F04 — Visual summary of a book

A reading app sends Madoo the summary of a book; Madoo returns a page to look at and remember, as a PDF and as an image: the idea in one sentence, the thread of the argument in three steps, four to six key ideas, three things to try and the author’s own words, with an illustration drawn by the AI and the facts of the work printed exactly as they are.

A reference run: 47 seconds, 5.14 credits; the page is 989 pt tall (at most 1160) because it follows its content. PDF — clickable: Leggi il testo originale opens the work · the AI structure · illustration · layout report.

The variant with one image per idea is F04B — Illustrated summary.

The case

Book → summary → visual summary. The visual summary replaces neither the book nor the summary: it helps to see its structure and remember it. The case belongs to publishing and education, but the same scheme works for an article, an internal report or a lecture.

The example uses La brevità della vita (De brevitate vitae) by Seneca, a public-domain text. The summary is written for the example; the Latin quotation is taken word for word from The Latin Library, which is also the link printed on the page. Wikisource gives the same sentence with perdimus instead of perdidimus: exactly the kind of detail that must come from a source chosen by a person, not from a model.

Input	Who provides it	In the workflow
Book summary	the reading app (or the editors)	Book summary (text, with a default value)
Facts of the work	the app’s catalogue: title, author and date, quotation with translation and reference, link to the original	Facts of the work (one JSON object)

What it shows

• The facts of the work do not pass through the AI. Title, author, date, quotation (original, translation and reference) and link reach the template as a second data object, next to the one written by the AI. The render merges them in order: first data_0 (the AI), then data_1 (the facts). For a key present in both, the fact wins: even if the model wrote a title, the catalogue’s one would be printed. The instructions also forbid the model to quote the author. See Templates in workflows.
• The structure is a contract. The model returns JSON with the idea in one sentence, three steps, four to six key ideas (theme, title, explanation) and three actions; utility/json_schema_validate checks the fields, the number of items and the lengths, computed from the space on the page. Output that does not fit stops the workflow before composing. See Make the AI write for the page.
• The page is one flow. Each block takes the lines it needs, and the list of key ideas takes the height of the rows it prints: with four ideas Prova da domani moves up, with six it moves down. A variable list sits in the middle of the page with more content below it. See Flowing layouts.

The template

595 pt wide and up to 1160 pt tall: the page follows its content and ends 36 pt below the footer, so with four ideas it is shorter than with six (Pages that grow with their content). Designed for the screen; Playfair Display for titles and the quotation, Montserrat for labels, Inter for text; night blue #1f2a44, paper #f6f2ea, terracotta #c2542d, sand #ebe2d1, gold #e9c46a. The fixed labels of the page are in Italian.

Area	Content	How it is built
Header	Sintesi visiva, title, author and date; illustration	a horizontal Layout: on the left a vertical Layout of texts (title up to 2 lines, then shrunk), on the right the illustration as the background of a 160 × 160 Layout with rounded corners
The idea in one sentence	big_idea	a Layout with a night-blue background and rounded corners; text up to 3 lines
The thread of the argument	step_1, step_2, step_3	three equal boxes (each text drawn for three lines, at least the drawn height) joined by arrows
The key ideas	key_ideas: 4 to 6 { tag, title, text }	a repeat inside the page Layout, drawn for six ideas at most; each row is a rule and a horizontal Layout: theme and title (2 lines) on the left, explanation (3 lines) on the right (Lists)
Prova da domani	actions: 3 { text }	a repeat of rows with a check box
Quotation	quote_translation, quote_original (optional), quote_source	a sand Layout; the Latin shows only when present (a condition)
Footer	author and date; Leggi il testo originale →	the last block of the flow (drawn at the bottom of the page it would keep the page long); the link reads the field source_url; when it is missing, text and link disappear

