InkPhD
User guide

How to write a scientific article with InkPhD

The complete workflow, from an empty project to a submission-ready manuscript with verified Vancouver citations. Follow the steps in order the first time; afterwards you can jump straight to the part you need.

Quick start in 30 minutes

The shortest path to a first draft grounded in real literature. Each step links to the detailed section below.

  1. 01

    Create the project

    Pick the document type, title and output language (3 min).

  2. 02

    Add 10–20 sources

    Upload PDFs or import from PubMed; wait for indexing (10 min).

  3. 03

    Generate the structure

    Let the AI propose sections, then edit the list (5 min).

  4. 04

    Write one section

    Use Fast mode for the first pass (5 min).

  5. 05

    Check the citations

    Re-verify and read the numbered reference list (3 min).

  6. 06

    Run Readiness

    See exactly what is missing before you go further (4 min).

Step 01

Create a project and choose the document type

Everything in InkPhD lives inside a project: its own evidence library, its own vector index, its own sections and exports. The document type changes the prompts, the default structure and the length targets, so choose it carefully.

  1. 01On the dashboard press "New project" to open the 3-step wizard.
  2. 02Step 1 — pick one of the 10 types: Doctoral dissertation, Research article, Systematic review, Meta-analysis, Clinical protocol, Grant proposal, Book/Monograph, Case report, Conference abstract, Editorial.
  3. 03Step 2 — enter the working title and a description of at least 30 characters (topic, population, design, what you want to show). This description feeds every later AI call, so be specific.
  4. 04Step 3 — choose the writing language (12 languages supported). The interface stays in English; only the generated text follows your choice.
  5. 05Dissertation and Book use a section-by-section workflow; all other types also offer a one-shot whole-article generator.

TipA vague description ("about pancreatitis") produces vague text. Write the description as if you were briefing a co-author who knows the field but not your study.

Step 02

Choose the reporting guideline (EQUATOR)

InkPhD ships 11 EQUATOR guidelines plus a composite Dissertation checklist — 426 checklist items in total. The primary guideline drives the compliance engine and the auto-fix suggestions.

  1. 01In step 3 of the wizard the AI proposes the top-3 guidelines with a fit score and a short rationale — pick the one that matches your design.
  2. 02Later you can change it in the "Guidelines" tab of the workspace.
  3. 03Typical mapping: RCT → CONSORT · observational → STROBE · systematic review → PRISMA · prediction model → TRIPOD+AI · diagnostic accuracy → STARD · case report → CARE · protocol → SPIRIT · economics → CHEERS · animal study → ARRIVE.
  4. 04For a PhD thesis use the composite "Dissertation" guideline together with the guideline of each sub-study.

TipChoosing the guideline early is what makes the later compliance score meaningful — it is not a formality.

Step 03

Build the evidence library

InkPhD never writes from memory. Every factual sentence is retrieved from documents you added, so the library is the single biggest driver of quality.

  1. 01"Library" tab → upload PDFs (full texts, guidelines, your own protocols). Files are stored in object storage and indexed in the background; the chunk counter tells you when a document is ready.
  2. 02"PubMed" tab → advanced search with MeSH terms, study-type and date filters, free-full-text-only and sorting. Import the hits you want with one click; InkPhD tries to fetch the open-access full text (Unpaywall / OpenAlex).
  3. 03Zotero: in "Settings" add a read-only Zotero API key + library ID to sync an existing reference collection.
  4. 04Mark methodological favourites as "style reference" — their voice is used as an example, never as evidence.
  5. 05Aim for 15–40 well-chosen sources per section. Ten full texts beat a hundred abstracts.

TipUpload the full text, not the abstract. Retrieval works on paragraphs — an abstract gives the engine almost nothing to quote.

Step 04

Research question, title, abstract, keywords

The "Research Question" tab holds the scientific spine of the manuscript. Every field has an inline "Generate" button that drafts a proposal grounded in your library and description.

  1. 01Write (or generate, then edit) the research question in PICO form: population, intervention/exposure, comparator, outcome.
  2. 02Add the hypothesis, the objectives and the specialty — these sharpen every later generation.
  3. 03Generate title candidates and the abstract only after at least one results-bearing section exists, otherwise the abstract has nothing to summarise.
  4. 04Keywords: generate, then trim to the MeSH terms your target journal indexes.

