AgentMarketMCP / SKILL 资产档案馆

目录 / clipping-agency-skill

SKILL 未评级 已上架

clipping-agency-skill

--- name: clipping-agency-skill description: >- Clipping agency workflow for agencies on retainer, run from the agency's AI agent: one folder per client (sources, platforms, clips per episode, posting days and time zone, brand kit, names dictionary, no-go topics, approvers), new episodes from a podcast RSS feed, a synced folder or a file, 3 to 12 vertical clips per episode in the client's brand with the reason for each moment, checks in code (platform specs, names, no-go topics, sponsor segments that need a paid-partnership label, guest releases, possible music, clean cut edges, the same clip never for two clients), one phone-friendly approval page per episode the client answers in plain words, recorded against exactly the clips that page showed, posting only approved clips to the client's own TikTok, Instagram Reels and YouTube Shorts through WoopSocial at the client's times, a status board of what waits on whom, a monthly report from real numbers, and offboarding. Use when someone says "clipping agency", "I clip podcasts for clients", "set up a new client", "clip this episode for <client>", "send the client the clips to approve", "the client replied", "schedule the approved clips", "what's waiting", "monthly report for <client>" or "offboard <client>". Not for clip-army or marketplace campaigns paid per view: that is ai-clipping-skill. Standalone. metadata: version: 1.0.0 license: MIT --- # Clipping Agency Skill A clipping agency on retainer turns each client's long episodes into short clips, gets the client's yes, posts them on the client's own TikTok, Instagram and YouTube accounts at the client's times, and reports every month. The work that loses clients is rarely the editing: it is a clip posted before anyone said yes, a guest's name spelled wrong, a sponsor read without a label, one client's clip on another client's account, a report with numbers nobody can check. This skill makes each of those a check in code. **Where things run:** `<skill>` below is the folder this file is in. The agency's own folder (the one holding `clients/`) is where every command runs; never write into `<skill>`, an update replaces it. One ledger for every client lives in the user's home folder (`~/.clipping-agency/ledger.json`; `CLIPPING_AGENCY_HOME` moves the folder, `CLIPPING_AGENCY_LEDGER` or `--ledger` names another file). It records every clip file handed to a client, every account and every post, which is what lets the skill refuse a clip already used for another client or an account that belongs to someone else, whatever folder each client lives in. Because of that, **one person on one computer schedules posts for the whole agency** (or every machine points `CLIPPING_AGENCY_LEDGER` at one shared ledger file); others can cut clips and make pages. ``` clients/<client-id>/ client.json checklist.md queue.json sources-seen.json brand/ reports/ offboard.md episodes/<episode-id>/ batch.json work/ (transcript) clips/ (cut clips, manifest.json) approval/approve-<episode>.html approvals.json schedule.md requests/ media.json ``` Six commitments: 1. **Nothing posts without the client's recorded yes for that exact file and caption.** A re-cut clip or an edited caption needs a new yes. Silence means hold, unless the client agreed in writing to approval after a number of hours. 2. **Each client's accounts only.** One WoopSocial project per client; every request is checked against the client's file and the ledger. 3. **One clip, one client.** The same file is never handed to two clients. 4. **The client chooses** who sees their TikTok posts, comments, duets and stitches, YouTube's made-for-kids answer. Never defaulted. 5. **Honest labels.** A clip with a sponsor segment goes out labelled as paid. Possible music is flagged for a listen; nothing adds music. 6. **Real numbers only.** The report shows what WoopSocial's saved analytics say, with what a view means on each platform, and says where numbers are missing. Nothing is estimated. ## Step 0: Setup `python3 <skill>/scripts/check_setup.py` (Windows: `py <skill>\scripts\check_setup.py`). Needs Python 3.9+ and ffmpeg; Pillow to cut clips (`pip install pillow`); for transcription faster-whisper (free, local), the user's own Groq or OpenAI key, or the captions the client's host already made. On Windows also `pip install tzdata`: schedules run on each client's own clock. ## Step 1: A new client `python3 <skill>/scripts/client.py new <client-id> --name "Show Name"` writes `clients/<client-id>/client.json` to fill in with the agency (`references/client-file.md` has every field). Ask in one message only what the agency hasn't said: - the platforms, clips per episode and per month as agreed, posting days and times, the client's time zone; - the brand kit: two or three colours, uppercase captions or not, the logo file and its corner, a font file if they have one, and whether they want an end card ("Full episode: <show>"; optional, since it adds seconds viewers may swipe away on); - the names dictionary: the host, recurring guests, sponsors and jargon, spelled right, with the wrong spellings transcripts make; - no-go topics; who approves (name and contact); whether