Fourteen sections, one for each working tool. Each section explains what the tool does, where to find it, what result the user gets, and which settings are available.
In the mini app (Mini App), the user uploads an image (PNG, JPEG or WebP up to 10 MB) or a short video (MP4, MOV or WebM) and creates an asset from it: an entity that stores the file itself together with all overlay settings, including position, size, transparency, outline, watermark removal protection, rotation and placement mode. The asset name becomes its label on the bot button, so the user decides what to call this tool: “Corner logo”, “Full-frame watermark”, “Client N”, “Urgent” - anything that fits the workflow.
One asset can contain several variants - for example, white and black versions of the same logo, or completely different images that use the same overlay settings. After processing, the bot sends the result and immediately offers another variant without re-uploading the source.
A logo and a watermark don't have to be applied one run at a time - a composition can be assembled in one pass from several assets on the same image or video: for example, a logo in the corner and a text caption at the bottom together. Layers are added, hidden and reordered right in the editor; the layer order decides what ends up on top and what stays underneath.
Each layer is configured independently: the corner logo has its own position and size, the frame-wide watermark has its own opacity and outline. The active layer is selected and adjusted on its own, without touching the rest. A composition holds up to five layers, one of which may be a video layer (for example, an animated logo over a static caption).
A composition doesn't have to be built with buttons. Once the files are sent, describe what you want in plain text — “logo 1, opacity 20 + logo 2 top left” — and the bot arranges the assets into layers itself. The result is adjusted the same way: change settings, add or remove a layer, reorder them. It works in batch processing too — one phrase applies to every file you sent.
If you need a second tool from the same image with different parameters - for example, a small corner logo and a dense translucent watermark across the entire frame - the asset can be duplicated with one button and configured separately. The Mini App will show two independent entities and, accordingly, two bot buttons.
An asset doesn’t have to come from somewhere else — you can create one right in the Mini App out of text or generate it. The “Create from text” button: type the wording, pick a typeface and weight, and add an icon beside it if you want. From there it behaves like any other asset — same placement modes, opacity, outline and its own button in the bot. A quick way to make a name-based watermark with your handle, or a caption linking to your social accounts, when there is nowhere and no time to draw a logo.
The “Generate with AI” button: describe the idea in words, add a reference if you like — and the mark arrives with a transparent background already, ready to overlay. One button saves it as an ordinary group asset, and it works on photos and videos right away. More on this in the AI tools section.
Icons come from a built-in set — around eighty services, from Telegram, VK and MAX to YouTube, TikTok, Boosty and Dzen, each in colour, black and white. If the one you need isn’t there, search covers the rest; you can also use your own icon from your existing assets.
Rotation and sway are not separate assets but settings of the same editor: the tilt angle is set with a slider and combines with any placement mode, including tiling, while the sway effect brings the logo to life on video without distracting from the frame itself.
A logo can be not only a picture but also a short clip — MP4, MOV or WebM, including with a transparent background (alpha channel). An animated channel emblem, a spinning mark, a live sticker — such an asset is overlaid on video just like a regular one: the same editor, the same positions, size and transparency. Additionally, you choose the clip behavior: loop the playback for the entire video or freeze the last frame after playing once — handy for logo intro animations.
For any mode, the user can adjust transparency, outline and noise. The outline color can match the logo, contrast with it or be chosen manually. Noise can be gray or colored, and can be placed over the outline or over the entire asset.
These parameters are not cosmetic. A noisy stroke and a dense repeated layer significantly reduce the chance that the watermark can be removed by public automatic watermark removal services. Random placement works differently: in video, the logo changes position at a specified interval, making it harder to cover with one mask or remove with the same action across the whole fragment.
There is no such thing as 100% protection, and Flowmark does not promise it. But outline, noise and random positioning make it noticeably harder for automatic services to remove the watermark cleanly.
After processing, the bot sends the file and a set of buttons for quick adjustments: change grid position, size and transparency, enable or disable outline, noise or tiling. The user decides which of these buttons should appear in the asset editor; a layout preview is visible directly in the Mini App.
Logo on video → · Animated logo → · Watermark on photos → · Protection from theft → · Mark for a channel →
By preset — “ceilings” are set: maximum fps, resolution and bitrate. Several profiles are available - from lightweight to close to original quality. If the file already fits into the “ceiling”, the video remains unchanged; only files that are actually heavier than needed are compressed.
