Use cases
AI for developers
By Ofrai Editorial (Use Case Research)
Engineering work is mostly tokens plus the occasional diagram. Ofrai puts DeepSeek next to Claude and GPT so you can spend Flash-class credits on loops and premium credits on the explanation you ship.
Chat and video use credits from your Ofrai balance.
Coding loop
DeepSeek V4 Flash is the default hammer. Step up to Pro, GPT, or Claude when the answer needs to be pasted into a design doc. Image Studio is there if you need an architecture doodle; it is optional.
Research and specs
Upload PDF/DOCX in Text Studio. Compare DeepSeek vs Claude when the document is long and you care about citation-style care versus raw throughput.
When to add chat-with-PDF
When the spec arrives as a PDF, switch into chat-with-PDF inside the same Chat studio and ask questions in the same thread as your coding iterations.
When to escalate from DeepSeek to Claude
Stay on DeepSeek Flash while you hammer tests. When you have a patch you would actually merge, escalate to Claude or GPT for the explanation you paste into a ticket. The thread is preserved.
AI for developers: a day in one ticket loop
Most development time is not typing, it is holding context. AI helps most when it removes the part of the ticket where you re-read the same files for the third time. A workable day looks like this: read the ticket, paste the relevant module into a Chat with PDF session when the spec is long, and ask for a plan before code. Then generate the small, boring part first: a migration, a test scaffold, a rename with call sites.
Save the interesting design decision for a human, and use the model for the parts where the specification is unambiguous. This ordering matters because code you did not write is code you have to review, so the cheaper the generation, the less review you pay for.
Prompts that compile the first time
Context beats cleverness. A prompt that returns code on the first try usually contains five things: the language and version, the framework and its major version, the exact file paths involved, the constraints you cannot violate, and the expected output shape. Example: 'TypeScript 5, Node 20, express 4. Paste src/routes/orders.ts and its test file. Produce a paginated GET handler matching the existing response envelope used in src/http/respond.ts. No new dependencies. Return a unified diff.'
The unified diff matters more than it looks: it makes review mechanical, and it makes it obvious when the model changed something you did not ask for. If the first answer is wrong, add a missing constraint rather than rewriting the whole prompt.
Choosing between DeepSeek, GPT-5 and Claude Sonnet
Use the models for different jobs. DeepSeek Flash is the cheap lane: file reads, small refactors, test scaffolding, anything where you will read every line anyway. DeepSeek Pro is the reasoning lane for a bounded problem such as tracing why a test fails across three files, or comparing two libraries against your constraints. GPT-5 is the strong generalist when the task needs breadth: unfamiliar framework, multi-file change, or a migration plan you will argue with.
Claude Sonnet is the lane for long context and careful revision, especially when you paste a large file and ask for a diff that touches only what you named. The practical rule is to spend the strong model where you would otherwise lose an hour, not where you would lose a minute.
Where AI-written code goes wrong
Four failures account for most bad output, and all four are fixable by changing the prompt rather than the model. Hallucinated APIs: the model calls a method that does not exist in your version. Fix by pasting the real signature and adding 'use only APIs that appear in the pasted files'. Silent behaviour change: it rewrites logic you did not mention. Fix by naming the files it may touch. Confident refactors that break edge cases: empty arrays, nulls, timezone edges.
Fix by asking for the test cases first, then the implementation. Plausible but incorrect handling of simultaneous tasks: ask it to state its assumptions explicitly at the end. And whatever you take, you still read the diff.
AI for developers: credits per ticket, not per seat
Credit spend tracks tickets, not people, which makes it easy to forecast. Budget a two-hour ticket as roughly one planning pass, two or three generation rounds, one review round and a short explanation written for the pull request. Multiply by the number of tickets in a sprint and compare against the cost of the same work done at full attention.
The habit that keeps this honest is to keep cheap models on cheap tickets and to reserve the strong models for the two or three tickets per sprint that actually carry risk. Log which model produced which diff, because after a quarter that record tells you where your real bottleneck is: generation, or review.
What to attach to a pull request
The review is the deliverable. Alongside the diff, attach three things: a one-paragraph summary of what changed and why, the tests you ran and their output, and any assumption the code makes about data or configuration. State explicitly which files you did not expect to change, because that is the first thing a reviewer checks. If a model wrote the code, say so in the description rather than hiding it; reviewers can then focus their attention on the parts they would have skipped anyway.
Keep the prompt in your library next to the finished branch so a future fix can be regenerated with the same context instead of rediscovered. This turns the assistant into a record of decisions rather than a novelty.
Who should adopt this, and who should stay in their IDE
It fits developers who work in an unfamiliar codebase weekly, who write tests they would rather not write, and who spend real time on migrations and boilerplate. It fits teams where documentation is written by the same people who write the code. It is a poor fit if your work is dominated by architecture decisions with heavy organizational context, because the model cannot see the politics of a system it has never read.
It is a poor fit in regulated or security-critical paths without a review culture, where a plausible wrong answer is worse than no answer. And if your team cannot afford to review carefully, more generated code is not more delivered code, it is more deferred reading.
A first hour with AI for developers
Try one ticket end to end before changing any habits. Pick a small, well-specified bug. Paste the failing test and the two most relevant source files into Text Studio. Ask for a plan first, read the plan, and only then ask for a unified diff. Apply it locally, run the test, and keep or revert based on the result rather than on how confident the answer sounded. Then paste the specification document into Chat with PDF and ask the three questions you would have asked a colleague.
One hour and a handful of credits tells you whether this loop shortens your tickets, and the reverted diff tells you exactly how much review you would have had to do.
Open studio
FAQ
- Is there an API?
- This site is Ofrai. Check docs or support if you need programmatic access; these pages describe the product studios.
- Why not only DeepSeek?
- Because writing and review still benefit from Claude or GPT. You can try them on the same Ofrai credits instead of buying another seat.
- How much does a generation cost?
- Each job lists Ofrai credits in the studio. Catalog credits are starting points; resolution and duration can change the final spend.
- How do I start on Ofrai?
- Open Image or Video Studio from this page, pick a recommended model, paste or adapt the starter prompt, and generate from your library.
- What is AI for developers on Ofrai?
- A token-first workspace: DeepSeek Flash for cheap loops, Pro or Claude or GPT for the explanation you paste into a ticket, chat-with-PDF for spec questions. Same Text Studio, same credits.