
What it is
One text, read once and set out many ways
A finished piece of long work tends to live in a single shape. The chapter stays a chapter, the talk stays a recording, the article stays on the one page where it first appeared, and everything else that might be made from it — the post, the letter to a mailing list, the script, the lesson — waits for an afternoon that never quite arrives. ContentSpark is built around that waiting. It takes one substantial text and sets it out again, in as many forms as a person asks for, while the text itself is kept whole beside everything made from it.
The work happens on one screen, which the app titles Architect, with the words Content Transformation Agent beneath. A person brings in a source: a manuscript or a chapter of one, a transcript, an article, a blog post, a page of notes. An AI reads the opening of it and writes down what it finds there, its themes and its tone, the audience it seems to address and the keywords a search engine might know it by. From that source a person can then ask for any of fifteen formats, from a short post to a sequence of five emails to a workbook, and the AI drafts each one with its own search title, description and keywords and, if the person wishes, a picture made to go with it.
ContentSpark does not publish.
The same source can become a course. A person describes the learners in their own words and chooses how many modules the course should hold, and an AI writes an outline of modules and lessons with its learning outcomes, then writes every lesson in turn, quizzes and exercises among them. The course is kept, read lesson by lesson beside its outline, and copied out whole as a single document whenever it is wanted somewhere else.
What it is not matters as much. ContentSpark does not publish. A person can mark a piece Draft, Review, Approved or Published and can place it in a queue against a platform, but the screen says plainly that actual publishing requires an integration, and nothing in the app sends a word anywhere. It does not measure how a piece fares in the world; the counts along the top of the screen count what the app holds. And the AI reads a set opening portion of each source rather than the whole of a long book, which is why the sensible way to work with it, as the sections below describe, is a chapter or a talk at a time.

A day with it
Thursday, a director and one chapter
The director of a small institute that teaches workplace mediation has a manuscript on her desk that the institute's founder finished in the spring: eleven chapters on how a difficult conversation is held open long enough to change something. The institute wants the fourth chapter, the one on listening for what a person needs rather than what they demand, to reach the managers on its mailing list before the autumn intake, and it has been talking for a year about a short course. She opens ContentSpark at nine and finds the one screen: four counts across the top, three tabs for Sources, Generated and Courses, and a search field beside a button that says Add Content.
She has learned to bring the book in a chapter at a time, because the AI reads the opening of whatever it is given, and a chapter fits inside that reading where a whole book does not. She presses Add Content, types the chapter's title, chooses Book/Manuscript from the list of types and drops in the Word file. The upload area says Extracting content while the file is read, and then the chapter's text fills the box with its length counted beneath it, a little under seven thousand words. She presses Upload & Analyze Content.
A new card appears in the Sources tab marked Analyzing, and within a minute it reads ready. It carries the chapter's title, its type, its word count, four tags drawn from the keywords the AI chose, and a line of themes. She selects it. A band opens above the grid with the AI's reading laid out in four parts: the first three themes, a sentence about the audience, a few words on the tone, and the number of opportunities it identified. Two buttons sit at the right of the band, Create Course and Transform Content.
She begins with Transform Content. The panel asks for an audience level, and she chooses Intermediate, since the managers on the list have all done the institute's introductory day. The box for AI images is already ticked and she leaves it. Under Select Output Formats she clicks the header of the Email group, which selects both of its formats at once, then ticks the long blog and the LinkedIn post in the groups above it. The panel counts four formats selected. She presses Transform, with the four beside the word, and the panel becomes a bar and a line of text that reads Creating 2 of 4: email sequence.
When the run finishes the screen moves her to the Generated tab, where four new cards wait, three with pictures across their tops. She opens the long post. It is rendered with its headings, and she reads it the way she reads a junior colleague's draft, slowly, with the chapter open in another window. In the fourth section the AI has made the founder sound more certain than she is about a study she only mentions in passing. She presses Edit, rewrites the sentence in the founder's own terms, and presses Save. The SEO tab shows a meta title of fifty-eight characters counted against sixty and a description just inside its limit. The Image tab shows the picture with the prompt the AI wrote for it set out underneath.
On the Distribute tab she marks the post Review, because the founder reads everything before it goes out, and places it in the queue for WordPress. Beneath the button the tab notes that actual publishing requires an integration, which she knows; for her the queue is a record of what is meant to go where. She copies the email sequence, the five letters one after another, into the institute's mailing software herself, and leaves the LinkedIn post for the founder to send in her own name.
After lunch she returns to the chapter's card and presses Create Course. The title is already filled in with the chapter's name and the words Complete Course, and she shortens it. In the box for the audience she writes three sentences about the managers: their first year in the role, the meetings they dread, what they have asked the institute for. She chooses five modules and presses Generate Course. The progress line names each module as it is written, and the panel takes its time, because every lesson is written by its own call to the AI. She answers two emails while it works.
The course arrives in the Courses tab with its modules, its lessons and its estimated hours, two learning outcomes on the card and a note of how many more there are. She opens it, finds the outline on the left, opens the third module and reads its quiz lesson, whose questions come with their answers. She presses Copy All and pastes the whole course into a document for the founder to mark up by hand. Before she leaves she opens the round button in the corner of the screen and asks ARCHITECT, the app's agent, which of the remaining formats would suit this chapter for the institute's podcast, and reads its answer before closing the panel. The counts at the top now read one source, four pieces, none marked Published, and one distribution.

