Captions From A Transcript . Agency OS

What it does

Takes a video transcript and returns up to eight copy-paste ready outputs: TikTok, Instagram, Facebook, YouTube Shorts, YouTube Long-Form, LinkedIn, a Threads chain and an X thread. You can ask for only the platforms you post on. Each one is written natively for its platform, with the primary keyword front-loaded into the zone that platform actually indexes, hashtag counts that follow that platform's rules, and character counts inside its limits. YouTube gets separate title, description, hashtag and tag fields. Instagram gets alt text and a format recommendation. It writes in your voice by taking the voice signal from the transcript itself first, ahead of any brand file. Before writing a word it settles the call to action: direct, soft or none.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

> write captions

> caption this

> post this video

> social captions

> write a caption for

When to reach for it

You have a transcript and need captions for every platform
Posting one video across TikTok, Instagram, Facebook, YouTube, LinkedIn, Threads and X
Your video ends "comment WORD and I'll send you the guide" and a comment automation is waiting for that word
You need YouTube title, description and tags written properly
You need Instagram alt text and a Reel versus Carousel call
Captioning a video that is part of a series
Captioning a value-only video with no pitch in it

When NOT to use it

If you actually want	Use this instead
Writing a social post from scratch with no transcript to work from	`legendary-copywriters`
Paid ad copy for Meta, Facebook or Instagram	`meta-ads-copywriter`
A VSL or sales video script	`vsl-architect`
Turning one input into a full multi-asset set of carousels, threads and posts	`content-repurposer`
Cutting a long video into vertical clips	`video-repurposer`
A long-form SEO article	`blog-writer`
The 30-day multi-tier posting plan the captions slot into	`content-system-architect`

Before you start

What you need	Why
A transcript, as text or a file path	This is the primary input and the primary voice source. Without it there is nothing to caption and the skill has no fresh voice signal to work from	Required
`brand/[workspace]/verified-claims.md` with entries marked `consent: ok-as-stated`	Numbers, results and stories spoken in the video still have to trace to the registry before they appear in a caption. A hook is not exempt	Required
`positioning.md` with your actual offer names and CTAs	Offer alignment pulls real offer names from it. Without it the skill has no offer stack to match the video topic against	Optional
`voice-profile.md` and any voice guide files in `knowledge/`	They back up the transcript voice. The transcript still takes precedence, so these matter most when the transcript is short	Optional
`keyword-plan.md`	It supplies the keywords the captions layer in. Without it the skill derives keywords from the transcript alone	Optional
The video file and ffmpeg installed, if you want visual context	It extracts a frame every five seconds to see screen recordings, whiteboards and text overlays. Without it the skill works from words only and cannot reference what is shown	Optional
Web search available	It researches current algorithm behaviour and hashtag patterns per platform. If unavailable it notes that research was skipped and falls back to its built-in rules	Optional

How it runs

Read the transcript in fullBefore anything else it reads the whole transcript, and reads extracted frames if you gave it the video. It pulls the core topic, the strongest hook, three to five key moments, stories, data points, the emotional tone and the keywords you repeat naturally.
Settle the call to actionThere are three modes. Direct is an explicit ask, such as 'comment GUIDE and I'll send you the guide' or a link. Soft is a nudge with no ask, such as 'save this for later'. None means the caption ends and nothing replaces the ask. If you name a mode, yours wins. If you said an ask on camera, the mode is direct and your exact words are kept, keyword included, so a comment automation still fires. Otherwise it proposes a mode in one line and carries on.
Match the ask to each platformIn direct mode, Instagram and Facebook (or wherever your comment automation is live) end on your spoken line word for word. Platforms where you confirm you reply to keyword comments by hand get the same line, with no promised reply time. Platforms with neither never promise a DM nobody will send: they get a link you confirm, such as the guide link in the YouTube description, or a soft ending. It asks you once if it does not know.
Align to an offer, or do notIt matches the video topic against your offer stack from positioning.md and picks the CTA style that fits. This runs alongside the research step. It will not force an offer where none fits naturally, and the alignment note is for you, not for publishing.
Research the platforms and the keywordsIt searches for current algorithm and hashtag behaviour per platform, then sets one primary keyword and three to five secondary keywords per platform, because the same video is searched differently on TikTok than on YouTube.
Write every caption in parallelIt launches one sub-agent per output. Each gets the content analysis, offer alignment, platform trends, keywords, voice profile and that platform's playbook, and writes independently as native content rather than as a reformatted copy of the others.
Run the quality gateChecks before delivery, including: does every caption sound like you rather than a social media manager, does every hook name its audience in the first line, is the primary keyword early on every platform, are hashtag counts and lengths within each platform's rules, is the chosen call to action mode honoured on every platform with your spoken line kept word for word, and does a value-only video contain no hidden pitch.
Run the claim checksIt lints the captions against your verified-claims registry and spawns a claim grader over the final text. A failure blocks the caption until the flagged claim is grounded or made generic.
Deliver everything at once and learnAll outputs arrive in a single response, never one platform at a time. It then asks whether you shipped as-is, made minor tweaks, rewrote significantly or scrapped it, and which platform performed best. That feedback goes to learnings.md, and any pattern seen three times becomes a rule.