TipTreat the AI proposal as a first draft by a diligent junior colleague: keep the structure, replace the claims you cannot defend.

Step 05

Build the structure (AI Structure Builder or reference PDF)

Instead of inventing a table of contents, let InkPhD propose one — or import the skeleton of a published thesis you admire.

  1. 01In the "Writing" tab open the Structure Builder. Tab "AI Generate" proposes 6–12 sections with a description, rationale, subtopics and 4–6 PubMed-ready MeSH keywords per section (the keyword chips open PubMed directly).
  2. 02Tab "Import from Reference PDF" parses a published dissertation (≤30 MB) and rebuilds its hierarchy up to three levels — front matter such as Abstract/Acknowledgements/References is skipped.
  3. 03On the review screen reorder, rename, indent/outdent, delete or insert sections, then press "Apply".
  4. 04Later you can drag sections in the left pane; dropping onto the middle of a card nests it as a sub-section.

TipA good structure is the cheapest quality gain in the whole workflow — fix the outline before you spend generations on prose.

Step 06

Scope sources to each section

Retrieval quality collapses when a 300-source library is searched for every section. Attach the right subset to each section instead.

  1. 01Each section card shows a "Library · N" badge — click it to open the two-column dialog (Attached vs project pool).
  2. 02Attach the sources that belong to that section; a collapsible list shows documents already attached elsewhere so you can reuse them.
  3. 03With sources attached, generation is strictly limited to them. With none attached, InkPhD falls back to the unassigned project pool (you can disable that fallback).

TipMethods sections work best with 5–10 very specific sources; discussion sections tolerate a wider set.

Step 07

Generate a section

The generator writes an argued, citation-carrying draft — not filler. You control length, retrieval mode and speed.

  1. 01Press "Write" on a section card (or "Generate section") and review the prompt: section title, your instruction, output language.
  2. 02Length slider: 1 000–8 000 words. The prompt enforces the floor, so 3 000 means 3 000 — expect 4–6 minutes for a long section.
  3. 03Retrieval mode: Dense (semantic), Hybrid (semantic + BM25 keyword, recommended) or BM25 (exact terminology). Optional LLM reranking for high-stakes passages.
  4. 04Fast mode (amber switch) halves the chunks and the word floor and uses a faster model — 30–90 seconds for a rough first pass.
  5. 05For non-dissertation types the Workspace dashboard also offers "Generate article" to produce title, abstract, all sections and references in one run.
  6. 06Generation runs in the background — you can keep reading or editing other sections while it finishes.

TipUse Fast mode to test whether the structure and the sources work, then regenerate the section properly.

Step 08

Edit: ghost text, Smart Continue, selection toolbar

The centre pane is a normal editor with three AI affordances that never overwrite your text without a preview.

  1. 01Toggle "Live edit" on a section: after a ~1 second pause at the end of the text a grey suggestion appears — Tab accepts it, Esc dismisses it. Auto-save keeps the draft safe.
  2. 02"Continue here…" (toolbar) or the hover "+ add here" button on any paragraph opens Smart Continue: give a one-line instruction ("add two sentences contrasting Bakker 2012 with van Santvoort 2010"), choose 1–5 sentences, then Generate → preview → Accept or Reject.
  3. 03Select any text to get the floating toolbar: Expand · Shorten · Rewrite · Academic tone · Add citation · Translate. Citations in the selection are preserved.
  4. 04Version history is kept for every section, so you can always compare or roll back.

TipSmart Continue with an explicit instruction is the single most useful tool for turning a generic paragraph into your own argument.

Step 09

Citations: numbered Vancouver style and verification

InkPhD may only cite sources that are actually in the retrieved evidence. Everything else is flagged instead of silently published.

  1. 01Internal citation tokens are rendered as sequential [1], [2], with [1, 2] collapsing for adjacent references and reuse of the same number for repeats.
  2. 02Each section ends with a numbered reference list: full title, authors, journal, year and a DOI / PubMed link.
  3. 03The Citation Verifier checks every emitted DOI/PMID against the retrieved evidence and reports a trust percentage; unresolved references are marked "needs review", never hidden.
  4. 04After adding new sources press "Re-verify" on the section so old citations are checked against the updated library.
  5. 05In the "Exports" tab you can reformat the whole reference list into APA 7, Vancouver, AMA 11 or BibTeX.

