Structured input, structured output
Fields for the constraints, a prompt for the subject. The formatting is applied by the app rather than requested in a sentence and hoped for.
AI SaaS case study
Alpha Drafts
Two jobs in one app: write the document, then ask the document questions. Both metered on credits, both ending in a PDF.
- IndustryWriting and productivity SaaS
- UsersStudents, professionals and academics
- DeliveredSeptember 2024
- ServicesProduct Design, Development, AI
Overview
Alpha Drafts writes long-form work to a brief — subject, length, formatting style, whether it needs citations — and hands back an editable draft rather than a wall of chat. When the draft is right, it exports as a PDF.
The other half runs the same relationship backwards. Upload a PDF and ask it questions: what does this say about X, where is the figure, summarise section three. The document stays on screen beside the conversation, so an answer can be checked against the page it came from.
Everything is metered on credits against a daily allowance, with three plans behind it — which is the part that turned a demo into a product.
- AI SaaS
- GPT-4
- Content Generation
- PDF Chat
- Rich Text Editor
- PDF Export
- Credit Billing
Project video
Content that writes itself.
The drafting flow, the PDF conversation and the credit model behind both.
Key capabilities
A brief, not a prompt
Type of piece, length, formatting style and citation handling are fields — so the person writing does not have to know how to ask.
The draft lands in an editor
Headings, emphasis, lists and alignment, in place. The output is a document being worked on rather than a message to copy out.
Chat with a PDF you uploaded
Ask a document what it says, with the page open beside the answer so it can be checked rather than trusted.
Credits, visible at all times
A daily allowance in the header of every screen. Nobody discovers the limit by hitting it.
A chat box will write an essay, and then the work starts: it is the wrong length, the citation style is not the one required, the formatting has to be rebuilt somewhere else, and turning it into a file to submit is a manual job every single time. The gap between "a model can write this" and "I have a finished document" is where the effort actually goes, and it is a gap made of small, repetitive, unglamorous steps.
A chat transcript is not a document
Formatting is lost on the way out, and rebuilding it in a word processor undoes most of the time saved.
The requirements come as an afterthought
Length, style and citations are decided before writing, and a prompt box asks for them last, if at all.
Usage nobody can see is usage nobody can budget
Unmetered AI is either a bill that surprises the operator or a limit that surprises the user.
The generator asks the questions a writer would be asked anyway — what are you writing, how long, in whose formatting style, with or without in-text citations — and only then takes the subject. What comes back is not a message but a document in a rich-text editor, ready to edit and one button from a PDF. The PDF conversation is the same design running the other way, and the credit counter sits above both so the cost of a piece of work is visible while it is being done.
One button to a PDF
Both halves end in the same place: a downloadable document. It is the step that makes the tool finish a job rather than assist with one.
Three plans, one meter
A day pass and two monthly tiers, all denominated in the same credits — so the difference between plans is a number rather than a feature matrix.
The offer, plans included
Advanced prompting, grammar fixes and downloadable PDFs at the top; the plans at the bottom, priced in credits rather than in features. A day pass exists because the first thing most people want is to try it on one document.
The brief on the left, the draft on the right
Type of piece, length, formatting style, in-text citations, references — the decisions a writer makes before writing, as fields. The result arrives in an editor with its formatting intact, and the credit counter sits above it the whole time.
The document, and the conversation about it
The PDF stays open beside the chat rather than being swallowed by it. That is the whole design argument: an answer about a document is only useful if the document is there to check it against.
Formatting that survives the trip to a file
Solution: Model output is plain text with markup conventions in it; a PDF needs real structure. The draft is parsed into a rich-text document first, so headings are headings and lists are lists — which is what makes the export a document rather than a screenshot of a chat.
Metering something whose cost varies
Solution: A long essay and a short email cost the operator very different amounts, so a per-request charge would be either unprofitable or unfair. Credits are deducted against a daily allowance and shown on every screen, which prices the work rather than the click.
Keeping a PDF conversation grounded
Solution: Answering from a document means retrieving from that document rather than from what the model remembers about the subject. The page stays visible beside the answer for the same reason — the user is the last check, and they can only be that if they can see the source.
I'm very happy and impressed with the freelancer's level of expertise and coordination. Excellent delivery, and simply the best!
Alpha Drafts
Client, via Fiverr
Questions
Long-form work to a brief — essays and papers are what it was built around, and the same fields cover reports, emails and articles. The formatting style and citation handling are part of the request rather than something to fix afterwards.
Yes. It arrives in a rich-text editor with its structure intact, so the normal work of rewriting a paragraph or changing a heading happens in place before the export.
Answers questions about the document you uploaded, with the document open beside the conversation. It is for finding and summarising what a file says rather than for generating something new.
Every generation and every PDF conversation draws on a daily credit allowance set by the plan. The remaining balance is in the header of every screen, so the limit is never a surprise.
Got a model and no product around it?
Let's build the part people pay for.