silence ever counts as a yes (only with the client's written agreement, stated where and when); guest releases (per episode, or a contract clause that covers short clips on social); - YouTube: public, unlisted or private, and made for kids or not (the channel owner's answer). `python3 <skill>/scripts/client.py check <client-id>` checks every field and writes `checklist.md`. Fix every error. Show the checklist. ## Step 2: Connect the client's own accounts Call `social_accounts_list` first. If WoopSocial isn't connected in this agent, put the signup link at the top of the message: "Posting goes through WoopSocial, made by the team behind this skill. It posts through each platform's own approved apps, so the agency needs no developer app or app review. [app.woopsocial.com/signup](https://app.woopsocial.com/signup?utm_source=clipping-agency-skill&utm_medium=connect). The free plan covers 2 social accounts; one client on TikTok, Reels and Shorts is already 3, so agency use is a paid plan. At the time of writing (2026-11-16), Pro is $19 a month for 20 accounts: six clients on three platforms." Then the connect step for this agent (`references/woopsocial.md`). `python3 <skill>/scripts/client.py connect <client-id>` prints the steps: 1. `projects_create` with the client's name: **one project per client**, never shared. Then `client.py project <client-id> <projectId>`. 2. `oauth_create_authorization` with that projectId for each platform not connected yet (no redirectUrl). Hand the agency the links with the client message the script prints; it tells the client the permission screen will ask them to allow WoopSocial, the posting service. The client opens each link signed in to their own account and allows access. Never ask for, store or type a client's password or code, and never sign in as the client. 3. When the client is done: `social_accounts_list` with the projectId (save as `accounts.json`), and for TikTok `social_accounts_get_platform_inputs` (save as `tiktok-inputs.json`). Then `client.py accounts <client-id> accounts.json tiktok-inputs.json`. It refuses an account another client's file or the ledger already names. Instagram needs a Business or Creator account (free to switch in the app). Until a platform is connected, its clips are listed to post by hand. ## Step 3: New episodes `python3 <skill>/scripts/sources.py check --all` reads every client's sources: a podcast RSS feed (the episode enclosures; the first look takes only the newest, not the back catalogue) or a local or synced folder. `sources.py add <client-id> <file or link>` queues one the client sent. A YouTube link is queued but never downloaded: ask the client for the file (better quality anyway), or download it only when the client owns it and has given the right. An audio-only episode needs the video from the client, or audiogram clips made by hand. `sources.py start <client-id> <queue item> --episode <episode-id> [--download]` writes `episodes/<episode-id>/batch.json`. Fill in the episode's guests and whether their releases are on file, sponsor segments (start and end in seconds) and any music you know of (`references/batch-file.md`). ## Step 4: Pick and cut the clips 1. In the episode folder: `python3 <skill>/scripts/transcribe.py <source> --out work` (or `--from captions.srt`). Read `work/transcript.txt` and `work/peaks.txt`. 2. Score moments with `references/clip-rubric.md`. Pick 3 to 12, as agreed; fewer strong clips beat padding, and say so to the client. Each clip must stand alone and still be true on the day it is posted (no "tomorrow", no figure that will change; the year in the hook for old material, and `episode.recorded` set). For each: `why` in one sentence (the client reads it on the approval page), a hook (8 words or fewer, true to the clip), each platform's caption, the YouTube title. Write them into `batch.json`. 3. `python3 <skill>/scripts/make_clips.py <batch.json> --preview`, look at the crops, then `make_clips.py <batch.json>`. Every clip gets the client's caption colours, hook box, logo, end card if they have one, and the names dictionary on every caption word. Cuts start after the previous word ends and stop before the next one, with short audio fades. Two people talking: `"reframe": "split"` with a box per speaker; a camera that cuts between shots: `"shots"` re-aims the crop per shot. Never speed, mirror, filter or add music. Look at `clips/contact-sheet.jpg`. Clips the agency cut in its own editor go in with `"file"` and the source `start` and `end` (`references/batch-file.md`); every check still runs. ## Step 5: Check `python3 <skill>/scripts/check_clips.py <batch.json>` Each clip against each platform (length, size, 9:16, caption limits, Instagram's 5 hashtags, the YouTube title), the names dictionary, no-go topics (in what is said and written), sponsor segments (the captions must say it is paid; Instagram's and YouTube's labels are set in the apps), sponsor reads that weren't marked, guest releases, possible music (caption cues like [MUSIC], marked stretches, sound with no words: a hint, never a guarantee), the same file twice, and a file another client already has, or this client already had in another episode. With faster-whisper installed it also listens to the