To a specific size — the target file size is set from one to two hundred megabytes, and parameters are selected automatically. This is useful when you need to fit a specific platform limit or stay under an autoplay threshold - the same “light” 10-15 MB, above which a video in a messenger feed no longer starts automatically and loses views.
In both modes, the system tries to preserve as much quality as possible within the given constraints. If the size restrictions are loose, a fast hardware codec is used; if the file needs heavy compression, the system can switch to slower software encoding for the sake of quality. The user sets the goal - a profile or required file size - and technical parameters are selected automatically.
For videos longer than three minutes or heavier than forty megabytes, preview mode is enabled automatically: one frame is taken from the middle of the video, and logo or caption overlay settings are checked on it in a couple of seconds. The full render starts only after explicit confirmation, so there is no need to wait a minute just to realize that the logo should move a centimeter higher.
Removing an audio track, converting video to GIF, and converting GIF back into a video with a quiet audio track (so Telegram does not treat it as a silent animation) are separate buttons in the action menu.
A horizontal video needs to go out vertically, a photo — as a square for the feed. A plain “center crop” cuts off faces and meaning. Automatic cropping works differently: Flowmark looks at what is happening in the frame and positions the window so the main thing stays visible. The tool used to work mostly from faces; now it also notices other things that matter — an object, an animal, an unfolding event.
Before the run you pick a format and what matters most: “Automatic” — the bot decides for itself, “People and faces” — it stays with the people and ignores other eye-catching objects, “Name the subject” — you describe in words what has to stay: “a burning car and smoke”, “the man in the blue jacket at the top”. Hints about position also work — “on the left”, “top right”, “in the bottom left corner”. If it cannot find what you described, the bot says so and falls back to ordinary automatic cropping. This function is still in beta.
With photos, the edges of the frame are checked before the subject search: white and black bars left over from someone else's crop are detected and removed, so the window is not built around empty strips. Ordinary frame content is not cut away — a bright sky or a dark background does not count as a bar.
With video, the tool thinks like an editor: the clip is automatically split into scenes, and the crop window is chosen anew for every cut. Inside a scene the frame does not chase the subject: while the main thing sits safely in view the window stays put, and if the subject does leave the edge of the frame it moves smoothly rather than snapping. You can also ask to remove black bars from the video if they were there in the source material. Available formats: 1:1, 16:9, 9:16, 3:4 and 4:3.
Diagram: the solid frame is the resulting crop window, the dashed rectangles are what was found in the shot. When the format changes, the window repositions to hold on to the main thing; if everything found cannot physically fit, the frame is centred on their weighted center.
The same mechanism picks the square frame for a round video message — which is why people stay inside the round mask. And it is usually the next step after text-based editing: first assemble the clip from the parts you need, then move the result into a vertical format for the feed.
Crop video for Reels and Shorts → · Automatic cropping in video messages → · Text-based video editing → · How faces are found →
Send a clip, press “Video editing” and write what to do with it: “keep 0:12 to 1:05”, “remove the part from 2:30 to 2:48”, “drop the first thirty seconds”. There is no notation to memorise — 0:45, 1.20, “from second 12 to 30” and “to the end” are all read the same way.
“Keep” and “remove” are opposite instructions, and the difference is read from the wording rather than from a checkbox in settings. One message can list several ranges at once: “remove 0:30 to 0:45 and 2:10 to 2:30” — the bot cuts both and joins what is left.
Send the video sources as an album or via /batch — and describe the order in words: “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; internally each keeps its own, so the order never gets mixed up, even when the files have identical names.
You can skip timecodes entirely: “keep the part where he talks about pricing”, “remove everything after he says ‘that’s all for today’”. In that case the bot transcribes the speech first and finds the spot itself.
A request can carry no numbers at all — cut a half-hour recording down to two minutes, keep what matters, drop the filler. The bot works out what the video is about and offers several meaningful versions of the edit to choose from: shorter or more detailed, closer to the topic or to the speaker. After that it is business as usual: review the plan, adjust it in words, assemble.
To understand a request written in plain words, Flowmark turns to a neural network — and only for that. It answers one question: what to keep and what to remove. The video itself is cut and joined by ordinary software: it transcribes speech, marks pauses, finds places where the clip had already been cut before, counts the exact seconds and verifies the result. Detecting earlier cuts is now noticeably faster, so a long clip takes less time to analyse. If you named exact timecodes, the neural network takes no part at all: the plan is built straight from your numbers.