What you get

A content analysis block: topic, audience, offer alignment, CTA detected, CTA mode, primary keyword and emotional tone
TikTok caption with target keywords, hashtags with rationale and character count
Instagram caption plus alt text, a Reel versus Carousel versus Static recommendation with reasoning, keywords, hashtags and character count
YouTube Shorts title, description, hashtags, 8 to 12 tags and target keywords
YouTube Long-Form title, description with keyword-labelled timestamps and CTA, 3 hashtags, 10 to 15 tags
LinkedIn post with a specific engagement question, hashtag rationale, character count and format note
Facebook post with hashtags, character count and a link placement note
Threads lead post with a numbered reply chain, and an X thread, with character counts per post
A captions-v1 schema artifact other skills can consume, and a log entry in brand/assets.md

Honest limits

Read this before you rely on it

No transcript, no captions. This skill starts from something already said on camera, and writing a post from an idea rather than from a recording is a different job. Non-English video is refused rather than guessed at: it asks you instead of auto-translating.
Specifics are capped by your registry, hooks included. If you say a number or tell a client story on camera and it is not in verified-claims.md with consent, the caption comes back generic with a proof gap flag rather than repeating it. The skill states plainly that a hook is not exempt, so expect your best spoken line to be the one that gets flagged.
The claim checks do not see chat. The linter fires on file writes, and captions are almost always copied straight out of the conversation into a phone or a scheduler. That means the model grader is the only tier that ran. Read the flags before you paste.
Platform rules are a snapshot, not a live feed. Character limits, hashtag counts and algorithm behaviour are researched at run time where web search is available, and fall back to built-in rules when it is not. Verify anything that looks stale.
It does not post, schedule or publish anything, and it does not cut the video. You get text, and how much of the research runs depends on where you run it: on Claude Code it uses sub-agents and web research, on Claude.ai it runs in manual mode with no parallel research.

Where people go wrong

The mistake	Do this instead
Feeding it a cleaned-up summary instead of the raw transcript	Give it the full unedited transcript. The transcript is the primary voice source and outranks your brand voice file, so a polished summary strips out exactly the phrasing and rhythm it needs.
Asking it to add a CTA to a pure value video	Leave it value-only. Shoehorned offers are explicitly against the skill's rules, and the quality gate checks that a value-only video carries no hidden pitch.
Overriding the CTA the video already contains	If you said 'comment AGENCY' on camera, that exact CTA belongs in the caption. The skill preserves it verbatim for a reason: it has to match what the viewer just heard, and your comment automation listens for that word.
Promising a DM on a platform with no automation	Tell it where the automation is live and where you will reply by hand. Everywhere else gets a link you confirm or a soft ending, never a promise nobody keeps.
Asking for one platform at a time	Let it deliver them all at once. The sub-agents run in parallel, so asking for them one by one costs you time and produces less consistent keyword coverage.
Putting the link in the Facebook post body	Keep outbound links out of the body. The quality gate checks this, and the link placement note tells you where to put it.
Treating the offer alignment note as caption copy	That block is for you, not for publishing. It explains why the skill matched the video to a given offer.

Worth knowing

The transcript outranks every brand file for voice, so the rawer the transcript the better the captions. If you were fired up on camera, the captions come back fired up. Send the verbatim transcript including the false starts and the asides, and send the video file too if you have ffmpeg, because the frames let the captions reference what was actually on screen rather than only what was said.