 10 posts overall, label each
positive/neutral/negativebased on tone of the post text. Keep this lightweight — one-token classification, no nested reasoning.
- Write
output/articles/[REPLACE: SKILL_NAME]-${today}.md:
# [REPLACE: KEYWORDS] — ${today}
## Volume
- X: N posts (vs 7d avg M)
- Reddit: N posts (vs 7d avg M)
## Sentiment
positive: X · neutral: Y · negative: Z
## Top posts
1. [Author · platform · timestamp]
"Excerpt or paraphrase."
→ URL
2. ...
- Notify via
./notifywith a 2-3 line summary:[REPLACE: KEYWORDS] — ${today} · N posts · sentiment skews positive/negative · top: . Full digest:. Silent on quiet days (volume < 25% of 7d average AND no negative-sentiment spike).
- Log to
memory/logs/${today}.md:
## [REPLACE: SKILL_NAME]
- **Volume**: x_posts=N, reddit_posts=N, vs_7d_avg=Δ%
- **Sentiment**: pos=X, neu=Y, neg=Z
- **Status**: SOCIAL_OK | SOCIAL_QUIET | SOCIAL_SPIKE (vol > 2x avg) | SOCIAL_DEGRADED
Network note
X / xAI requires XAI_API_KEY; a bare $XAI_API_KEY on a curl line is refused by the Bash analyzer, so call ./secretcurl with the {XAI_API_KEY} placeholder (the key is injected via requires:). Reddit's JSON endpoint is keyless but rate-limited per IP — WebFetch is the fallback when curl returns 429.
Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (KEYWORDS=...), uses if / for, tests a key, joins commands with ;, && or ||, redirects into a file (> .reddit-cache.json), or puts a shell variable on a ./secretcurl line (-d "$PAYLOAD", $KEYWORD), and each refusal costs a model turn. So make one call per keyword with the keyword written in literally, pipe each request body into ./secretcurl ... --data-binary @-, save replies with -o /tmp/, and write the article and the log with Write/Edit.
Constraints
- Bot filter is critical. New accounts with high posting velocity dominate any keyword and are almost always inauthentic. Strict drop.
- Volume is more honest than sentiment. A
SPIKE(volume > 2x 7d avg) is a real signal; sentiment shifts within normal volume often aren't. - Engagement filters scale.
MIN_LIKES = [REPLACE: MIN_LIKES]is a starting threshold — raise it as the topic gains attention so noise stays out.