The bot may move a boundary by a fraction of a second — so a word is not cut in half and no quarter-second stub is left at a seam. It then shows both figures, yours and the resulting one, and explains why; and if you are confident in your timecodes, the “Strictly my timecodes” button will restore the original.
Before assembling, a revision arrives in the chat: which parts remain, where the cuts will be and why a boundary moved. You can accept it, or correct it by sending the bot a new message — “start a second earlier”, “drop the second part” — and what you have already agreed on is not lost in the process.
While reviewing the revision you can pick the cropping you need straight away and remove black bars: after the edit, automatic cropping steps in and keeps the right thing in shot. Automatic editing uses short audio transitions that remove clicks at the cuts, while audio and video stay exactly the same length.
It does not build montage “edits”: no picking a soundtrack, no cutting to the beat, no inventing a narrative. It adds no effects or video transitions, does not generate video or music and does not retouch the image. That is a deliberate choice: the tool trims and joins, it does not invent.
Automatic editing is currently in beta — an unusual request can still trip it up. If the result is not what you expected, 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.
Trim and merge video with plain words → · Crop video for Reels and Shorts →
When the user sends a voice message, audio file, podcast, interview or video with speech, the “Audio to Text” button appears in the action menu. Before starting, the user selects whether diarization is needed (speaker separation), whether timecodes are needed, and what form the result should take - line by line or coherent paragraphs.
The language of the recording can be set explicitly. By default the system detects it automatically, but for mixed-language recordings, strong accents or rare terminology, forcing the language noticeably improves accuracy — and prevents cases where a couple of English phrases “convince” the recognizer that the whole recording is in English.
For interviews with background music, podcasts with jingles and videos with a soundtrack there is a dedicated mode: before recognition, the recording is split into stems — voice separately, music separately — and only the vocal track goes into processing. The music stops “polluting” the recognition, and the text comes out noticeably cleaner. The same mechanism is used in speech censorship (see the next section).
If the file was sent by mistake or the result is no longer needed, the process is stopped with a button while it runs, right next to the live progress. The finished text comes back as a reply to that very message with the file, so in a chat with several tasks it is clear what the result belongs to.
The main transcription is performed locally, on the service server side. The recognized text is additionally checked for typical hallucinations - repeating phrases and fragments that can appear during long pauses - so they are not included in the final document.
The finished transcript is provided in one of three formats:
If, after receiving the result, you need a different format, convert TXT to DOCX, switch from line output to paragraphs, add or remove timecodes, there is no need to re-upload the source and wait for a new transcription. A formatting button sits right next to the finished result: the bot reformats the already recognized text using the new parameters in seconds, without spending credits on repeated recognition.
On top of the same transcription, Flowmark can build animated subtitles over a video: the bot recognizes the speech, aligns the words to timecodes and renders captions in a chosen style. The style can be changed for free — without repeated recognition — and an SRT subtitle file can be downloaded along with the video. The tool is enabled in the mini app (Mini App); think of it as an extra option next to the main transcription.
Transcription is a separate tool and can be disabled from the menu if it is not needed for the workflow. This also works at the group level, where the owner can hide the button for the entire team, and at the individual level for a specific user. The same applies to any other Flowmark tool: the menu does not have to show anything extra.
Audio to text → · Interview transcription → · Transcribing recordings with music →
Silences obscene words in video and audio files. Under the hood, the tool uses a transcript with word-by-word time alignment: the system knows the exact start and end of each word and mutes exactly the necessary segments - not “a second around” and not a whole phrase.
The dictionaries cover Russian and English - several thousand words along with common spellings, typos and transliterations. Before comparison, the text is normalized: case, punctuation, Cyrillic and Latin characters are reduced to a common form. Additionally, bigrams are checked - pairs of words, each of which individually is harmless, but together they form an obscene expression.
The base dictionary is not dogma. In the mini app (Mini App), under “Tools”, each group has its own censorship dictionary: you can exclude words that are acceptable for your team (so they stop triggering falsely) and add your own — slang, competitor brands, anything. There is also a live check: type a phrase and see exactly what the production matcher would mute in it, with your rules applied.
Along with the processed file, the bot sends a “Muted: N” button. Behind it is a list of every triggered word with context and timecodes. If something was muted in vain, the word can be excluded right there — and the file rebuilt for free: recognition does not run again, only the audio track changes. After the rebuild, the bot offers to add the excluded words to the group dictionary so they never trigger again.
