Send a clip to the bot, press “Video editing” and write what to do with it: “remove the part from 2:30 to 2:48”, “drop the first thirty seconds”, “keep 0:12 to 1:05”. No editing software, no syntax to memorise. The neural network is only here to understand the request — the seconds and the cuts are computed by ordinary software, so the numbers you name stay yours.
Promo access opens every Flowmark tool for 30 days — editing, cropping, branding, transcription and more. Right in Telegram and MAX.
Editing bench · pick a command and watch the result
The clip was shot for this demo — three failed takes in a row. The pale bands on the timeline are where the person is speaking; the bot marks them itself so it never cuts mid-word. Inside the bot everything works the same way, except you type your own sentence instead of tapping a button.
Cutting the “testing, testing” at the start. Dropping the pause while someone searched for a word. Removing the bit where a stranger walked into frame. Joining two takes because the good ending was in the first one and the good opening in the second.
For those ten seconds you normally have to open an editor, wait for the import, hunt for the right spot on the track and then wrestle with export settings. Flowmark handles it inside the chat instead: you describe the result in words, the bot shows you what you are going to get, and sends back the finished file.
It also works when there is more than one source: send several clips as an album or via /batch and describe the assembly order in an ordinary sentence.
Sources can be referred to the way people actually talk: “the first video”, “from the second up to 0:47”, “the third one in full”. Internally each keeps its own number, so the order never gets mixed up — even when the files have identical names.
Two failed takes · make one good clip out of them
The numbers inside the highlighted parts show the order in the finished clip, not the order of the sources. The bot works the same way: it first sends you what the result will be, and only assembles the video once you agree.
The neural network in this tool does exactly one job: work out what you meant when you wrote a sentence in plain words. Everything else — listening to the recording, finding the boundaries, counting the seconds and assembling the video — is done by ordinary software, each piece at its own step. Here is the whole path from your message to the finished clip.
Plain text in the chat, the way you would write it to a colleague: “remove 2:30 to 2:48”.
Separate programs transcribe the speech down to the word, mark where someone is talking and where the pauses are, and find the places where the clip had already been cut before.
speech · pauses · existing cuts
The only step it takes part in. It answers the question “keep or remove, and what exactly” — and it can only refer to things that were actually found at the previous step: a specific word, phrase or scene. Anything invented fails the check and is thrown away.
if you named exact seconds, this step is skipped
It turns the understanding into precise boundaries, moves a cut away from the middle of a word where needed, and verifies: are all the named parts there, is the order intact, is there a quarter-second stub left at a seam.
A short breakdown arrives straight in the chat: which parts remain, where the cuts will be and why a boundary moved. You can accept it, or correct it with the next message: “start a second earlier”.
The parts are cut and joined, the audio is matched in length. If you asked for a vertical frame, automatic cropping steps in here too.
Name exact seconds and the video is cut precisely along them. The neural network simply isn’t allowed to rewrite them.
If a cut was nudged half a second so a word wouldn’t be chopped, the bot shows both figures — yours and the resulting one — and explains why.
“Drop the second part”, “start a bit earlier”. What you have already agreed on with the bot isn’t lost in the process.
0:45, 1.20, “from second 12 to 30”, “the first minute and a half”, “to the end”. The bot parses it and shows how it read your request.
“Keep 0:12 to 1:05” and “remove 0:12 to 1:05” are opposite instructions, and the difference is read from the wording — not from a checkbox in settings.
“Remove 0:30 to 0:42 and 2:10 to 2:25” — the bot takes both ranges into account and shows them in the chat before it starts assembling.
“Keep the part where he talks about pricing”, “remove everything after he says ‘that’s all for today’”. Here the bot transcribes the speech first and finds the spot itself.
“Third one first, then the first, and from the second take the start up to 0:47.” Source numbers follow the order you sent them in.
No numbers needed: the bot works out what the recording is about and offers several meaningful versions of the edit — shorter or more detailed. Then you review the plan and adjust it in words, as usual.
In the same message you can ask for a vertical frame and no black bars: the bot will edit the clip and immediately re-frame it, keeping the right thing in shot.
So there are no false expectations — here is what it does not attempt. This is a deliberate choice rather than a “not yet”: it trims and joins, it does not invent.
It does not build montage “edits” — no picking a soundtrack, no cutting to the beat, no inventing a narrative. It adds no effects or transitions: cuts are straight, without flashes or dissolves. It does not generate video or music and does not retouch the image — for pictures there is a separate AI editor.
The tool is already useful in real work, but an unusual request can still trip it up. That is stated inside the bot as well, not only here.
The bot offers the /report_auto_tools command — it attaches the technical details of the last operation to your report, so the fix arrives sooner. Writing to support works too.
The same bot can recrop a clip to a vertical format, put a logo on it, compress it for a platform limit or transcribe the speech — the whole media workflow in one place:
No. 0:45, 1.20, “from second 12 to 30”, “the first minute and a half” — the bot reads almost any notation and shows how it understood you.
No. The seconds you name are kept as given: the neural network is only there to understand the request, while the numbers, the order of the parts and the cuts are computed by ordinary software.
The bot may move a boundary by a fraction of a second — so a word isn’t cut in half, or a stub isn’t left at a seam. It then shows both figures, yours and the resulting one, and explains why. “Strictly my timecodes” restores the original.
Yes. Send them as an album or via /batch, press “Video editing” and describe the order in words: “third one first, then the first, and from the second take the start up to 0:47”.
You can point at the speech or at what happens: “keep the part where he talks about pricing”, “cut the beginning while he is silent”. The bot then reads the recording first — transcribing the speech and marking the scenes — and that is a separate paid step (1 credit per minute, minimum 1) announced before it starts.
No. The tool does not build “edits”, pick soundtracks, add effects or transitions, or generate video. It trims and joins carefully — no more and no less.
At every seam the audio fades to silence within hundredths of a second and comes back just as fast — a short audio transition that keeps the cut from clicking. Duration is not eaten away in the process: audio and video stay exactly the same length, even across a long sequence of cuts.
2 credits per minute of source video; re-rendering the same material costs half. If the bot has to find the right part from your words, that adds 1 credit per minute. The promo link opens 30 days of extended access.
Because unusual phrasings can still trip it up, and it is fairer to say so upfront. If the result isn’t right, the bot offers the /report_auto_tools command — it attaches the technical details of the last operation to your report, so the fix comes faster.
