Help

Every step of a session, and what to do when one of them stops.

Before your first session

To begin, say this in your chat:

cut my video

Any way of saying it works — a stream, a podcast, an episode, a recording. Fetch takes them all the same way.

Claude will ask before Fetch does each thing. Tap Always allow and it stops asking.

That is Claude checking with you, not Fetch being unsure. You can change it any time in Claude's settings.

1 · Start

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_REGISTER)

Say “cut my video”. That is the whole setup. You do not need to say how long it is, what shape it is, or which way up — Fetch reads all of that from the file itself, and reads it better than anyone remembers it.

Please do not guess. Anything you say about the video is taken as fact and used instead of the file. A rough guess at the length makes every estimate wrong, and a wrong guess at the shape can put a blurred panel across footage that never had one.

Setting up is free — a session is only spent when you approve a render.

You get a session code like unk-260901: your channel slug, then the date. Every clip from it is numbered from that code — unk-260901-01. Those codes never change and never get reused, so a clip you posted in March can still be matched to its numbers in December.

Over four hours? Split it. A session tops out at four hours, and two sessions rank their clips separately, which is what you want anyway.

2 · Send your video

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_UPLOAD)

Claude asks one question: is the recording in Google Drive, or on the device you are chatting from?

On this device

Claude gives you a link. Drop the file on it, or tap to choose — it works the same on a phone. The bar shows real bytes as they move, because it is reading the actual upload rather than guessing. The file goes straight from your browser to Fetch's storage; it never passes through the page, and the page holds no keys. The link expires after a few hours — ask for a new one, it is free and instant.

Already in Google Drive

Say so. If your Claude has your own Google Drive connector, it searches your Drive, shows you what it found, checks the file is shared, and hands Fetch the link. You do not go and copy anything.

You never connect your Drive to Fetch — your Drive talks to your Claude, which you already trust. Fetch only ever receives the one file you pick.

No listing of your files, no folder names, no search results. One link, one file.

Without a Drive connector, paste the share link into the chat. It has to be readable by someone who is not you: in Drive, right-click the file → Share → under General access choose Anyone with the link → leave it as Viewer → Copy link. You can set it back to Restricted the moment the fetch finishes.

On a phone, Drive is the faster path

This is worth knowing before you start a multi-gigabyte upload over cellular. If your recording is a replay on your phone — a TikTok LIVE replay, an Instagram Live, a screen recording — the fastest route to clips is almost never "upload it from the phone":

1. Save the replay to your phone. On TikTok: your profile → the LIVE replay → Save video.
2. Share it into Google Drive. The phone uploads it once, in the background, and the Drive app is happy to do that over Wi-Fi later.
3. Share the file — Anyone with the link → Viewer — and copy the link, or just tell Claude it is in your Drive.

Why it wins: the Drive app uploads in the background, survives the screen locking and the app being closed, and retries on its own when the signal drops. A browser upload does none of those, and 4 GB over cellular dying at 90% is the worst ten minutes this product can give you. After that, Fetch's own fetch is server-to-server and takes minutes wherever you are.

If the recording is still on YouTube

Fetch will not go and get it — not the video, not the captions, not the title. It records the link and tells you the two downloads to make in YouTube Studio:

1. Studio → Content, find the video, row menu (⋮) → Download. That is the recording.
2. Subtitles → your language → Download → .srt. That is the caption track.

Then drop both on the upload page. The caption file is worth the extra ten seconds: with it Fetch skips transcription entirely, your names are already spelled the way you spell them, and you save around forty minutes on a two-hour recording. If the download lands in your Drive, say so and Claude pulls it from there.

3 · Write the words

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_TRANSCRIPT)

If you did not send a caption file, Fetch writes one on its own server — roughly a third of the recording's length, so about forty minutes for two hours. Nothing is required from you while it runs. Close the chat if you like. Say "where are my clips?" in a new one later.

This is the step that uses credits — one a minute of the recording, taken once. Before it starts, Fetch tells you what it will cost and what you have left.

If you have not bought credits yet, the estimate is a rough window rather than a countdown — see lanes for why.

4 · Check the words 📝

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_CAPTION_CHECK)

Before anything is cut, Claude shows you the names, handles and community terms the transcript produced, and you correct anything mis-heard. This step exists because hook lines, titles and burned-in captions are all lifted from these words — a name fixed afterwards means cutting the clip again.

Corrections here are free and unlimited, and Fetch remembers them: a term you fix once seeds the transcription of your next session, so the same name is not wrong twice.

5 · Pick your clips

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_CUT_REVIEW)

The step that takes your attention. Claude reads the candidate windows Fetch found — not the whole transcript — and proposes clips, ranked strongest first, each with its hook line, its length, and where in the recording it came from.