You no longer pick a “for speech” or “for music” mode: the recording is always split into vocal and instrumental tracks first, only the vocals are censored, and then the tracks are mixed back together. The music stays intact — no “holes” in the arrangement where words were muted — and recognition stops stumbling over background noise.
Fast or slurred speech, background noise and a strong accent can be too much for a single recognizer: a word may stay open or come out distorted. That is why Russian recordings are also processed by a second recognizer tuned for Russian. It re-checks the audio and, when it finds a match the first one missed, moves it onto exact time boundaries. Censoring in difficult fragments became noticeably cleaner.
There are two options to choose from: a standard beep tone over the found words or silence at the required intervals. Each task uses one mode, selected before processing starts.
In addition to processed audio or video, Flowmark can generate FCP 7 XML - standard exchange format, which is supported by most professional editing software (Premiere Pro, DaVinci Resolve, Final Cut, Avid). The XML contains pre-arranged intervals with the selected muting method - beep tone or silence - and exact source timecodes.
Along with the XML, the archive includes a separate audio track with the beep tone or silence, pre-cut to the required timecodes. The original video or audio is not included in the archive: it is assumed to be available to the person who sent the file to the bot. The editor opens the XML in their editing software, links it to the source and immediately receives ready-made markup for further editing: re-voice, replace fragments, cover with a jingle or leave it as is.
One “⭕ Video note” button turns a regular clip into a Telegram video message. It is not just a center crop: the square frame is chosen by smart automatic cropping — by where the faces are in each scene — so people stay inside the round mask.
The effects run from quiet to loud: a colour frame and any label around the rim, a glow and a pulse along the edge — and further on fire, a snow globe, Matrix, the VHS effect, an old terminal. Most come with more than twenty colour combinations and three strength steps, so the same effect can be a barely visible accent or the main thing in the shot. Five stylings read the loudness of the track and move along with the speech: they settle down in the pauses and flare up on stressed words.
You can set the effects up in the mini app: the preview animates live and shows the effect on your own frame — what you see is what you get. A styling can also be added to an already recorded video note.
A video note is a square video whose corners Telegram hides behind a round mask. Flowmark designs that hidden area too: under the mask sits a blurred, darkened version of the frame, and in the corner — an animated paper plane with a label curving along the arc.
Both the plane and the label can be replaced with your own. The label is a short line of up to twenty-four characters: an author name, a channel title, a link. In place of the icon goes any image of yours or a short clip, and the size is adjustable too. It is uploaded once in the mini app and from there lives as an ordinary Flowmark asset — the same file also serves as a logo on video and as a watermark.
None of this shows in the Telegram feed, but when a video note is saved and forwarded on as an ordinary video — and that is a common scenario — the styling is visible in full, along with whose message it is.
Flowmark applies several face detection algorithms in sequence: first, the most accurate neural network detector; then a second pass with an expanded frame area; then models aimed at rotated and partially turned faces; and finally, a classic detector as a fallback. The goal is simple: a profile face, an angled face or a poorly lit face should be detected as reliably as a standard frontal portrait.
In most real working situations, detection is reliable: frontal portraits, profile faces, angled faces, poor light and faces in a crowd. In the hardest shots, where a face is blocked by another object or almost completely turned away from the camera, accuracy can be lower; for such cases, manual adjustment is available.
The bot sends a preview frame with the detected faces, each of them numbered. You can hide all faces at once, hide only selected faces, or exclude some from processing. For each person, the mode can be configured separately: pixelation, blur or a colored rectangle; full face or eye area only; effect strength; manual offset and resizing if the detector is slightly off.
For blurring, soft edge feathering is additionally available - without harsh rectangular border around the face.
Yes, Flowmark has neural networks. By the laws of the genre we are supposed to shout about this on the first screen of the website — instead, we calmly tell you about it in section nine. These are auxiliary tools in beta: useful, but the service is not built around them.
For single photos, the action menu has an “AI Editor” button. The user describes the change in words — “remove the extra object”, “make the background wintery”, “replace the sign” — and gets a new image. Edits run as a chain: under the latest result there is a panel with every image of the session — the original photo and the numbered variants. One of them is ticked, and the bot says it in words: “your next request will change: variant 2”. To go back to the original or continue from another variant, press the button under the message — no forwarding files and no replying to old messages.
An intent check runs before generation: a stray remark, a bare number or general feedback like “not this” no longer goes to the model and costs no credits. The bot asks you to be specific, or offers an “Another take on the same request” button when the prompt itself was fine. A short clarification (“make the jacket blue”) or a plain “yes” picks up the context of the previous request.
Up to five images at once can be sent as input — they go in as a regular batch and are used together: for example, a product photo, a background reference and a style sample. Every text request and every image passes pre-moderation before generation starts; requests to remove other people's watermarks and logos are blocked.
You don’t have to choose a model — that has been decided for you. Flowmark runs on GPT Image 2 and Nanobanana 2, picked for how closely they follow the instruction and for how they handle text, layout and typography — the things generators usually fall down on. That makes the editor safe for demanding graphics: fixing the wording on a layout, replacing a sign in the shot, putting together a cover or a card for a channel.
The second decision is just as simple: your request is not rewritten. Many services quietly pad the description with “magic” words for a flashier picture — here it goes to the model as it is. That makes it clearer what was actually sent, and easier to phrase the next correction.
No separate menu is needed for this: just tell the bot what you want to see — an ordinary sentence, in your own words. The bot works out that you mean a picture from scratch, shows how it understood the request and offers a “Generate image” button. Nothing runs silently until you confirm the action.
Under the finished image there are “Another version”, “Edit image” and “Create new” buttons: repeating the same prompt and generating from a new description both ask for confirmation before anything is charged, while “Edit image” moves the picture into the AI editor. The description is checked before the run and the result after it; requests to remove other people’s watermarks and logos are blocked, just as in the editor.
The reverse task: there is no logo yet, but branding is already needed. In the mini app, in the asset editor, there is a “Generate with AI” button: you can describe an idea in words and get a finished logo with a transparent background right away — one button saves it as a regular group asset, and it immediately works on photos and videos. Two modes: B/W (a white mark that reads on any dark frame) and color; a reference image can be attached. We honestly consider this tool a quick start rather than a replacement for a designer — but for a first watermark it is more than enough.
And one thing that holds for every AI function: the output is produced automatically, may contain errors and distortions, and is not an opinion or position of Flowmark — it is made at the user's request. Checking the result before publication, including against the rules of the country where it will be shown, is the user's job; this is set out in § 10 of the offer.
This is not just “text in the corner”, but a separate tool: nine-point grid positioning with custom X/Y padding, font family and style selection (regular, bold, italic, thin, etc.), text color and transparency, and a configurable background box with color, transparency, corner radius and line spacing.
You can add an icon to the caption — from the built-in set of social networks and services, or your own, from the group assets. A messenger logo next to an account address, a channel mark before a name: the icon is placed to the left or right of the text, sits above the background box and automatically scales to the line height. It is the same set as in assets built from text — around eighty services in colour, black and white, plus search when the one you need isn’t there.
The catalog contains more than twenty font families from Google Fonts: from strict Roboto, Manrope, Onest, IBM Plex Serif and Playfair Display up to characteristic Lobster, Pacifico, Caveat, Press Start 2P, Rubik Glitch and Climate Crisis. All fonts are licensed under the SIL Open Font License, so they are suitable for commercial use.
Padding from the frame edges is set separately for each side. This is useful, for example, for a geotag or caption that should “peek out” from the bottom or side edge, rather than hang as a separate rectangle in the air. When the caption moves to the opposite side of the frame, the margins are mirrored automatically, so the same geotag on the right edge does not have to be reconfigured from scratch.
If needed, the height of the background box can be fixed: the block stays identical for any text and does not “jump” when the caption changes to longer or shorter text. If fixed height is disabled, the background box adjusts to the content.
A group can have several named caption presets, with one marked as the default. Users can switch between presets with buttons directly in the bot. In the mini app (Mini App), each preset has a live preview: the caption is shown on an image before saving, the aspect ratio can be changed (16:9, 1:1, 9:16), a test background can be selected, and a separate button zooms into the exact area where the caption sits.
Flowmark does not only work with files one at a time. Several photos or clips can be sent at once and processed as a single task — with the same settings and one button for all of them.
Smart batch — the user sends files one after another, and the bot detects the sequence, groups the files together and processes them as one task.
Manual batch - the /batch command: files can be uploaded at any pace, a live “Received: N” counter is visible in the chat, at the end there is a “Confirm” button.
Album - a regular album of several photos in Telegram or MAX is automatically treated as a batch, without separate commands.
The batch is downloaded in parallel, rendered simultaneously and comes back as a single message with a collage navigator. You can open any item from the collage as a separate file and adjust its settings without changing the rest.
The batch size depends on the group plan: starter plans allow a small set, while higher plans allow dozens of files at a time. Exact numbers are listed in the plan table.
