[REPLACE: SKILL_NAME] Skill
[REPLACE: CHANNEL_PLATFORM] 频道 [REPLACE: CHANNEL_NAME] 摘要——热门 [REPLACE: TOP_N_THREADS] 个话题帖 + 未决问题
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. Override the channel name. If empty, summarises [REPLACE: CHANNEL_NAME].
Today is ${today}. Read the last 24h of activity in [REPLACE: CHANNEL_PLATFORM] channel [REPLACE: CHANNEL_NAME] and produce a community digest.
Required secrets per platform
| Platform | Secrets | Notes |
|----------|---------|-------|
| discord | DISCORD_BOT_TOKEN, DISCORD_CHANNEL_ID | Bot must be in the server with View Channel + Read Message History. |
| telegram | TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID | Bot must have been added to the chat (private chats and groups both work). |
| slack | SLACK_BOT_TOKEN, SLACK_CHANNEL_ID | Bot needs channels:history (public) or groups:history (private) scope. |
Read the "Keys set for this run" line in your run notes, never a shell test (those are refused and cost a turn). If the secrets for [REPLACE: CHANNEL_PLATFORM] are not on it, log COMMUNITY_NO_TOKEN and exit cleanly.
Steps
- Resolve channel - the channel is
${var}if set, else[REPLACE: CHANNEL_NAME]. A channel or chat ID is not a credential, so./secretcurldoes not fill a{DISCORD_CHANNEL_ID}style placeholder: print the ID with its own call,node -e 'console.log(process.env.DISCORD_CHANNEL_ID)'(orSLACK_CHANNEL_ID), and write the printed value into the call literally (thebelow). The exact API call depends on[REPLACE: CHANNEL_PLATFORM], one call each:
- discord:
./secretcurl -s -o /tmp/community-messages.json -w 'discord http=%{http_code}\n' -H "Authorization: Bot {DISCORD_BOT_TOKEN}" "https://discord.com/api/v10/channels/<CHANNEL_ID>/messages?limit=100"
- telegram -
getUpdatespolling has limits; better to read the bot's stored offset state frommemory/topics/[REPLACE: SKILL_NAME]-tg-offset.json(Read it, then write the offset in literally):
./secretcurl -s -o /tmp/community-messages.json -w 'telegram http=%{http_code}\n' "https://api.telegram.org/bot{TELEGRAM_BOT_TOKEN}/getUpdates?offset=<OFFSET>&limit=100"
- slack - get the window start with its own call,
date -u -d '24 hours ago' +%s, and write it in as:
./secretcurl -s -o /tmp/community-messages.json -w 'slack http=%{http_code}\n' -H "Authorization: Bearer {SLACK_BOT_TOKEN}" -G "https://slack.com/api/conversations.history" -d channel=<CHANNEL_ID> -d oldest=<OLDEST> -d limit=200
- Filter to the last 24h — drop messages older than
now - 24h. Drop bot messages (most platforms exposeis_bot/bot_id). Keep replies.
- Cluster into threads — Discord and Slack expose explicit thread/parent IDs; Telegram doesn't. For Telegram, cluster by reply-chain hops.
- Score threads — for each thread:
- Reach — unique participants × log(message count).
- Question pressure — does the parent message end with
?or contain words like "how", "why", "anyone", "stuck"? +5. - Recency — within the last 12h gets full marks.
Pick the top [REPLACE: TOP_N_THREADS] threads.
- Detect open questions — scan all parent-level messages. If a message ends with
?and has zero replies after 6 hours, mark it asOPEN_QUESTION. List them separately so the operator can chase.
- Write
output/articles/[REPLACE: SKILL_NAME]-${today}.md:
# [REPLACE: CHANNEL_NAME] — ${today}
**Volume**: N messages from M participants (vs 7d avg of K).
## Top [REPLACE: TOP_N_THREADS] threads
1. [Author · timestamp] "Parent message excerpt..."
→ N replies, M reactions
→ Permalink
...
## Open questions (no reply > 6h)
- [Author] "Question text..." → permalink
## Volume sparkline (last 7 days)
▁▂▃▅▇▆▄
- Notify via
./notifywith a 3-line summary:
*[REPLACE: CHANNEL_NAME] — ${today}*
N messages · M participants · K open questions · top thread: <one-line>
Full digest: <url>
Skip the notification on quiet days (volume < 25% of 7d avg AND zero OPEN_QUESTIONs).
- Log — append to
memory/logs/${today}.md:
## [REPLACE: SKILL_NAME]
- **Channel**: [REPLACE: CHANNEL_PLATFORM]/[REPLACE: CHANNEL_NAME]
- **Volume**: messages=N, participants=M, vs_7d_avg=Δ%
- **Threads picked**: [REPLACE: TOP_N_THREADS] (of K candidates)
- **Open questions**: N
- **Status**: COMMUNITY_OK | COMMUNITY_QUIET | COMMUNITY_DEGRADED (api errors)
Network note
Discord and Slack need their bot token in the Authorization header; Telegram puts it in the URL path (/bot{TELEGRAM_BOT_TOKEN}/). A bare $TOKEN on a curl line is refused by the Bash permission analyzer - so call ./secretcurl with a {ENV_NAME} placeholder instead. The token (injected via the skill's requires:) is substituted inside the helper and never appears on the command line, as in the step 1 calls.
Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (CHANNEL=...), uses if / for, tests a key, joins commands with ;, && or ||, redirects into a file, or puts a shell variable on a ./secretcurl line ($DISCORD_CHANNEL_ID), and each refusal costs a model turn. So write IDs, offsets and dates into each call literally, save replies with -o /tmp/, and write the offset file, the article and the log with Write/Edit.
Constraints
- Privacy. Don't quote DMs or anything from a private channel verbatim into a public notification — paraphrase, attribute by display name only, never include user IDs.
- The "open questions" section is the most useful output. Even on a quiet day, if there's an unanswered question, surface it. That's the operator-action signal.
- Avoid bot-on-bot loops. If this skill notifies into the same channel it reads from, filter your own bot's messages out before scoring.