Every arc Fetch found is shown, including the ones it is not proposing to render, so nothing is hidden from you. You add, drop and reorder freely; none of it costs anything.

Then the gate: for each clip you see the hook, the line it opens on, the line it lands on, any word that looks like a mis-heard name, and every filler word that will be struck from the burned-in captions. Edits at the gate are unlimited and free. On your first session the gate cannot be skipped — afterwards you can turn it off.

6 · Check the look

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_LOOK)

Before anything is made, you get a picture of each clip. Not a description of it — the real cover frame, with everything drawn on it that decides how the clip will look.

Where your words will sit. A dashed strip shows the exact band the captions will occupy. You choose: bottom, top or the middle. Fetch also looks at what is under that strip, and says so when the words would cross the busy part of the shot — the person talking, the thing being shown. Safe from an app’s buttons and safe from your own picture are two different questions, and this step asks both.

What each app covers. A solid line marks the area that survives on TikTok, Instagram and YouTube alike. Each app’s own edge is drawn faintly beside it, so when something is covered you can see which app is doing it. Anything actually hidden is marked, and only the part that is hidden.

The thumbnail. The picture is the first frame of the finished clip, so what you approve here is what people scrolling will see.

Change anything and Fetch redraws before you can say go. A picture you approved can never be out of date — if you move the words, the old picture is withdrawn rather than left sitting there looking authoritative.

7 · Polish your clips

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_RENDER)

Your clips are made — you pressed the button at step 6. Nothing is charged here either: the credits came off when the video was read, and making the clips is what they were for.

Each clip renders as a still cover frame, then the joined beats, then a short end card; roughly a minute per clip. A clip with a hard readiness failure is refused with its reason rather than made badly — see readiness.

This is the step where you listen. Fetch cuts on the words, and a word list marks each line a fraction late — so now and then a clip opens or closes a hair inside a word. It is the one fault you cannot see, only hear, and it is why this step exists. Watch one, listen to the start, and if a word is clipped, say so: Fetch extends that piece and re-makes it.

This is also where anything belonging to one clip gets changed — its thumbnail, its captions — while step 6 set the look for all of them at once. Five changes per clip are free.

8 · Get your clips

📹 Walkthrough clip for this step — link pending (see SETUP-NEEDED: FETCH_VIDEO_DELIVER)

Download links for every clip and its thumbnail. Each link works for 7 days — ask again and you get a fresh one. Ask for the posting schedule and Claude spaces them so your own clips do not compete with each other, in your timezone.

Ask for the post copy and it lands on the same page: for each clip, a title, a caption and the tags, with a button that copies each one. Fetch writes none of it by itself and will not invent a view count. Paste your link back in when a clip goes up, and the page remembers where it went.

Sound

Fetch never puts a platform’s sound into a file. Their terms do not allow it, and it would not count as using the sound anyway — a track baked into a file is invisible to them, so the clip never reaches the sound’s own page.

What Fetch does instead is tell you the second a sound should land, measured on your own footage. Then:

• A trending sound. You add it in the app when you post. Tell Fetch where its drop lands and it will move your payoff to meet it.
• Music you make or hold. Hand Fetch the file and it goes under your voice, quiet where you talk and up in the gaps.

Music you supply is licensed by whoever you got it from, to you. Fetch adds nothing of its own, so nothing it hands you carries a claim it did not already have.

Write the words on your own computer

You do not need this. Fetch writes the words for you on every plan, including free, and it costs you nothing. This page is for someone at a desktop who would rather not wait in line — it needs a command line, so it will not work on a phone.

Your machine already has the hardware. If you would rather spend twenty minutes of your own CPU than wait in the free lane — or you simply want your recording to stay on your disk — transcribe it yourself and hand Fetch the captions file. Fetch then skips the whole step.

Entirely optional. Nothing about Fetch requires this, and the free lane transcribes for you either way.

The simple way: an app you click

No terminal, no commands. Both of these are ordinary applications that turn a video into a subtitle file.

How long does it take? Roughly a quarter to a third of the recording's length on a recent laptop — so 30–45 minutes for a two-hour recording — and much less if you have a gaming GPU. It runs in the background; you do not have to watch it.

The technical way: one command

If you are comfortable in a terminal, the script below does the same job and picks up a graphics card automatically. It writes captions.srt next to your recording. Then drag both onto the upload page, exactly as above.

• It never contacts Fetch or AnswerDog, and never sends your recording anywhere.
• The only network access it makes is downloading the transcriber and its language model, from pypi.org and huggingface.co — and it tells you what it is about to install, how big it is, and where it comes from, and waits for you to say yes.
• Every step is commented in plain language, with links to look things up.

