Use cases
AI for writing
By Ofrai Editorial (Use Case Research)
Writing work is a loop: draft, dislike a paragraph, rewrite, cut it in half. Ofrai keeps Claude Opus 5 and the GPT models on the same credit balance so that loop does not need a new tool every time you change your mind.
Chat and video use credits from your Ofrai balance.
Drafting
Give the model the brief, the reader, and one example of something you have already written. The example does more work than the brief. Without it the model writes for an imagined generic audience; with it, the model writes for yours. A draft built on your own past paragraph is closer to your voice than a prompt that lists adjectives.
Rewriting
Rewriting is where the model is strongest, because the constraint is already written. Name the audience change, the length change, and the register change separately, and ask for one of them at a time. Asking for all three in a single instruction gives you a rewrite that is mediocre on all three axes, because the model splits its attention across the differences you listed rather than resolving them one at a time.
Structure
Feed in raw notes and ask for an outline before any prose. This is the cheapest step in the whole process and it is where most of the editing time usually goes. An outline you can reject in a minute costs nothing; three drafted paragraphs you then unpick costs a quarter of an hour. Structure the argument before you decorate it.
Voice and tone
The model can match a voice it is shown and will drift from one it is not. Keep a file of three or four paragraphs you would sign your name to, and put them in the prompt whenever the piece matters. What you are really doing is teaching the model which sentence shapes you use, which transitions you tolerate, and which clichés you never use, because a list of banned words is a weaker instruction than an example of the register you want.
AI for writing: a week of one document, three ways
Pick one document that matters and give it a week, because that is long enough to find the workflow and short enough to finish. On the first day, write the brief and ask for an outline only; do not let it draft. On the second, take the outline you kept, expand it yourself for the paragraphs you know cold, and paste those paragraphs back in as a voice reference before asking for the rest.
On the third, do a rewrite pass with single-axis instructions, one at a time, and keep every version rather than overwriting, so you can see which change actually improved the piece. On the fourth, generate a second version of the whole thing from the outline with a different instruction for the opening, and compare the two openings side by side, because a strong first paragraph carries more weight than three paragraphs of careful editing.
On the fifth, do the last edit yourself and mark the sentences you would never have written, because that list becomes the voice file for next time. By Friday the document is done, and the asset you have gained is a repeatable order of operations rather than a single finished draft. A week on one document is what makes the second document take an afternoon instead of a day.
Prompt patterns that keep your voice
Four patterns do most of the work. The voice-carrying pattern: 'Here are three paragraphs I wrote. Match this register exactly, including sentence length and how much I hedge. Now draft the section on {topic}.' The constraint-carrying pattern: 'Rewrite this. Keep every fact, name and number identical. Change only the opening so it starts with the disagreement rather than the agreement.' The cut pattern: 'Cut this by a third without losing any fact. Show me what you removed so I can put it back if it mattered.'
The unstick pattern, for when you have written the same paragraph four times: 'List five different angles on this section, one sentence each, no full draft yet.' The last one is the most useful and the most skipped, because generating angles costs almost nothing and it breaks the loop that comes from trying harder on a paragraph that is not working. The cut pattern's request to show what it removed matters too: a cut you cannot audit is a cut you cannot trust.
Choosing between Opus, GPT-5.6 and GPT-6.1 for writing
Opus is the editor, not the drafter, for most people. It is what you want when the piece is nearly right and the problem is structural, because it will tell you that the second paragraph repeats the first and that the ending argues a smaller claim than the opening promised. GPT-5.6 is the speed, and it is genuinely fast: use it for the volume work, the three subject lines, the five variations of a paragraph, the outline of a piece you have not decided on yet.
GPT-6.1 is the one to point at work that has a fixed shape, because a product description, a release note and a support reply all have a structure the model can be held to. The working order is to use the fast model for the parts you will throw away and the deep one for the parts you will keep.
Most people over-use the strong model on drafts they were going to rewrite anyway, and then under-use it on the final hour of editing, which is where it actually changes the piece.
Where AI writing goes wrong
The failures are recognisable once you have seen them. The first is generic register: a piece that opens with a statement about how fast things move, or closes by saying the future belongs to those who prepare. That is the model writing for nobody, and it happens when the prompt asks for a good article instead of showing a sample. The second is invented substance: statistics, studies, customer names and quotes that are plausible and unsourced, which is the most dangerous failure because it reads as authoritative.
The third is over-polishing, where a plain sentence arrives dressed in three subordinate clauses and your meaning survives but your voice does not. The fourth is dropped constraints, where the rewrite that was told to stay under a certain length, avoid a phrase, and keep the factual claims came back having quietly broken one of them.
The defences are unglamorous and effective: show an example every time, check every number against the source, compare revisions instead of trusting the latest one, and put the constraints at the end of the prompt as well as the beginning.
AI for writing: credits per draft, not per tool
Writing tools usually sell you a seat, and a seat charges the same whether you wrote one paragraph or thirty. On Ofrai the charge follows the work, so a long piece and a short reply cost what they actually consume. This changes which model you reach for. Because a throwaway outline is a fraction of the cost of a final draft, you can afford to generate three outlines and keep the best, which is the cheapest quality improvement available in the whole process.
It also means the expensive model is worth spending on precisely when the words will be read, and not when they will be deleted. The balance is shared with the rest of the studio, so a writing-heavy week and an image-heavy week draw from the same pool rather than requiring separate subscriptions.
What to take from the model
Keep the three things the model is actually better at, and leave the rest to you. Keep the outline, because a second set of structural options is worth more than any amount of line editing. Keep the variations, because three ways to open a piece is a decision you can make in ten seconds and would otherwise spend twenty minutes on. Keep the cut list, because knowing which sentences were expendable tells you where the argument is thin.
What you should write yourself is the sentence that carries the point, the claim you are willing to defend, and the sentence that a colleague would remember. The model produces competent connective tissue very fast, and it is the connective tissue that is replaceable. The paragraph a reader quotes is the one you should have written.
Who should adopt this, and who should not
Adopt it if your work has a lot of first drafts in it. If most of your time goes into producing a version that then gets cut, the model pays for itself quickly, because the cut is exactly the part it is good at. You will feel it most in the places where you are stuck on structure rather than on language.
Stay out of it if your value is in the exact words, which is the case for a manifesto, a founder letter, or anything whose voice is the point; a model will produce a competent version and the competent version is the problem. Be careful with personal correspondence for the same reason, and with anything where being wrong is expensive. The honest test is whether you would be embarrassed if the model turned out to be better than you at your job.
If the answer is yes, use it to draft and rewrite, and keep the final pass.
A first hour with AI for writing
Do not start with a document you have to deliver tomorrow. Start with something you already published, because you know what good looks like and you can judge the output without a deadline distorting your judgement. Paste in the piece, ask for an outline in the shape you would have used, and then ask for three alternative openings. Read the openings against the original, not against what you wish the original had been, and notice which one you would have written yourself.
Then take a paragraph from the piece and ask for a single-axis rewrite, and check whether the fact survived. That last check is the one that decides whether you keep using this. In an hour you will have a sense of whether the model is a useful second pair of hands for your writing or an uncanny stranger, and that sense is worth more than any prompt you can copy from someone else.
Open studio
FAQ
- Will it make my writing sound like AI?
- It can, and the cause is almost always the same: you did not show it an example of your own writing. A prompt that asks for a good article gets the model's idea of a good article, which is the flattest thing it knows. A prompt that includes three paragraphs you wrote gets your sentence length, your hedging, your tolerance for repetition. Keep a small file of your best paragraphs and paste them in every time; that one habit removes most of the generic register.
- Can it keep my brand voice across a whole piece?
- Across one piece, easily, if you re-anchor it as you go. Across a series, treat the voice file as the source of truth: put your own paragraphs in front of the model at the start of each session and check the sections that came back against it. A model drifts toward its own house style when it has been writing for a long time in one conversation, so start a fresh one for each document rather than carrying context forward.
- How much does a draft cost?
- It scales with the model you use and the length of the output, and the live per-model pricing is shown in the studio. The cost-effective pattern is to use the fast model for outlines, variations and anything you plan to discard, and the stronger model for the final structural pass. Most of the credit spend in a writing session usually goes into drafts you never use.
- How do I start on Ofrai?
- Open Text Studio and paste in something you have already written, then ask for three alternative openings and an outline. Keep whichever opening you would have written and throw the rest away. That is the entire first session, and it is enough to tell you whether the model is useful for your voice or only for other people's.
- What is AI for writing on Ofrai?
- One workspace with Claude Opus 5, GPT-5.6 Sol, GPT-6.1 Sol and a writing tool, on a shared credit balance. You choose the model per task instead of per subscription, and the balance carries over into the image and video work you do in the same studio.