Field	Type	Required	Comes from
big_idea, step_1, step_2, step_3	text	yes	the AI’s JSON, validated (data_0)
key_ideas, actions	lists	yes	the AI’s JSON, validated (data_0)
title, byline, quote_translation, quote_source	text	yes	Facts of the work (data_1)
quote_original, source_url	text	no	Facts of the work (data_1)
illustration	image	yes	the generated illustration (its own port)

Design choices worth copying:

• A theme instead of a number. Repeat rows do not know their position; instead of having the model write “01, 02…”, each idea has a theme of one or two words, which says more.
• The steps have the same height. Each text in the three boxes is drawn for three lines with the rule at least the drawn height: the boxes stay aligned whatever the length of the sentences.
• The schema limits come from the page. Ideas: title up to 44 characters (two lines), explanation up to 130 (three lines); the idea in one sentence up to 150 (three lines).
• The instructions ask for less than the schema accepts. About 10% less (38 characters for the steps, 44 in the schema): a model that overshoots by a few characters stays on the page instead of stopping the run. Without this margin a run stopped on the third step (step_3 [maxLength]: Value should be at most 44 characters) — the contract did its job, but it should happen rarely.
• Three sample sets — Seneca, cinque idee, Quattro idee, senza latino and Il più lungo possibile (every text at the schema’s limit, six ideas): the last proves that even the worst case fits the page — about 940 pt with four ideas, 1036 with five, 1129 in the longest case (five ideas, four without Latin, the longest case).

The workflow

Book summary ──► Build the visual summary (AI, JSON) ──► Summary fits the page (JSON Schema) ──data_0──► Compose the page ──► Visual summary — PDF
                                                             └─► Illustration prompt ──► Draw the illustration ──illustration──►  (Render     ──► Visual summary — image
Facts of the work ──data_1──►                                                                                                    Document   ──► Layout report
                                                                                                                                 Template)

Node	Type	Why
Book summary	input/text	The summary, as the app has it
Facts of the work	input/json_value	One object with the verified facts: an app fills it from its catalogue, without six separate inputs
Build the visual summary	ai/text_generation, response_format: json_object, temperature 0.4	Finds structure and ideas; the rules forbid invented facts, quotations and repetition between sections
Summary fits the page	utility/json_schema_validate, inline schema, mode fail	Fields, number of ideas and actions, lengths: the template must not hope that the text fits
Illustration prompt	utility/extract	Takes the illustration instruction out of the JSON as text
Draw the illustration	ai/text_to_image, 1_1	A square illustration in the page’s palette, without text
Compose the page	design/template_render, output both, max_repeat: 6	data_0 the AI structure, data_1 the facts (they win), the illustration on its port
Outputs	output/pdf, output/image ×2, output/json ×2	PDF and image of the page, the reusable illustration, the approved structure, the layout report

Reading a run: Approved summary is the structure the AI wrote; the Layout report shows status: passed, repeatItems: 8 (five ideas and three actions), two images loaded, no warnings. If the structure does not fit, the run stops on Summary fits the page with every violation, before the illustration is paid for.

Cost: about 5.14 credits per run — the structure about 0.14, the illustration 5; validation, extraction and composition are free.

Install it

The example installs into your workspace with an API key, through the public API only — see Installing an example:

node madoo-install-example.mjs f04 --api <your Madoo API URL> --key <key.json>

It has no files to upload: it publishes the template and the workflow. To run it, open the workflow and paste the object from run-inputs.json into Facts of the work, or add --run (about 5 credits).

Change it

• Another book: change the summary and the facts of the work (title, author, quotation, reference, link). For a modern work with no different original language, leave out quote_original.
• Another kind of text: an article, a report, a transcribed lecture — the instructions speak of “a book or an essay”, but the structure (idea, steps, ideas, actions) holds.
• Another language: the model writes in the language of the summary; the fixed labels of the page (Sintesi visiva, Le idee chiave…) are in Italian in the template, so translate them in a copy of the template.

Files

manifest · template · workflow · run inputs · sample sets preview