Verify it never phones home

Do not take our word for it. Either watch what it opens while it runs:

sudo lsof -i -P -n | grep -i python

…or simply cut it off: run it once with the network on, then turn Wi-Fi off and run it again on another file. It works, because after the first run everything it needs is already on your disk. You can also search the file itself — there is no code in it that opens a connection to anything.

Checksums

Check what you downloaded is what we published:

shasum -a 256 src/fetch_local/__main__.py

Why there is no Fetch app to install

We will not ship you an installer, a launcher, or a desktop program — not now and not later. An unsigned download that wants permission to run is the single most alarming thing a piece of software can ask of you, and asking for it would undo everything else on this page. The two apps above are other people's, chosen because they are established and open; the script is a text file you can read.

Credits and your Claude plan

A session is one recording, start to finish, however many clips come out of it. It costs one credit for every minute of the recording, taken once when Fetch reads it. Your first hour is free. After that, $3 buys six hours.

Fetch is deliberately light on your Claude usage: the transcript never enters the chat. A two-hour recording's cues would be more than 13,000 tokens; the candidate windows Claude reads instead are about 1,200. A whole session is designed to fit comfortably inside a Claude Pro window, which is the point of building it this way.

Lanes

The free hour is the whole product in the slow lane — not a trial, not a watermark, not fewer features. The only thing buying credits changes is your place in the queue.

What "slow lane" actually means: your job waits behind paid work and then runs identically. A free job that has waited 24 hours jumps the entire queue, ahead of paid work — without that rule a busy week would push free jobs back indefinitely and "slow" would quietly become "never".

Buying once puts you in the fast lane for good. A purchase is a fact, not a subscription that can lapse and demote work you are already waiting on.

That is also why the free lane is quoted as a time of day rather than a number of minutes: the wait is other people's jobs, whose end the server cannot see, so a precise countdown would be invented. Your lane is fixed when the job is queued, so upgrading does not reorder work already waiting — and a lapsed card never demotes a job you are already waiting on.

Widescreen gameplay, made upright

If you recorded in widescreen with your face in a corner, Fetch stacks it: your face on top, the game below, filling an upright frame. Your face pane is generous — big enough to read an expression, which is the half people actually watch.

We ask two things once, when you set up: which corner your face is in, and where the action sits — left, centre or right of the game window. Both are remembered.

Plainly: these are fixed, smart crops, plus your preview. Fetch does not follow the action around the frame yet — there is no auto-tracking. It crops to the third you told us about and holds it steady, and you see a picture of every clip's cover before you approve anything. A steady crop you checked beats a clever one you did not.

The eight formats

Every clip is cut to one of eight shapes, and the shape decides the beats. Ask Claude for the methodology any time — it is published, versioned, and the same one every session uses.

These eight are seed values: chosen from published short-form study plus editorial judgement, and not yet validated against a body of your own results. The methodology is the same for every account, and it says so rather than implying otherwise.

Readiness checks

Before a clip renders, the server runs deterministic checks against the methodology: is there a resolution, does the hook land in time, is the runtime inside the platform windows, is the text inside the safe zone, is there one idea rather than three.

A hard failure refuses to render and tells you which check and why — that is the system declining to make you something bad. A soft failure renders and is reported, and you decide whether it matters. Every modification Fetch makes to your words — a trim, a tidied caption, a rewritten hook — is disclosed in one line before it happens. Nothing is changed silently.

For developers

Nothing in this section is needed to use Fetch. Every step of a session works in the chat, in a browser, or on a phone — if you are ever told to open a terminal to finish something, that is a bug, and we would like to hear about it.

If you are driving Fetch programmatically, register_stream returns the signed upload URL, its method, and the exact headers the signature covers. A PUT that omits Content-Type, or sends a different one, is rejected with 403. The equivalent from a shell:

curl -X PUT -H 'Content-Type: video/mp4'   --upload-file /path/to/recording.mp4   '<the signed url from register_stream>'

The connector itself speaks MCP over Streamable HTTP at https://mcp.answerdog.ai/mcp, with OAuth 2.1 and dynamic client registration.

Principle zero

Fetch never touches your accounts. No platform logins, no posting, no scheduling, no reading a platform's API, no scraping — for video, captions, titles or view counts.

Everything Fetch knows about your channel, you told it. That is why the YouTube path is two downloads instead of a paste, and why your Drive connects to your Claude rather than to Fetch. It is enforced in the code, not just promised: the server has nowhere to put a platform credential, and the build fails if a table grows a column that could hold one.