
How generation works
How the service turns a topic or a found item into a ready post: text, image, model and key selection.
1Two generation paths
A post can be created in two ways:
- 🔗 “From sources” — the service takes a found item and rewrites it for your brand.
- 🧠 “From the model’s knowledge” — a post is created on the day’s topic without an external source (facts, date references).
Holiday greetings are generated separately.
2Human writing style (no «AI tells»)
ALL posts automatically get a single editorial quality standard — regardless of topic or project context:
- ✍️ a hook/lead-in up front (not raw statistics from the first line), one developed idea and a real ending;
- 🚫 no clichés or «AI tells» (officialese, «it's important to note», empty intensifiers, symmetric lists);
- 📏 if a post runs longer than the platform limit, the service trims it with a cheap second pass, keeping facts and links.
This is about HOW it's written (clarity and voice), without inventing facts. One-off style wishes are best turned into «Editorial rules» (see «Project info»).
3Following links for substance (digests)
Sometimes the source post is a “link roundup”: a couple of lines plus a list of links to articles, with no substance of its own. So the post isn't just a rephrasing of headlines, during rewrite the service opens the links from such a post (the first 2) and pulls in the gist of those articles as extra material — the model then writes from facts, not from link titles.
- 🔗 Triggers when the source post has 2+ external article links (social networks, images and files are skipped).
- ⏱️ Hard-capped in size and time so token spend and latency don't balloon.
- ⚙️ Enabled by default; it can be turned off on the server.
4Choosing text and image models
In the «Информация о проекте» → “Generation models” section you set up model choice:
- 📝 the text model and 🖼️ the image model are picked separately;
- 🧩 models are grouped by provider country;
- ⚙️ “Auto” picks the first available provider (by key availability).
5Models per task (savings without quality loss)
In the “Which model does what” block you can assign different models to tasks:
- 🧠 “Content posts” — main posts, use a strong model: quality over price.
- 🔍 “Analysis & utility” — trends, holiday check, auto-plan; use a cheap/fast one: there are many such calls, savings are noticeable and the simplification is invisible to the user.
- 🧩 “Helpers” — project improvement, a mid model fits.
“Default” = the project model; “Apply to all tasks” returns every task to it at once. The “✨ Economy” button applies a preset picking the cheapest model AMONG THOSE YOU HAVE A KEY FOR (with no suitable key the task stays on the project model — an unavailable provider is never set). Review and save. Each group has a hint with pros and cons.
6Economy mode
The “Economy mode” toggle shortens the context sent to the model (a shorter own text of the blocks and a smaller slice of context files) — this reduces input-token spend. Downside: texts may be a bit simpler. Turn it on when price matters more than maximum detail.
7Keys: your own or the service
In “My keys” mode your personal API keys from the profile are used (stored encrypted). Subscription service keys are planned. If no suitable key is set, the service uses another available provider or the local mode.
8Custom model (OpenAI-compatible)
You can set your own endpoint, model name and token right in the project settings — for text and/or images. Any server with an OpenAI-compatible API works. Pick “Custom model” in the model list and fill in the fields in the project card.
9Context the model sees
The prompt includes:
- 🧩 tone of voice, audience and forbidden topics;
- 📣 marketing context and extra instructions;
- 📝 the text of uploaded context files (
.md/.pdf); - 🗓️ the project’s language and region;
- for “from the model’s knowledge” posts — also the date, weekday and the day’s topic.
The more precise your project settings, the better the result.
10Images and format
By default the image is generated in the channel’s social-network format (e.g. 1:1 for Telegram, 9:16 for Shorts). In the post plan you can set the format manually via the “Image format” field: the wider the ratio, the SHORTER the image — if a post “takes up half the screen”, pick “Wide 16:9” and the image becomes noticeably more compact; “Portrait 4:5” and “Vertical 9:16” are taller. The image is cropped to exactly the chosen format (no empty bars), and the same format is used for auto-animation into video. Changed the format when posts are already queued? Nothing is lost: right before sending, the service checks the image against the plan's current format and, if they differ, brings it to the new one — regenerating the image for the new frame (if an image model is available), otherwise neatly re-cropping the existing one. So posts prepared in advance go out in the current format.
Next to it is “Image style”: by default “Brand” (the project's visual style). For scientific and explanatory posts choose “Accurate illustration of the subject” — then the image is drawn as an EXACT depiction of the actual subject (correct anatomy/structure/proportions) rather than an abstract metaphor, and the brand “mood” style is not applied to such images (it is exactly what often pushes science posts into “glowing abstraction”). Note: the model can't render labels or captions in the image — the bet is on the accuracy of the picture itself, without labels. Your brand visual style from the project settings is added to every image prompt so all images stay consistent.
11If there are no keys or the model is unavailable
The service doesn’t fail: text is written by a local deterministic rewriter, and a gradient placeholder with a caption is used instead of an image. This way demos and tests work entirely without keys — only the quality will be template-like.
12Model request log
Every real call to a model (with the actual request and response) is visible in the «Запросы к моделям» section — handy to check exactly what was sent and what the model returned.
All purposes are logged:
- 🔗 source rewrite and posts from model knowledge;
- 📝 refinement and project improvement;
- 🔍 trends and relevance analysis (embeddings during source sweeps and when adding a source);
- 🖼️ images, import, templates and plan autofill.
So every sweep that ran topic selection adds a “Relevance analysis” entry to the log.
13Exact model name (Opus / Sonnet, etc.)
In “Project info” you can set the exact model name for text and images (e.g. claude-opus-4-… instead of Sonnet, or gpt-4o). Take the name from the provider console; empty means the current default model is used. Handy when a provider has several tiers (Opus/Sonnet/Haiku) or a new model ships — the list never goes stale.
14Generation limits (cost control)
To control model spend, the project settings let you set draft limits:
- ⏱️ a maximum per 24 hours and per 7 days (
0= unlimited); - ⛔ when a limit is reached, generation (both manual and scheduled) pauses until the window passes;
- 📊 current usage is shown next to the fields.
It counts generated drafts (by generation events) — deleting a post does not refund the quota.
15Language of posts and images
The post language is set in “Plan”; the model is explicitly told to write the title and body strictly in that language, and the image prompt requires any lettering to be in the same language.
Didn't find an answer? Write to us — we'll help.
Contact support