The skill behind this guide: Sebastian - Academic Writer AI Skill. Argument structure, disciplinary register, and feedback on drafts you wrote - $14.99, one payment, yours permanently.
View the Sebastian skill →Claude can restructure an argument you already made, tighten prose you already wrote, and attack your reasoning so you can find the holes before a marker does. It cannot be trusted with a single citation: it will produce a real-sounding author, a plausible journal, a volume, a page range and a well-formed DOI for a paper that was never written. And generating text you then submit as your own is misconduct under most institutional policies, whatever tool produced it. This guide covers the method for catching fabricated references, where the integrity line actually sits, and four prompts you can copy.
Why does Claude invent citations that look perfect?
Because a citation is one of the most patterned strings in written English. Surname, initial, year in brackets, title in sentence case, journal in italics, volume, issue, pages, doi:10. followed by a prefix and a slug. A model that is very good at continuing patterns will produce something with every one of those features whether or not the underlying paper exists. Nothing in the output looks different when it is invented - no hedge, no lower confidence, no missing field.
The failure has a shape worth knowing. Fabrication is most likely exactly where you most want a source: the claim that is specific, convenient, and sits at the hinge of your argument. If the literature genuinely covers something, the model has seen the real papers and is more likely to name one. If nobody has published the study your argument needs, it will supply one anyway. So the references you are least inclined to double-check - the ones that arrived precisely as ordered - are the ones most likely to be fiction.
Related traps: real author, real journal, wrong paper. Real paper, invented page numbers. Real paper, real finding, but the finding is about a different population or a different timeframe than the sentence you attached it to. A DOI that resolves is not proof either, because a well-formed DOI can point at an unrelated article.
What is the actual method for getting references right?
Reverse the order. Do not ask a model for sources and then check them; that has you verifying a list where most items are plausible and some are not, which is slow and easy to abandon at 2am. Instead:
- Ask the model to mark the claims, not supply the sources. Have it go through your draft and flag every sentence that needs a citation, with the type of source that would support it and the search terms to find it.
- Find the sources yourself - PubMed, your library discovery layer, Google Scholar, the journal site. Read enough of each to know what it actually claims.
- Then have the model check your draft against the source text you paste in, sentence by sentence, and tell you where your paraphrase drifts past what the source says.
Step 3 is the one people skip, and it catches the more common academic problem: not an invented reference, but a real reference stretched to cover more than it says. Our citation manager guide covers formatting and reference hygiene; the research assistant guide covers the searching half.
Prompt 1 - mark the claims, supply no citations
Prompt 2 - check your paraphrase against the real source
Which parts of the workflow are defensible, and which are not?
Institutional policies differ, and yours is the one that governs you - read it before you read anyone's general advice, including this. But the line most policies draw is between help with expressing your thinking and substitution for your thinking.
Generally defensible, and generally still disclosable:
- Reorganising an argument you built - section order, where evidence sits, where the thread breaks.
- Tightening prose you wrote, without new claims being introduced.
- Explaining a convention you do not know: what a discussion section is for, how a systematic review is structured, why your field hedges the way it does.
- Arguing against you on purpose, so you can find the weak link yourself.
- Language help if you are writing in a second language - the same category as a proofreader, which most institutions permit and many require you to acknowledge.
Generally not defensible:
- Generating paragraphs you submit as your own writing.
- Taking any reference from a model without resolving it against the source.
- Producing a literature review of papers you have not read.
- Having it interpret results you did not interpret, then presenting the interpretation as yours.
For journal submission specifically, the ICMJE recommendations are explicit: a chatbot cannot be listed as an author, because authorship requires responsibility for the accuracy and integrity of the work and a model cannot hold that. Authors are asked to disclose AI-assisted technologies at submission and describe how they were used, and remain responsible for verifying every AI-touched line. Many journals and funders now have their own version of the same rule. Assume disclosure is required and check, rather than assuming silence is fine.
One more thing, said plainly: “will a detector catch it” is the wrong question. AI-detection tools produce both false positives and false negatives, they disproportionately flag non-native English writing, and no serious integrity decision should rest on one. The question is whether the work is yours, and whether you have declared what you used.
Holds your field, level, citation style and thesis across a long project, so feedback stays inside your discipline instead of resetting to generic every chat. It marks claims that need sources rather than inventing them.
View Sebastian - Academic Writer AI Skill →How do you install the Sebastian skill?
The download is a ZIP with SKILL.md at the root of the archive - not inside a nested folder, which is the usual reason an upload fails. In the Claude desktop app, open Customize → Skills, upload the ZIP, and toggle it on. Claude then loads it whenever your request matches what it covers.
Skills are available on Free, Pro, Max, Team and Enterprise plans. The real requirement is code execution: on Free, Pro and Max, turn on Code execution and file creation under Settings → Capabilities. On Team it is on by default at the organisation level, and on Enterprise an owner enables it in organisation settings - which is why the Skills menu is sometimes greyed out on a work account. That is an admin switch, not a plan problem.
In ChatGPT or Gemini there is no upload step: open SKILL.md, copy the contents, and paste them into custom instructions. You lose automatic triggering and keep the method.
How do you get structure help without handing over the argument?
Give the model your thesis and your evidence, and forbid prose. If it cannot write sentences, it cannot write your essay - it can only tell you where the load-bearing parts of your own reasoning are missing.
Prompt 3 - argument map, no prose
Prompt 4 - register edit that cannot add claims
Ready-made versions of the argument-mapping prompt above: sections, evidence placement, and the counter-argument blocked out before you write a paragraph.
View AI Essay Outline Generator →Where does this fail?
- Anything it has not been given. Ask about a paywalled paper you did not paste in and it will answer from the title. The answer will sound like a reading.
- Agreeableness. Ask “is my argument sound?” and you will usually be told yes. Ask it to write the harshest reviewer report your paper could receive instead - same model, far more useful.
- Disciplinary voice. A history monograph and an experimental psychology paper hedge, cite and structure differently. Left generic, output lands somewhere between the two and reads as neither. This is what a loaded skill fixes and a fresh chat does not.
- Novelty. It cannot tell you whether your contribution is new. It has no reliable view of what your subfield published in the last few months, and it will not say so unprompted.
- Long documents. By chapter six it has lost what chapter two claimed. Work section by section and restate the thesis every time.
- Statistics and results. It will produce confident interpretations of numbers it cannot verify. Never let it read a result you have not read yourself.
Prompts for the drafting stage once your outline and evidence are lined up - thesis, evidence order, and a conclusion that follows from what came before. Your ideas, your sources, better scaffolding.
View AI Essay Writer →In summary:
Never take a reference from a model - have it mark the claims, find the sources yourself, then have it check your draft against the real text. Map an argument with the AI Essay Outline Generator ($9), draft with the AI Essay Writer ($9), or keep field, level and citation style loaded across a whole thesis with Sebastian - Academic Writer AI Skill ($14.99). More in the Learning & Research skills collection. All work in Claude, ChatGPT and any AI chat, with a 30-day money-back guarantee. Check your institution's AI policy before you use any of it.
Sebastian - Academic Writer AI Skill
One file that keeps your field, level, citation style and thesis in view across a long project - structuring your argument, tightening your prose, and marking what needs a source instead of inventing one. No subscription. Yours permanently.
KissMySkills is a marketplace of 853 AI skills, 158 prompt packs, 55 agents & free tools for Claude, ChatGPT & any AI chat.