Should I vibe code
Summarizes sources and organizes notes for research-heavy writing
Its premium feature is still called GPT3 Summaries. That is the entire build-versus-buy analysis.
?
Their verdict, the Basic (Professional) price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-05.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Start with the pricing page, because it settles the build-versus-buy question faster than any risk analysis can. Genei still sells GPT-3 summaries and GPT-3 paraphrasing as its premium tier, at £29.99 a month, under a footer reading © genei 2021. The app loads. Nobody appears to be home. That inverts the usual argument on this site — the case for buying is normally that somebody else keeps the thing alive, and there is no evidence of that here. So build it, and be clear-eyed about what you are building. Summarisation is one API call and always was. The work is PDF extraction, which is where all your accuracy quietly goes, and reference generation, which is the part that manufactures claims with a straight face. A summary of a paper you did not open, carrying a citation you did not check, ends up in a literature review with your name on it. That is the only way this tool hurts anyone, and it is a discipline problem rather than a code one.
What actually breaks
not "if". the specific failures.
- PDF extraction, which is where the accuracy goes: two-column layouts interleave, tables flatten to one line, ligatures come out as mojibake, and scanned appendices come out as nothing at all
- Reference generation, which is a citation-shaped string produced by a model — plausible authors, plausible year, plausible DOI, and no guarantee that the paper exists
- The summary of a paper whose conclusion is hedged, where the qualification is the finding and the qualification is exactly what a summariser drops
- Confidentiality, when the manuscript you are peer-reviewing under embargo, or the paywalled PDF your library licensed, goes out to a third-party model endpoint you never thought about as a recipient
- The note graph, which is the actual retention mechanism and the part that takes weeks rather than an evening
- Cost, if you re-embed the whole library every time you change chunking — a metered embedding endpoint plus a corpus is a bill with a loop around it
- The vendor, in this specific case, which is why this entry exists at all: the buy option is a product that stopped moving three model generations ago and is still charging monthly
Is that you?
the verdict is a default, not a law
- It runs locally over your own papers and the output is a draft you will read against the source
- Every generated claim carries a page number you can click, and clicking it is a habit rather than a feature
- You are summarising to decide what to read properly, not to avoid reading
- Citations come from a metadata lookup — Crossref, OpenAlex, the DOI — and never from the model
- References are generated by the language model rather than resolved against a real bibliographic source
- Anything it produces goes into submitted work without you opening the paper it came from
- You are uploading manuscripts under peer review, unpublished data or licensed PDFs to a third-party endpoint without checking what you agreed to
- It is a shared service for a lab or a class, at which point somebody else is trusting a summary you cannot vouch for
If you build it anyway
the checklist, then the prompt that enforces it
- Resolve citations, never generate them. A DOI lookup against Crossref or OpenAlex is cheap, deterministic and cannot invent an author.
- Anchor every sentence in a summary to a page and offset, and make the page reachable in one click. An unanchored summary is a rumour.
- Run extraction as its own step with its own output you can inspect. Most bad summaries are bad extractions wearing a nice font.
- Keep the source PDFs untouched next to the derived data. The notes, embeddings and summaries are all regenerable; the library is not.
- Decide deliberately where inference happens. If any document is confidential, embargoed or licensed, run a local model or do not run it at all.
- Record which model and prompt produced each summary. Without that, you cannot tell a year-old note from a current one, and you will not remember.
- Never let the tool write the sentence that goes in the paper. It suggests, you type — a rule that costs nothing and removes the only real hazard here.
I am building a research summariser: I drop in PDFs, it summarises them,
extracts keywords and helps me write. The output will end up near work that
has my name on it, so build for verifiability rather than fluency.
1. Never generate a citation. Resolve every reference against Crossref,
OpenAlex or a DOI lookup, and if it cannot be resolved, say so instead of
writing one that looks right.
2. Extraction is its own step with inspectable output. Show me the extracted
text before any summary, because most bad summaries are bad extractions.
3. Anchor every sentence you produce to a page and character offset in the
source, and make that anchor clickable. Refuse to emit an unanchored
claim.
4. When a source hedges — "in this sample", "may indicate", "not
statistically significant" — carry the hedge into the summary. Dropping
qualifiers is the failure mode here.
5. Ask me once, before any upload code exists, whether any document is
embargoed, under peer review or licensed. If yes, run inference locally
or stop.
6. Keep the original PDFs untouched and treat summaries, embeddings and
keywords as derived data regenerable with one command.
7. Record the model, prompt and date on every generated artefact, so I can
tell a stale note from a current one a year later.
8. Cache embeddings by content hash and warn me before any operation that
would re-embed the whole library at cost.
9. Do not write prose for me to paste. The tool suggests and links; I type
the sentence that goes in the paper.
10. Out of scope unless I ask: multi-document synthesis, a chat interface
over the whole library, and anything that rewrites my text.
11. Finish by telling me what a maintained alternative costs, and note that
Zotero is free and does the library half properly.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
Honestly, not this one. The usual argument for buying — someone else maintains it, someone else absorbs the model churn — needs evidence that someone else is still there, and a premium tier advertised as GPT-3 in 2026 is evidence of the opposite. If you want to pay for something in this space, pay for a tool that is still shipping, or pay for nothing: Zotero is free, handles the library and citation half properly, and has a plugin ecosystem that covers the summarising half with whatever model you prefer. Spend the subscription on API credits instead and keep the PDFs on your own disk.
The only irreplaceable asset is the PDFs and the notes you wrote by hand, so keep both in a directory tree with readable filenames and let everything else be derived. Summaries, embeddings, keywords and the note graph should all regenerate from that tree with one command; if they cannot, you have built a system you are trapped in rather than a tool. Export notes as Markdown with the source file and page in the front matter, and export the library as BibTeX with real DOIs. Both formats will still open in a decade, which is more than can be said for the product this entry is about.
Mature open-source reference manager that already solves the library, metadata and citation half properly.
Retrieval and question answering over scientific papers with citations resolved back to the source passage.
Questions
Is Genei actually dead?
Not formally, and this entry does not claim it is. The marketing site is up, the pricing page renders, app.genei.io loads and the trial signup is live. What is missing is any sign of movement: the premium features are named after GPT-3, the copyright line says 2021, and the feature list has not tracked three years of model releases. That matters here because the buy side of this decision is a bet on someone maintaining the thing. Treat the price as current and the roadmap as absent.
How is this different from the Afforai entry, which is the same build?
Afforai's entry is about retrieval quality: whether the chunk your searcher returned actually supports the sentence the model wrote on top of it. This one is about the two ends of that pipeline — extraction going in and citation generation coming out — and about a build-versus-buy calculation where the incumbent has stopped moving. Read both if you are actually building one; they are the same evening's work with different ways of embarrassing you.
Why does uploading a PDF count as a risk at all?
Because of which PDFs. A research tool's library is not a folder of blog posts — it is manuscripts you are reviewing under embargo, preprints shared in confidence, licensed journal downloads and sometimes your own unpublished data. Every one of those goes to a third-party endpoint the moment you point a hosted model at it, under terms you agreed to without reading, and that is a confidentiality question with a professional dimension rather than a technical one.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
A citation is a claim, not a formatting flourish. Your retriever cannot tell you whether the page actually says it.
The vendor switched the lights off in 2025. Whatever you build is now the only version of this that exists.
Saving is the weekend. Finding the thing you saved six years ago is the product you will not finish.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice