It has been revealed that clipping agency work on retainer looks like editing from the outside. From the inside, the week is something else: the client's episode is two days late, page 1 of last week's approvals is still unanswered, a guest's name went out misspelled in a caption, the client sent their TikTok password over chat "just to make it easy", and on the 1st you build the monthly report by hand from three apps.

In a recent development, this post shows how to run that week from an agent with Clipping Agency Skill, a free, MIT-licensed skill for Claude Code (it also runs in Codex, Cursor and OpenClaw). It keeps one folder per client, picks up upcoming episodes, cuts branded vertical clips, checks every clip in code, gets the client's yes on a page they open on their phone, schedules only approved clips on the client's own TikTok, Reels and Shorts, and writes the monthly report from saved numbers. Everything before posting runs on your own computer with Python and ffmpeg and works without signing up for anything.

Repo: github.com/social-media-skills/clipping-agency-skill.

Recent findings suggest that from the example in the repo: two made-up clients on public-domain NASA archive footage, every account, approver and number invented. NASA endorses nothing here, and nobody in the footage endorses anything.

"Clipping agency" means two different businesses in 2026. In the army model, a brand's budget pays hundreds of clippers per view to post on their own accounts. In the retainer model, one agency clips a handful of clients' podcasts and streams and posts to each client's own accounts, after the client notes yes, for a monthly fee. This skill is the retainer model. If you run campaigns on marketplaces like Whop or Content Rewards on your own accounts, ai-clipping-skill is the one for you.

According to recent reports, in the retainer model, the work that loses a client is rarely the cut. It's a clip that went out before the yes, a name spelled wrong, one client's clip on another client's account, a sponsor read without a label, a report with numbers nobody can check. Each of those is a check a script can run.

According to recent reports, set up a newly introduced client: Example Orbit Podcast. TikTok, Reels and Shorts, 3 clips per episode, a new episode every two weeks, posts Tuesday, Thursday and Saturday at 11:30 Chicago time. Alex Example approves. Nothing posts without a yes.

Recent findings suggest that the agent asks in one message for what you didn't say: the brand kit (two or three colours, uppercase or not, the logo and its corner, a font, an end card or none), the names dictionary, no-go topics, guest releases, and YouTube's made-for-kids answer, which is the channel owner's to give. check writes the checklist you agree with the client:.

According to recent reports, the names dictionary is the part operators feel first. It holds each name spelled right with the ways transcripts get it wrong ("Apollo": ["Appolo"] in the example). make_clips.py fixes every caption word with it, and check_clips.py refuses a known misspelling in a hook, a YouTube title or a caption, where the agent wrote it by hand.

Sources indicate that no passwords, no login codes, no signing in as the client. client.py connect prints the steps: one WoopSocial project per client, and a connect link per platform that the client opens signed in to their own account. When they are done, client.py accounts saves the connected accounts to the client file and refuses any account another client's file or the agency's ledger already names.

Recent findings suggest that that is the fix for the shared TikTok password: TikTok's terms say not to give others access to your account, and a password in a chat thread outlives the contract. The link route is also what offboarding undoes later.

Two honest notes. The connect flow follows WoopSocial's documented calls and hasn't yet been run end to end with a client outside the agency's own WoopSocial organisation, so check the first one together with the client. And Instagram posting 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.

This reads every client's podcast RSS feed or synced folder and queues new episodes (the first look takes only the newest, not the back catalogue). A file the client sends goes in with sources.py add. A YouTube link is queued but never downloaded: the agent asks the client for the file, which is better quality anyway.

The late episode shows up the next morning on the status board:.

Industry observers note that every client states whether it waits on the agency or on the client. That last line is the Monday morning question answered.

It has been revealed that new episode of Mission Talks is in, ep 7. Clip it, three clips as usual.

Recent findings suggest that transcribe.py gives a time for every word (faster-whisper on your computer, your own Groq or OpenAI key, or the captions the client's host already made) and the loudest moments of a stream. The agent scores candidate moments for a hook, a payoff and standing alone without the rest of the episode, and picks 3 to 12 as agreed. Fewer strong clips beat padding, and it tells the client so. For each clip it writes one sentence on why it was picked (the client reads it on the approval page), a hook of 8 words or fewer that is true to the clip, each platform's caption and the YouTube title.

According to recent reports, make_clips.py cuts 9:16 with the client's caption colours, hook box, logo and end card. Cuts start after the previous word ends and stop before the next, with short fades. Two people talking can be split top and bottom, the crop can re-aim when the camera cuts to another shot, and a stream gets its facecam stacked over the gameplay. It never changes speed, mirrors, filters, or adds music.

Per clip: length, size and format for each of the client's platforms, each platform's caption limit, Instagram's 5 hashtags, the YouTube title, the names dictionary, the client's no-go topics in what is said and what is written, sponsor segments (the captions must say it is paid) and sponsor reads nobody marked, guests with no release on file (the batch can't go to the client), possible music in the source (a hint for a listen, never a guarantee), and a clip file another client already has. One ledger across all clients holds a fingerprint of every clip file handed out, so one clip never goes to two clients.