Bringing it in
The source, and the AI's first reading
Everything in ContentSpark descends from a source, so the first screen a person meets after Add Content is the one that makes a source. It is called Upload Source Content, and it asks for three things before it will go on: a title, a type and the text. The types are six, and they say what the app expects to be given. A source can be a book or manuscript, a blog post, an article, a transcript, a set of notes, or something else, and the choice is kept on the record and shown on its card so that a transcript of a talk is never mistaken later for a chapter of a book.
The text can be pasted straight into a wide box, or a file can be given and its text taken out of it. The file field reads PDF, DOC, DOCX and TXT. When a file is chosen it is uploaded and read, the field says Extracting content while that happens, and the text appears in the box where a person can look at it before anything else is done. If the document carries a title of its own and the title field is still empty, that title is filled in too. Beneath the box the words are counted as the text changes, and the count is kept with the source when it is saved.
Then the AI reads. Upload & Analyze Content saves the source first, marked as analysing, and sends the opening of it, its first eight thousand characters, to a language model with a plain request. The model is asked for three to five main themes, a profile of the audience the text seems written for, five to seven key takeaways, ten to fifteen keywords for search, the opportunities it sees for turning the text into other forms, the tone and voice of the writing, and what sets it apart. The whole of that reading is stored on the source. The first five keywords become the source's tags, and its status becomes ready.
Not all of the reading is shown. The band that opens when a source is selected carries four parts of it: the first three themes, the audience, the tone, and a count of the opportunities found. The card carries the themes and up to four tags, with a small figure for any beyond them. The key takeaways and the points of difference are kept in the record but appear on no screen, which is worth knowing, because the agent described further down can read them when it is asked.
The Sources tab is the library of all of this, newest first. Each card shows its status, its title, its type and its length, its tags and themes and the date it was made, and selecting one opens the band with Deselect, Create Course and Transform Content. When there is nothing yet, the tab says so in a sentence and offers Add Content again. The one search field in the header reaches every tab at once: it finds a source by its title or by any of its tags, a piece by its title or its format, and a course by its title, narrowing the cards as a person types.

The source stays whole. Every draft made from it carries the way back, so nothing descended from the text forgets where it began.

