一键导入
local-schedule
In-session scheduling via CronCreate. For "at <time>", "in <duration>", "remind me", "tomorrow". Defer to /schedule for remote.
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。
菜单
In-session scheduling via CronCreate. For "at <time>", "in <duration>", "remind me", "tomorrow". Defer to /schedule for remote.
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
| name | local-schedule |
| description | In-session scheduling via CronCreate. For "at <time>", "in <duration>", "remind me", "tomorrow". Defer to /schedule for remote. |
Route time-anchored task requests to CronCreate instead of the remote /schedule skill.
| Task needs… | Tool |
|---|---|
| Files / processes / cwd / state from THIS session | CronCreate (this skill) |
| To survive Claude restart | CronCreate with durable: true |
| To run as a remote agent independent of this session | /schedule |
| "Every Nm run /slash-cmd" — bare slash-command on an interval | /loop (tighter match for that exact pattern) |
| Pure prompt with no local dependency, user didn't specify | CronCreate (default — cheaper, faster, no remote spin-up) |
If unsure, prefer CronCreate. Local scheduling can do everything /schedule can do for an idle session; the reverse is not true. The exception is /loop — it's purpose-built for repeatedly invoking a slash command, so when the user's phrasing is essentially every <interval> /<command>, defer.
Call CronCreate directly. The tool's own description has the cron syntax and runtime details — read it before composing the call. The pieces that matter most:
0 9 * * * is 9am wherever the user is.recurring: false with minute/hour/day-of-month/month all pinned. For "at 22:00 today": resolve today's date, build MM HH DOM MON *, set recurring: false. The job auto-deletes after firing.recurring: true (the default) and auto-expires after 7 days. Tell the user about the 7-day cap when they ask for an open-ended recurring job.0 or 30. Every model defaults to :00 and the fleet collides. Use 57 8 * * * or 3 9 * * * for "around 9am", 7 * * * * for "hourly". Only honor :00/:30 when the user names that exact time.durable: true only when the user explicitly asks the task to survive a session restart. Default in-memory mode is correct for "tonight" / "tomorrow morning" / "every 5 min for a while".The user's prompt-submit hook injects the current wall clock as Message time: YYYY-MM-DD HH:MM:SS Weekday. Use that as the anchor — do not guess.
recurring: false.0 22 ... (user named :00 exactly, so :00 is fine).57 8 or 3 9, recurring: false.28 9 * * 1-5 (off-minute; user can correct if they meant 9:30 sharp).If the message-time hook is absent, ask the user for the current time rather than guessing — cron mistakes are silent until the fire moment.
prompt fieldThe prompt fires as a fresh user turn in this session at fire time. Write it as if the user just typed it: enough context that the model can act without re-reading scrollback. Example:
Run the regression test on branch <name> that we pushed at <time>. Logs go to <path>. Report pass/fail and any new failures vs main.Don't write "do the thing we discussed" — fire-time context is just this session's history; the model can read it but a self-contained prompt is more reliable.
Tell the user, in one line: what was scheduled, the resolved fire time in their local clock, and the cron expression you used. They need to see the cron string to catch a misread time. Save the returned job ID if you might need to cancel it later (CronDelete).
Hand off to /schedule only when the user explicitly indicates one of:
durable: true isn't enough (e.g., they want it on a true cron schedule for weeks/months).In those cases, mention you're switching to /schedule and why, so the user can correct you if the locality assumption was wrong.
Use the `bilibili-api-python` async library (400+ endpoints) for any Bilibili programmatic operation — videos, users, favorites, danmaku, live rooms, comments, search, uploads. This skill should be used whenever user asked for any Bilibili operation.
Fetch authoritative, current, version-specific documentation for libraries, SDKs, APIs, and CLI tools via Context7. MUST TRIGGER before writing or editing any non-stdlib library call — including well-known libraries you think you know, since your prior knowledge may be stale. SKIP for internal/private libraries, stdlib code, raw HTTP/filesystem, shell built-ins, or stable primitives whose signatures haven't changed in years (e.g. `json.dumps`, `os.path`). Prefer this over web search for library docs.
Query AI-generated overview of GitHub repositories via DeepWiki MCP. Use when needing a fast high-level understanding of a repository's architecture and design decisions without fetching full repository.
Periodic distillation: scan the last 5 days of high-quality Claude transcripts in the current project, dedup facts seen across sessions and against existing docs, and propose surgical edits to the right reference docs. Wider-than-daily window avoids 00:00 boundary loss. This skill should be used when the user says "distill recent", "distill session", "distill last week", or asks "what should we learn from recent work".
Extract atomic flashcards from documents/conversations and write them into local Anki via AnkiConnect, using a React-rendered paper-note card template with markdown, collapsible code, KaTeX, and mermaid. Triggers on "flashcards", "flash", "抽认卡", "anki".
This skill should be used when the user sends an image and asks to "analyze this image", "describe this picture", "what's in this image", or any request requiring visual understanding of images. Provides image analysis using Zhipu GLM-4.6V multimodal model.