Text
Nearly every workflow writes some text: the confirmation a customer receives, the caption under a picture, the listing of a product, the summary of a contract, the same notice in three languages, the JSON a layout or an app reads, the prompt that tells an image model what to draw. This section explains how to produce that text — and how to read text out of pictures, documents and videos — so that it is right every time, not only most of the time.
Three ideas run through every page.
Assemble what can be assembled. A confirmation is a fixed sentence with a name, a date and an amount in it; an address is three lines joined by commas. Four nodes — Text Template, Format as Text, Join Text, Replace Text — build that kind of text exactly, in milliseconds, for 0 credits, and write money and dates the way each language does. A model is needed only for what has to be written: a description, a summary, a translation, a judgement.
A model’s answer is a proposal. It is good writing, but it may run long, add a plausible detail, drift from a brief or answer in another form. The pages show how to make it reliable: ask for data instead of prose, check it against a contract, keep figures out of the text and fill them in afterwards, and recompute what the model read.
Every page has a lab. Each technique comes with a small workflow you can install and run — a hotel confirmation, a bookshop newsletter, a garden centre’s emails, an antiques catalogue, a supplier’s quote, a ferry notice in three languages — so every claim on these pages was measured on a run, mistakes of the models included.
Where to go
Section titled “Where to go”| Your task | Read |
|---|---|
| Build text from data — a message, an address, a price in the customer’s language, a cleaned or masked text — without a model, for free | Text without a model |
| Have a model write a text: prompt and system prompt, Max Tokens and reasoning, which model | Writing with a model |
| Get data, not prose, from a model: JSON answers, checking them, reading their fields | Structured answers |
| Read pictures: alt texts, attributes, a condition report, a category for routing, a shop listing | Reading pictures |
| Read a video or a document: a summary, chapters, the figures of a quote | Reading video and documents |
| Decide on a text without writing one: which team for a message, what kind of document, whether a draft is good enough — answers with a probability | Routing and grading text |
| Produce the prompt of an image, video or audio step from data | Prompts for other AI steps |
| Write in the reader’s language; translate without breaking names, markers and figures | Languages and translation |
| Know what a text step costs, and why the charge differs from the estimate | What text costs |
| Find out why a text step failed, was skipped or wrote something wrong | Troubleshooting |
| See it all at work on a real job, and install it | Examples |
Captions and subtitles — timed text — are in Video and audio; reading and deciding on JSON values in Data and decisions.
The nodes by family
Section titled “The nodes by family”| Family | Nodes | Cost |
|---|---|---|
| Assemble text | Text Template, Format as Text, Join Text, Replace Text | 0 |
| In and out | Text Input, Text Output, JSON Output | 0 |
| Write with a model | Text Generation | by tokens, usually hundredths of a credit |
| Read pictures | Image Analysis, Image Classifier, Product Description, Image to Prompt | by tokens, a few hundredths per picture |
| Read video and documents | Video Analysis, Document Analysis | by tokens: about 0.31 credits per minute of video, under 0.1 per page |
| Check and read the answer | Validate JSON Schema, Extract JSON Value, Calculate | 0 |
The exact inputs, outputs and settings of every node are in the catalogue (get_node_type over MCP, or
the node types API).