Setting it out
Fifteen formats, each with its fields
Transform Content is where one source becomes many pieces. The panel sets the fifteen formats out in five groups, and a click on a group's header selects all of its formats or clears them again. The first group is the blog, with a short post of three to five hundred words and a long one of one to two thousand. The second is social media, with five formats: a post for LinkedIn, a thread for X, a caption for Instagram, a post for Facebook and a script for TikTok.
The third group is email, with a single letter or a sequence of five. The fourth is long form, which holds a course module, a podcast script and a video script. The fifth holds documents: a presentation, a workbook and an assessment. Beneath the groups the panel counts what has been chosen and offers Clear all, and the button at the foot carries the count in its own label, so that a person knows before pressing it how many pieces are about to be written.
The prompt is the AI's, not the person's, and it is kept and shown beside the picture so that a person can see exactly what was asked for.
Each format is given to the AI with an instruction written for it, and the instructions are specific. The thread for X is asked for in five to seven posts, each under two hundred and eighty characters. The Instagram caption ends on ten to fifteen hashtags, and the TikTok script opens on a hook in its first three seconds and stays under a minute. The LinkedIn post is told to use emoji sparingly and to close on a question. The sequence of five emails is asked to build letter by letter, each with its own subject line. The video script carries cues for what should be seen and suggestions for cutaway footage. The assessment holds ten multiple-choice questions, five true-or-false and three short answers, with an answer key that explains itself.
One choice applies to the whole run: the audience level. Beginner asks the AI to define its terms and give more context and examples. Intermediate asks it to balance depth with practical application, and Advanced allows technical language and finer distinctions. General Audience, the default, asks for language a broad readership can follow. The level is stored on every piece made in that run, and the piece's header shows it afterwards. Each piece is written from the first twelve thousand characters of the source, which is the second reason a chapter, rather than a book, is the right size to bring in.
With every piece the same call to the AI returns its search fields: a title for the piece, a meta title of fifty to sixty characters, a meta description of one hundred and fifty to one hundred and sixty, and five to ten keywords. It also returns a prompt for a picture. When the box marked Generate AI images for content is ticked, as it is when the panel opens, a picture is made from that prompt in a professional, modern and clean style, shaped for social media or for a blog by the format it belongs to. The prompt is the AI's, not the person's, and it is kept and shown beside the picture so that a person can see exactly what was asked for.
The pieces are written one after another, and the panel shows which: Creating three of five, with the format's name and a bar that fills. A picture that cannot be made is let go, and the piece is kept without it. If a piece itself cannot be written, the run stops there and says so, and every piece already finished in that run is kept. Each is saved as a draft, linked to the source it came from, and the screen moves to the Generated tab to show them.

The course
A source read as modules and lessons
Create Course, on a selected source, opens a panel titled Transform to Course. Its title field is filled in with the source's name followed by the words Complete Course, ready to be changed. Beneath it is the one field the panel will not go on without: a description of the audience, in the person's own words. The placeholder asks who the course is for, their level of skill, their aims and what they want to come away with, and the AI is given whatever is written there. Then a person chooses the size of the course, three, five, seven or ten modules, and the panel lists what will come back before Generate Course is pressed.
The course is made in two movements. First the AI reads the opening fifteen thousand characters of the source and writes the outline: a description of the course, five to seven learning outcomes, and the chosen number of modules, each with a title, a description, and three to five lessons. Every lesson in the outline is given a title, a short description, an estimate of its minutes and a type, which is one of video, text, quiz or exercise. Then the second movement begins, and each lesson is written in turn by its own call to the AI, from the opening eight thousand characters of the source, with an introduction, its main sections and its key points. A quiz lesson comes with its questions and their answers. An exercise lesson comes with instructions for the practice.
This takes as long as it should. A course of ten modules can hold fifty lessons, and the panel says which module is being written as it goes. When the last lesson is done the minutes are added up into hours, rounded to a tenth, and the course is saved whole and marked complete. If a lesson cannot be written partway through, the panel says so and nothing of that course is saved, so there is never half a course in the library pretending to be a whole one.
The Courses tab keeps every course as a card: its status, its title and description, the number of modules and lessons, its hours, its first two learning outcomes and a note of how many more there are. Opening a course shows its outline on one side, the modules opening one at a time, each lesson marked with a small sign for its type and its length in minutes. Before a lesson is chosen the reading pane shows every learning outcome together, which is the clearest view of what the AI took the course to be for. Choosing a lesson shows its number, its module, its type and its minutes, and then its content, rendered with its headings. Copy All gathers the whole course into one Markdown document on the clipboard, description and outcomes and every module and lesson in order, ready to be pasted into whatever will teach it.
Two things are worth saying plainly about levels and assessment. The course takes its audience from the person's own description; the four levels from Beginner to Advanced belong to the pieces made in Transform Content, where an assessment of eighteen questions with an answer key is one of the fifteen formats. Within a course, the testing lives in the quiz lessons the outline calls for. And because the course is drawn from the opening of the source, a whole book brought in at once yields a course about its first chapters. Brought in a chapter at a time, it yields a course for each.

The desk
Editing, marking, queueing, asking
The Generated tab is the desk where every piece waits. Each card shows the piece's picture with its format laid over it, or, where there is no picture, the format and the status side by side. Below sit the title, the first hundred and fifty characters of the text, three of its keywords and the date it was made. Three quiet controls open the piece, copy its text, and open a menu whose last line is Delete, which asks for confirmation before anything is removed.
Opening a piece shows its title with its format, its status and its audience level, and four tabs. Content shows the text rendered from Markdown, with Copy and Edit. Edit turns the text into a field a person can change line by line, with Cancel and Save, and Save writes the new text to the piece's record, so that the AI's draft becomes the person's own version and the draft itself is not kept beside it. SEO shows the meta title counted against sixty characters, the description counted against one hundred and sixty, and the keywords set out one by one; where a field is empty it says Not set. Image shows the picture, the prompt it was made from, and Download Image, which opens the picture in a tab of its own.
The fourth tab, Distribute, keeps the status. Four buttons mark a piece Draft, Review, Approved or Published, and the mark is written to the record and carried on the card. Below them is Schedule Distribution, and it does exactly one thing. A person chooses a platform from eight: four social networks — LinkedIn, X, Instagram and Facebook — two places to publish writing, WordPress and Ghost, and two mailing services, ConvertKit and Mailchimp. Schedule then writes a record that this piece is meant for that platform, marked pending, with no date. The tab says beneath it that actual publishing requires platform integration setup, and that is the truth of it: nothing is sent, and Bluesky, Mastodon and Pinterest are not among the platforms at all. The queue is a list of intentions a person keeps, and every piece still leaves the app by Copy.
The four counts along the top of the screen are just as plain. They count the sources held, the pieces made, the pieces marked Published, and the distribution records written. They are honest counts of the app's own records and nothing more; no view, click or reply from any platform reaches them, and the fields that might one day hold such things are empty.
Beside all of this there is a conversation. A round button in the corner of the screen opens a panel for talking with ARCHITECT, the app's agent, an AI given instructions on turning long work into courses and pieces, on search fields and image prompts, and on writing for beginners or for experts. Its replies appear in the panel as the conversation moves, set with their headings and lists, and the panel can be widened to fill the screen or narrowed again. Each opening of the panel starts a new conversation. The agent is given permission to read and change the app's records, to create and update sources and queued distributions, and to create, update and delete pieces and courses, so a request made in conversation can change what is on the desk. Its instructions speak of tracking performance and suggesting times to post, but nothing in the app gives it numbers to do either with, and its answers on those are its own reasoning, not measurements.

Fifteen formats, four marks, eight names in a queue, and not one word sent anywhere until a person copies it out.

In the suite
Where it sits among the nineteen
ContentSpark is one of the nineteen applications of B-Corps, installed on the house's own Mac by its own method and running against its own cloud backend, on a desk that starts each application, opens it, lets it rest, and every two minutes restarts any that has stopped. It keeps its own sign-in, as each of them does today. A stock-take of its four records and its two functions sits in the suite's register of every application's records, functions and agents, and the house's reading room has read it with the rest.
One connection is in the code. When ContentSpark opens, it asks Luminous Core, the house's register of its applications and the people who use them, for the signed-in person's profile by email, and makes one if there is none; and after every piece and every course, it records there that something was made, by whom, and of which format. It reads the register's address and its key from the app's settings, and it sends no text, only the fact of the work.
The other connections are arrangements of use, and they are the reason ContentSpark sits in the publishing layer. A person who keeps Chapter to Blog open beside it has a long post for each chapter of a book there, and can bring any one of them in here as a source, to be set out as a thread, a letter and a lesson. A person who keeps The Luminous Book open has the house's voice written down in it; the AI in ContentSpark is told only to keep the voice of the source, so the written voice is what a person reads each draft against before marking it Approved.
SquareDash is the marketing workspace for websites, with its pages, articles and campaigns. A long post finished here, copied out, can be placed there as an article, and from SquareDash the house's bridge carries articles to its Ghost publication as drafts. The queue in ContentSpark names Ghost among its platforms, but the bridge belongs to SquareDash, and the copy button is what crosses between them. Resonance Canon reads each book as a course with others and a coach; the outline of modules and lessons built here is a plain skeleton a person might bring to that room. Audiobook AI prepares a book for listening, and the podcast script drafted here comes from the same text. GrowthWrite runs a studio's own blog, OmniWriter keeps a writer's desk and a book's page, and Luminous Bookflow carries a book through a publishing house in software; ContentSpark is the place where one part of that book is set outward into every other form at once.

Who it is for
The person with one good text
It is for the person who has written something substantial and knows it deserves more readers than one form will ever give it. An author with a finished chapter and a publisher's spring list. A teacher who has given the same talk for years and finally has a transcript of it. A coach whose notes from a workshop could be a short course. An editor at a small press whose list of books has never once reached the press's mailing list in a form a person would open. Each of them has the text already. What they lack is the afternoon, multiplied by fifteen.
What it replaces is that afternoon. Without it, the same chapter is rewritten by hand for each place it might go, in a separate window, with the search fields worked out on a scrap of paper, the picture found somewhere else, and the course outline started in a document that is never finished. Every version drifts from the source and from every other version, and none of them remembers where it came from. ContentSpark keeps them together. Every piece and every course carries the source it was made from, every piece carries its own title, description, keywords, picture and mark, and all of them sit on one screen where one search finds them.
It counts what it holds, not what the world did with it.
What makes it different is how plainly it says what it is doing. Its AI works in the open: the reading of the source is shown on the source's own band, not kept out of sight. The prompt behind each picture is set beside the picture. The search fields are counted against their limits so that a draft can be checked by eye. The opening portion each call reads is a fixed measure, which a person can work with once it is known. Nothing leaves without a person copying it, and a person's own edit replaces the AI's draft rather than sitting beside it, so the version kept is always the one somebody chose.
It does not claim to know what will be read. It counts what it holds, not what the world did with it. It drafts quickly and leaves the judgement where it belongs: with the person who knows the founder's voice, the managers' dread of a Monday meeting, and the sentence that went one step further than the study it cited.
How it is offered
Installed for an organisation, and kept
ContentSpark is offered to an organisation in three ways: on its own, as part of the publishing floor, or inside the whole of B-Corps, where every application on these pages is installed together, set up in the organisation's own voice and with its own records, and kept running for as long as the organisation keeps it. The house carries each application into its suite for good, so what is set up now is still there, and still mended, in the years after.
The fee reflects the work of installing and shaping it, and the size of the organisation it serves. It is set with each organisation, after a conversation about what the application will carry for it.
Arranged in conversationContentSpark · B-Corps

The care it takes
What it keeps, and what it leaves to a person
ContentSpark asks a person to sign in before it shows anything, and a person who is not registered for it sees a page that says so in plain words and names what to try next, from checking the account to asking for access. The source is kept whole, with its type, its length and the AI's reading of it, and nothing made from it overwrites it. Every piece and every course points back to the text it came from, so that the source is never lost behind its descendants.
Its manners are quiet ones. A picture that cannot be made does not cost the piece. A run that stops partway keeps what it finished, and a course that cannot be finished is not saved as if it were. A piece is removed only after a person confirms it. The search fields come with their counts. A status changes when a person presses it, or when a person asks the agent to change it, and a queue entry claims no more than it is. The one thing it reports to the house is that a piece or a course was made and by whom, without a word of the text.
And it is direct about the AI. Each reading, each piece, each picture and each lesson is the work of a language model or an image model, and this page has said so at every turn; on the screens, the box marked Generate AI images for content and the agent who answers in the corner make it plain where a model is at work. A person is left with the decisions an AI should not make: what is true to the source, what the founder would sign, which piece goes out, and when.
At the end of the day the chapter is where it was in the morning, whole, in its card at the top of the library. Around it now are a long post, a letter in five parts, a short note for LinkedIn and a course in five modules, each carrying its way back to the same seven thousand words.
B-Corps
