Generate Text - Any LLM

The Generate Text - Any LLM node sends your text to an AI language model and returns the model's answer as text. Use it to write, rewrite, summarize, translate or describe. Choose a model, write instructions and a request, and wire the answer into any node that accepts text. The node was called Prompt from September to October 2026, and Generate Text before that. Workflows that contain the older AI Agent node (type ai-writer) open with this node in its place.

When to use it

• Write or rewrite text: captions, scripts, product descriptions, headlines, translations.
• Turn one idea into several image prompts, and generate one image per prompt.
• Describe or summarize a connected image, video or audio clip.
• Summarize or rewrite content from Web Scrape, Telegram Channel Feed or Transcribe.
• Return structured data, such as JSON, for Extract Field.

Quick start

Inputs

The Generate Text - Any LLM node has three inputs on its left edge.

When a field has typed text and its input is also connected, the typed text wins. To use both, write {} in the typed text where the text of the Prompt input should go.

The node has two outputs:

Settings

The toolbar under the node repeats the most used settings: the model, the reasoning effort, the preset and the number of runs.

Models

The Generate Text - Any LLM node can run the language models below. The AI Model picker in the settings panel always shows the current list.

• Gemini 3.7 Flash is retired. The AI Model picker no longer offers it. A node or API call that already names gemini-3.7-flash keeps running. Gemini 3.8 Flash usually answers it, and with Advanced mode on, Gemini 3.7 Flash itself answers. Choose Gemini 3.8 Flash for new work.
• Video and audio. Only Gemini 3 Flash, Gemini 3.6 Flash and Gemini 3.1 Pro read video and audio references. With any other model, a connected video or audio reference is ignored.
• Temperature. Only these models use Temperature: the Gemini models, Claude Haiku 4.5, Claude Sonnet 4.6, GPT-5.2 and GPT-5.4. Every other model ignores it.
• Reasoning Effort. Gemini 3 Flash, Gemini 3.1 Pro, Claude Haiku 4.5 and GPT-5.2 have no effort levels. The two Gemini models gain levels in Advanced mode. Gemini 3.6 Flash and Gemini 3.8 Flash offer Low and High, and more levels in Advanced mode. Gemini 4 Argon and Gemini 4 Carbon offer Low, Medium and High. GPT-5.4 and Kimi K3 stop at High, and GPT-6 Astra, Grok 4.6 and Grok 4.7 stop at Very high. A level above what a model offers runs at the model's highest level.
• Kimi K3.1 Pro offers Low, High and Max. Medium runs at Low, and Very high runs at High. On Auto, the model reasons at its own default, Max.
• None on Kimi and GPT-6.1 Terra. On Kimi K3 and Kimi K3.1 Pro, None runs at Low, their lowest level. GPT-6.1 Terra offers Low to Max and has no None level, so None runs at Auto.

Which model to choose

• Economy models are plenty for rewriting, captions, translations and image descriptions. Start with the default, Gemini 3.6 Flash.
• Standard and premium models cost more per token. They are worth it for long, complex writing and for reasoning over many rules.
• For a video or audio clip, choose Gemini 3 Flash, Gemini 3.6 Flash or Gemini 3.1 Pro.
• For a Claude model at its usual rate, leave Reasoning Effort on Auto. Any other level runs the call on Anthropic's own service, at that service's rate, see Credits.

Choose the cheapest model that meets your quality bar, and compare two models on the same request before a large batch.

Credits

A run is charged for the tokens it actually used, at the rate of the model that ran it. The tokens are the prompt (the instructions, the request and the references) and the answer. The full rules are in How LLM calls are charged.

• The Run button shows an estimate. Before a run, the Run button shows ≈ and a number. This is an estimate of a typical call on the selected model, not a fixed price.
• A hold comes first. When the run starts, Nodaro holds enough credits for the longest answer the model may write. If your balance is below the hold, the run is refused before it starts. After the call, the run is charged what it used, which is usually much less than the hold.
• Longer text costs more. A long prompt, many references or a long answer cost more on any model. A premium model costs more per token than an economy model.
• The reasoning level has no price of its own. A higher Reasoning Effort usually writes more reasoning tokens, and those tokens are charged at the model's rate. The Very high and Max labels say "may bill one tier up", but their calls are charged per token like every other level.
• GPT-6.1 Terra has one estimate. Its Run button shows the same estimate at Auto and at every Reasoning Effort.
• Some calls run on the maker's own service. This happens in Advanced mode, and when you choose any Reasoning Effort other than Auto on a Claude model, because Claude's reasoning levels take effect only there. It also happens when the default service fails and the call is tried again. The call is then charged at the rate of the maker's own service.
• Gemini 4 Argon always runs on Google's own service. It is charged at that service's rate on every run, at every level, with or without Advanced mode. On a self-hosted install it needs a GEMINI_API_KEY. Without the key, the run fails with a "provider unavailable" error.
• Number of runs multiplies the estimate. Each answer is charged on its own, and the Run button shows the total estimate.
• Failures. A failed call that the model's maker still billed is charged for that usage, and the error says how much. This includes a refusal and an answer that was cut off. A failure that was not billed is refunded.
• Stopping a run. If you stop a run while the answer streams in, the run is charged for the prompt and the text written before you stopped it.

Split the answer into items

The Generate Text - Any LLM node can turn one request into many downstream runs. Ask the model to separate the parts of its answer with a line that reads ===NEXT===, for example: "Write 4 image prompts for this product. Separate them with ===NEXT===." The Items output splits the answer at each marker, and a node connected to Items runs once per item.

After a run, the settings panel also lists the items under Generated Prompts, with two buttons for working on the canvas by hand:

• Create N Image Nodes adds one Generate Image node per item, already connected and filled with its prompt.
• Generate All N Images runs every node that the first button created.

Use these buttons while you explore. For a workflow that runs on its own, connect the Items output instead.

Presets

A preset sets the instructions and the output format.

The three writers separate their prompts with ===NEXT===, ready for the Items output. They need a reference image, such as an Upload Image or a Generate Image node, connected to References. Without one, the panel shows a warning and the node refuses to run.

To keep your own instructions, click Save as preset. When one of your presets is selected, Update saves your changes into it, and the delete button removes it. Your presets also appear in Settings. Read more in Presets.

Use variables in the prompt

You can put the value of another node in User Prompt or Instructions (System Prompt) by writing its label in curly braces. {Product} becomes the text of a node labeled Product, and {Mood} becomes whatever the Mood picker is set to.

• A default value. {tone || friendly} uses the connected node labeled tone when there is one, and the word "friendly" when there is not.
• The connected input. {} inserts the text that arrives on the Prompt input at that point.
• One run per item. When a list, such as the output of Split Text, arrives on the Prompt input, the node runs once per item. Each item replaces User Prompt. Put the task in Instructions (System Prompt), which applies to every item. Pre & post text also wraps every item.
• Highlighting. A variable is cyan when a matching node exists and amber when nothing provides it yet. Amber is only a warning: a variable with a default still works.
• A missing node stops the run. If a variable names a node that does not exist anywhere in the workflow, the run is refused before any credits are spent.

Read more in Prompt variables.

Ask for structured output

To use the answer as data, ask for a format in the instructions. For example, write "Answer only with a JSON object with the keys title, caption and hashtags". A low Temperature keeps the format steadier on the models that accept it.

On the node, the view button switches the answer between raw text and a rendered view. View as JSON shows valid JSON as a colored object, and View as Markdown renders any other answer. Connect the Text output to Extract Field to pull one field, such as caption, out of the JSON.

Read and reuse the result

The toolbar at the top of the result on the node has:

• The view button, to switch between raw text and the rendered view.
• Copy text and Download.
• Open log, the full log of every run, with the model and the preset of each.
• Show outputs, when the node has more than one result: a strip of numbered results to page through. The selected result is the one that the outputs pass on.

Each result names the model and the preset that produced it. Drag the magnifier at the bottom-left corner of the node to enlarge it for reading.

When the answer is cut off

If the model reaches Max Tokens before it finishes, the run fails with a "cut off" error. A partial answer is never saved or passed on. The run is charged for the text the model wrote before it was cut off, because the model's maker billed it. To fix it, ask for a shorter answer in the instructions, lower Reasoning Effort, or raise Max Tokens.

Some models always reason before they answer, and their reasoning counts toward Max Tokens. On these models, Nodaro raises a small Max Tokens value to a minimum, so that the answer keeps room:

Nodaro also raises a small Max Tokens value on any model set to Very high or Max. DeepSeek V4.1 Flash does not accept a length limit, so Max Tokens does not limit its answer.

When a run fails with "Content policy violation"

The model's safety or recitation filter withheld the answer. On a Gemini model that runs on Google's own service, this includes an answer stopped because it repeats existing text. Nothing is saved, and the run is not tried again on another model. The run is charged only for what the model's maker billed for the refused request, and nothing when it billed nothing. Rephrase the request, or ask for an original answer instead of reproduced text.

Tips

• Put the rules in the instructions. Format, tone, length and "never" rules belong in Instructions (System Prompt), and the request belongs in User Prompt. Good instructions give more consistent answers.
• Adjust Temperature to the task. Keep 0.7 for most text, lower it for facts and structured output, and raise it for brainstorming.
• Compare answers side by side. Set Number of runs to 2 or more, then page through the results with Show outputs.
• Write reusable prompts once. Save instructions you use often with Save as preset, or keep a house style in Pre & post text.