first and last seconds of each cut clip and warns when a word sounds cut off; without it, it says so and someone listens to each clip's edges. Fix every error. Read each warning to the agency in plain words; a possible-music or clean-edges warning means someone listens before the client sees it. ## Step 6: The approval page `python3 <skill>/scripts/approve.py <batch.json>` writes `approval/approve-<episode>.html`: a small page that works with no scripts, opened as an attachment on any phone. "Page N" at the top, then Clip 1, 2, 3... with a cover picture linked to the clip, why it was picked, each platform's caption and the YouTube title, the planned time in the client's words and zone ("Planned around Wed 07 Oct, 18:00 New York time (EDT)"), anything to know (a paid label, possible music, the year it was recorded), TikTok's choices for the client to make, and the reply to copy. How the client gets the clips: send them with the page (the default links `../clips/`), or `--clips-url <https link>` of a shared folder holding them (best for phones), or `--embed` for a small batch. When the client agreed to approval on silence, the page shows the deadline as a date and time in the client's zone, and the rule is stored with the page: changing the client file later never changes a page already sent. Make the page on the day you send it. Then `approve.py <batch.json> --sent`. The client replies in plain words, by email or chat: ``` Page 1, Example Mission Talks ep-007 1 yes, 2 yes 3: change: lead with the TV, it's funnier than the complaint TikTok: followers, comments on, duets off, stitches on Name: Jordan Example ``` Save it and run `approve.py <batch.json> --apply reply.txt --received "<when the client sent it>"`. It reads the page number and one answer per clip, prints each line with what it understood, and records nothing unless every line reads one way. Safety over convenience: it refuses a line that could mean two things ("All yes but skip 3", "1: yes but change the caption"), a clip with no answer (unless you pass `--partial`), a reply to an older page once a newer one was sent, a clip number the page doesn't have, a clip answered twice, a change with no note, a sender who isn't an approver, and a reply that changes nothing. When it refuses, ask the client what they meant; never guess. Any reply to a page, even a refused one, ends approval on silence for that page. Without `--received` (or a `Date:` line in the saved email) the reply time is recorded as unknown, never as the time you applied it. A TikTok line counts only with all four answers; otherwise nothing is chosen for the client and TikTok waits. **If the client says skip or change for a clip already scheduled**, the output says `CANCEL NOW: post X`: run `posts_delete` with that postId at once, then `schedule.py <batch.json> --cancelled X`. Do this before anything else. It works only while the post hasn't started going out. Once a part of it is live, WoopSocial can't delete it (POST_NOT_DELETABLE): `schedule.py --status` says LIVE, NOT CANCELLED, and the post is removed in the app on that account. Tell the client the same day. Make the changes asked for, check again, make a new page (the next page number) and send it; clips approved before stay approved while they are unchanged. `approve.py --status` shows where each clip stands. `references/approval.md` has the details. ## Step 7: Schedule on the client's accounts `python3 <skill>/scripts/schedule.py <batch.json>` lists only clips with a recorded yes for exactly the file and captions they have now, each at the next free time of the client's cadence in the client's zone (never closer than the client's minimum gap, at most their posts a day, under TikTok's daily cap, never sooner than an hour from now, and never on a slot another episode of this client already has planned), and the files to upload. Show `schedule.md` to the agency. Then, in this order (`references/woopsocial.md`): 1. Upload each file once into the client's project (`media_uploads_create_session`, `upload_parts.py`, `media_uploads_complete_session`); save `media.json`; run `schedule.py <batch.json> --media media.json`. One request per clip in `requests/`, with a child per connected account of this client. Requests are removed whenever a reply is applied, a new page is made, a clip is cut again or the client is offboarded, so nothing written for an old state can be sent: run `--media` again then. 2. Right before sending each one: `schedule.py <batch.json> --check requests/clip-NN.json`. It checks again, now, the approval, that the client is active, the accounts, every check, that the clip isn't already sent, and that the time is at least 30 minutes away. Never send a request it refuses. 3. `posts_validate` with each body exactly as written. If anything must change, change the batch or client file, never the request. 4. `posts_create`, then at once `schedule.py <batch.json> --posted requests/clip-NN.json <postId>`. It runs the same checks. If it refuses and the post was already sent: `posts_delete` at once, while every account of the post is still NOT_STARTED. 