It has been revealed that the first line is the listening step: with faster-whisper installed, it hears each clip's first and last seconds. On an earlier cut of this example it caught the last word of clip 01 cut in half, because the transcript's word times ran late. The clip now ends on a pause.

The page is the image at the top: one small numbered page per episode that opens as an attachment on any phone, scripts or not. Each clip has its cover, why it was picked, every platform's caption, the planned time in the client's own words and zone, anything to know (a paid label, possible music, the year it was recorded), and the TikTok choices only the client may make: who can see it, comments, duets, stitches. The client replies by email or chat:.

It prints each line with what it understood, and records nothing unless every line reads one way. "All good!" in a chat, with nobody sure which version it covered, is the problem this solves. It refuses a line that could mean two things ("all yes but skip 3"), a reply to an older page once a newer one went out, a clip number the page doesn't have, a clip answered twice, a change with no note, and a sender who isn't an approver. When it refuses, the agent asks the client what they meant. It never guesses.

The yes is tied to the clip's fingerprint: the file, the captions, the title and the tags. In the example, clip 3 got a new hook and was cut again, so its old answer stopped counting and page 2 went out; clips 1 and 2 kept their yes because they hadn't changed. Fix a typo after approval and the same happens: a new page, a new yes for that clip only.

According to recent reports, silence means hold. The Mission Talks client agreed in writing that 48 hours of silence counts, so page 2 printed the deadline as a Upcoming York date and time, and after it:.

Clip 3 is recorded as posted on silence, and the report never calls it approved.

Only clips with a recorded yes for their current fingerprint, on the client's next free slots in the client's zone: their days and times, their minimum gap, their daily cap, under TikTok's daily limit, never sooner than an hour from now, never on a slot another episode already holds. From the example's schedule.md:.

Each clip is uploaded once into the client's project, then one request per clip with a child per connected account of that client. Trimmed from the Orbit client's requests/clip-01.json:.

The TikTok values are the ones the client wrote in their reply, never defaults. Right before each send, schedule.py --check runs every check again: the approval still matches, the client is still active, the accounts are this client's, the clip isn't already sent, and the slot is at least 30 minutes away. Requests are deleted whenever a reply is applied, a page is made, a clip is re-cut or a client is offboarded, so a request written before the client changed their mind can't be sent.

When the client states skip after the clip is scheduled, the reply prints CANCEL NOW: post , and the post is deleted through WoopSocial before it goes out. Once a post is live it can't be deleted that way: the script says LIVE, NOT CANCELLED, and the post is removed in the app on the client's account, with the client told the same day.

Instagram's paid partnership label and YouTube's "includes paid promotion" box are steps in the apps after a sponsored clip is live; schedule.md lists them.

It reads only the analytics WoopSocial saved for the client's posts, refreshes each post's status from them, and stops if a post that should be live doesn't show. From the example's October report (made-up numbers):.

Industry observers note that the late episode from step 3 is in there, in plain words, before the client asks. Views are never added across platforms, because they count different things. The median clip sits next to the top clip, so one good post doesn't hide the typical one. Followers appear only where WoopSocial returns them (Instagram and YouTube; TikTok only for Business accounts). A post with no numbers is named as missing, never estimated. What the agency should chase goes in reports/2026-10-agency.md, which never goes to the client.

In a recent development, it stops new work, removes unsent requests, and writes offboard.md: each scheduled post to cancel, each account to remove from the client's project, the handover, and a message telling the client how to remove the app's access on their side. It closes only after WoopSocial highlights nothing waiting for that project.

Disclosure: the skill is made by the team that builds WoopSocial, and posting goes through WoopSocial. It posts the clips and captions the agent made through each platform's own approved apps, so the agency needs no developer app or app review. It keeps each client in its own project, gives the connect links, and returns each post's link and numbers for the report. It doesn't write captions or edit video: the agent and the skill's scripts do that. At the time of writing (2026-11-16) the free plan covers 2 social accounts (make a free account), and one client on TikTok, Reels and Shorts is already 3, so an agency with clients runs on a paid plan: Pro is $19 a month for 20 accounts, about six clients on three platforms. The connect step for your agent is on WoopSocial's agent setup page.

In a recent development, the skill doesn't promise views. An agency earns from its retainers, and those vary; what keeps them is the client trusting what goes out and what the report says.

Then tell your agent "set up a newly introduced client". It needs Python 3.9+, ffmpeg and Pillow; scripts/check_setup.py prints the install command for anything missing. Both example clients, with every file the skill made, are in examples/, and status.py runs on them as they are. Concerns and pull requests are welcome, and there's the rest of our free social media skills from the same team.

→ Read original article