Each group has its own set of assets, caption presets, settings for tools and button layout in the bot. A personal group is created automatically on first login; you do not have to add anyone to it, and you can work with it alone.
The group has a shared working limit: the owner chooses the plan, and the participants spend available credits and use the available features within this group, without separate subscriptions for each person.
There are three roles in a team. The Owner creates and configures the group, manages the plan, invites and removes participants. Administrator helps with management: edits the brand kit and adds members, but does not have access to billing and cannot delete the group. A Participant uses ready-made tools without access to settings.
The owner generates an invite link in the “Team” section of the mini app (Mini App) or with the /invite command. The link leads to a landing page with “Open in Telegram” and “Open in MAX” buttons; regenerating the link automatically cancels the previous one.
In the “Team” section the owner and administrators get team usage: how many credits went out over a chosen period, how the spending is split across participants and tools, plus a chronological feed of operations. It is clear where the limit goes, without guessing whether it is time to change the plan.
Every participant has their own statistics too, in the “Profile” section: personal spending, the number of operations and credits per tool, including those from the shared team pool.
The group stores one brand kit, and the team chooses where to use it: in Telegram, in MAX or on both platforms at once. You can send an invite in one messenger and accept it in another. Switching between several of your groups is done with the /switch command or from the “Team” section in the Mini App.
The mini app (Mini App) is Flowmark's visual control panel inside the messenger. It is where users configure assets, variants, presets, bot buttons, captions with preview, video options, face hiding, team, profile and plan.
The chat remains for frequent, quick actions: send a file, press a button, get the result. The Mini App is for less frequent, more detailed tasks: create an asset, customize the caption font, rearrange buttons in the desired order, or share an invite.
The full group brand kit - assets, variants, presets, captions, button layout and tool settings - is packed into a .flowpack file. It works both as a backup and as a transfer tool: you can roll back to a previous version of the settings or move the assembled brand kit to another group.
Before importing, a preliminary check is performed: the bot parses the file and shows how many assets and presets will be loaded, whether there are any name conflicts, and what changes are worth knowing about in advance. During the actual import, a backup of the current state is created automatically; it can also be downloaded and saved.
Close the messenger, open it a day later - the bot remembers the last branding parameters, the active group and recent actions. Several users can process files independently of one another at the same time: heavy operations go into the background queue, the bot remains responsive, and long operations show live progress in the chat - downloading, encoding, batch processing.
When several photos, videos or audio files are processed at the same time, their settings, progress and results stay separate — in batch processing, in AI editing and in the editors alike. When several operations are open in the chat, free text is tied to an explicit reply to the right message; if it is unclear which file a command belongs to, the bot asks instead of guessing.
An ordinary message that doesn't look like a task is handled by the help autoresponder: it looks for an answer by meaning and keywords and points to the right tool — in the bot or in the Mini App.
Files up to 2 GB are supported: the standard 50 MB messenger limit is handled on the service side, while for the user, sending files looks the same as usual. Logos are scaled with premultiplied alpha, which eliminates the light halos around transparent PNG edges that often appear with standard resizing.
Flowmark does not permanently store submitted photos, videos or audio. The source file is needed only during processing and is deleted immediately after the result is sent. As a safety net, an automatic mechanism runs at the infrastructure level: it constantly tracks the age of all temporary files and forcibly destroys everything older than 24 hours — even if an operation was interrupted or abandoned halfway. Only group settings, user-uploaded brand kit assets and operation history are stored: the minimum needed for the bot to behave predictably the next time it starts.
Flowmark uses soft limits: if a video is longer or heavier than the plan allows, the bot warns you and offers to run the operation at an increased rate (×2 of the regular credit charge). The decision is always the user's — nothing is charged or started without confirmation.
The bot and Mini App interface is available in Russian and English. Language is switched in profile settings and stays synchronized across both platforms. Before any settings import, Flowmark automatically creates a backup of the current state; you can save it and return to the previous configuration at any time.
Everything described above works today and has been tested on a daily editorial flow: several thousand operations per month. There are no hidden “next version” features in this breakdown. If something is not mentioned here, it is not available in the bot yet.
The service is updated regularly: some tools are refined, and new ones appear in response to user requests. If the function you need is not on this list, if something behaves differently than expected, or if you simply have an idea, tell us. The bot has a feedback form: open it with the /report command or from the "Profile" section in the mini app. If email is more convenient, write to us: We read everything and answer directly.