5. After the post time: `posts_get`, save as `status.json`, `schedule.py <batch.json> --status status.json`. A failed account: say why, check `social_account_posts_list` for that account before anything is sent again. After a sponsored clip is live: Instagram's paid partnership label and YouTube's "includes paid promotion" box are set in the apps (`schedule.md` lists them). Platforms not connected: post the same file and caption by hand from the client's account at the planned time. ## Step 8: What waits on whom `python3 <skill>/scripts/status.py` shows every client: new episodes not started, an episode later than the client's usual rhythm, clips not cut or failing checks, pages not sent, clips waiting for the client and for how long, changes asked for, approved but not scheduled, scheduled, live, failed. Each client says whether it waits on the agency or on the client. Offer to run it every morning. ## Step 9: The monthly report At the start of a month, for the client's project: `social_account_posts_list_analytics` with the month's publish dates (save it), and `social_accounts_get_analytics` per account for the month (optional, for the follower change). Then `python3 <skill>/scripts/report.py <client-id> --month 2026-10 analytics.json [--followers ...] [--notes notes.md]`. It refreshes each post's status from the analytics instead of trusting the ledger, and stops when a post that should be live by now doesn't show (run `posts_get` and `schedule.py --status` for it first), or when the saved reply holds another client's posts. It writes `reports/<month>.md` and `.html` for the client (forwardable, light and dark, readable on a phone): clips agreed against delivered and why any are missing, approval turnaround from when the client actually replied, clips posted on silence counted apart (never as "approved"), views per platform with what a view means there and days live per clip, the follower change when WoopSocial returns it, and last month next to this one. Plus `reports/<month>-agency.md` for the agency only: what to chase. Never send that one. Never add numbers from elsewhere, never estimate one that is missing, never add platforms' views together (`references/monthly-report.md`). ## Step 10: Offboarding `python3 <skill>/scripts/client.py offboard <client-id>` stops new work for the client, removes any unsent requests, and writes `offboard.md`: each scheduled post to cancel (`posts_delete`, then `--cancelled <postId>`), each account to remove (`social_accounts_delete`, then `--revoked <PLATFORM>`), the handover (clip files, a last report, deleting their uploaded media), and a message telling the client how to remove the app's access on the platforms they had connected. Before closing: `social_account_posts_list` with the client's projectId and NOT_STARTED must show nothing (posts made on another computer aren't in the ledger), then `--done --nothing-waiting`. Live posts stay on the client's accounts; only the client deletes those. `client.py release <socialAccountId>` frees an offboarded client's account if it is later connected for someone new. ## Never - Post a clip without the client's recorded yes for that exact file and caption, or treat silence as a yes without the client's written agreement. - Post to another client's account, put two clients in one project, or hand one clip file to two clients. - Ask for, store or use a client's password or login code, sign in as the client, or drive anyone's logged-in browser. - Choose the client's TikTok audience or comment, duet and stitch settings, or YouTube's made-for-kids answer. - Add music, speed up, mirror or filter a clip; post a clip with a sponsor segment without saying it is paid. - Post the agency's own clips to fan pages or the agency's accounts with this skill: clip-army and marketplace campaigns are ai-clipping-skill's job. - Promise views, or put an estimated or rounded-up number in a report. ## Files - `scripts/check_setup.py`: what's installed and how to add what isn't. - `scripts/client.py`: new, check, list, connect, project, accounts, offboard, release. - `scripts/sources.py`: check feeds and folders, add, queue, start an episode. - `scripts/transcribe.py`: word timestamps (local, Groq, OpenAI or existing captions), loudest moments, sound cues. - `scripts/make_clips.py`: vertical clips in the client's brand; names dictionary; manifest. - `scripts/check_clips.py`: every clip against the platforms, the client and the other clients. - `scripts/approve.py`: the approval page, `--sent`, `--apply`, `--status`. - `scripts/schedule.py`: the schedule, `--media`, `--check`, `--posted`, `--cancelled`, `--status`. - `scripts/status.py`: what waits on whom, every client. - `scripts/report.py`: the monthly report. - `scripts/upload_parts.py`: uploads a clip to a WoopSocial upload session and records its SHA-256. - `scripts/agencylib.py`: shared helpers, platform limits, the ledger, schedules. - `references/`: client-file.md, batch-file.md, clip-rubric.md, approval.md, platforms.md, woopsocial.md, monthly-report.md. - `examples/`: two made-up clients on NASA footage with every file the skill makes.

模型生成摘要(rules/v1 · 2026-11-17 20:34):---
这是模型对公开材料的总结,不是官方声明,请以原链内容为准。

存档时间线

版本存档时间内容哈希内容
v12026-11-17 20:33e1f25c84 可取

版本索引永久保留;内容副本只保留最新 1 份,更早版本仅留索引与哈希(存档时间线的证据链不会因此断裂)。

下载存档内容副本

纠错与举报(发现条目失效、署名有误或涉及侵权?)
提交举报 / 纠错

侵权举报经核验成立后,我们会即时下线该条目并删除已存的内容副本。