TipNever submit a manuscript with "needs review" references. Either attach the real full text to the library or delete the claim.

Step 10

Copilot and the Evidence panel

The right-hand Intelligence Panel has five tabs — Copilot, Evidence, Citations, Reviewer, Compliance — and always works on the section you have open.

  1. 01Copilot: a multi-turn scientific chat over your own corpus. Ask "which of my sources contradict the primary outcome?" or "is my sample size defensible?" — answers carry [1], [2] citations into your library.
  2. 02Evidence: search your library directly, switch retrieval mode, seed the query from the open section, and read ranked excerpts with similarity scores before you write.
  3. 03Citations / Reviewer / Compliance: verification summary, inline AI peer review with per-issue auto-fix, and the EQUATOR checklist with missing-item generation.
  4. 04The three panes are resizable and the Intelligence panel can be collapsed; on a phone the workspace switches to a Chapters / Editor / AI tab switcher.

TipUse the Evidence tab before generating a difficult section: if the engine cannot find good chunks, neither will the writer.

Step 11

Quality checks: statistics, gaps, novelty, originality, grounding

These are the checks a reviewer will run on you. Run them first, on your own draft.

  1. 01Statistical Reporting Checker (Analyze → Stats): audits 10 items — sample size, 95% CI vs p-only, multiple-testing correction, missing data, subgroups, adverse events, baseline balance, primary vs secondary outcomes, exact p-values, ITT vs per-protocol — with verbatim quotes and fixes.
  2. 02Research Gap Detector (Analyze → Gaps): finds what your own library covers but your text ignores, and drafts a copy-pasteable paragraph citing the relevant papers.
  3. 03Novelty Checker (Analyze → Novelty): scores each claim 0–10 against the library — novel / incremental / restated / contradicted.
  4. 04Claim Grounding (PaperQA2-style): splits the section into atomic claims and labels each supported / partial / unsupported, with the supporting excerpt and a suggested search for the gaps. One click starts a self-feedback refinement pass.
  5. 05Originality check: semantic similarity against your own library, paragraph by paragraph, with an adjustable threshold — a warning tool, not a plagiarism verdict.

TipFix statistics findings first. They are the most common desk-reject reason and the cheapest to repair.

Step 12

Defense Panel and Readiness Score

Two orchestrators that answer one question: is this defensible yet?

  1. 01Defense Panel: five AI reviewers — Clinician, Statistician, Methodologist, Journal editor, Committee member — run in parallel and return their headline concern, likely criticisms, weak areas, unsupported claims, missing literature and a suggested defence answer you can rehearse.
  2. 02From a finished defence report press "Apply suggestions inline" to get 8–18 Grammarly-style accept/reject edits, each anchored to a verified quote from your text.
  3. 03Readiness Score: one 0–100 number per section from seven weighted signals (citation trust 15% · citation density 10% · EQUATOR 15% · statistics 20% · gap coverage 15% · novelty 15% · length 10%) plus up to 12 priority-sorted action items.
  4. 04Project Readiness Overview: word-weighted score for the whole manuscript, five readiness buckets and cross-section action items — click any item to jump to the section.

TipDo not chase 100. Above ~85 the remaining points usually require new data, not new writing.

Step 13

Export and bibliography

The "Exports" tab turns the workspace into a submission package.

  1. 01Formats: Markdown, HTML, DOCX, PDF and LaTeX (DOCX/PDF from the paid plans, LaTeX from Researcher upwards).
  2. 02Numbered citations and the reference list are rendered identically in every format.
  3. 03Choose the citation style (APA 7 / Vancouver / AMA 11 / BibTeX) before exporting; DOIs and PMIDs are resolved through Crossref and NCBI.
  4. 04If a Python project is attached, the export also contains a "Software" reference block.
  5. 05Approve each section ("human reviewed") before export — the manuscript then carries an explicit statement that a human verified the AI-assisted passages.

TipExport a DOCX early and read it outside the app. Formatting problems and repetitive phrasing are far more visible there.

Step 14

Article from a Python project

For software and model papers: attach a repository and let InkPhD extract verifiable technical facts with file-and-line provenance.

  1. 01"Python project" tab → connect a GitHub repository (URL, branch/ref, token for private repos) or upload a ZIP archive.
  2. 02Pick the component you are writing about; relevant files are filtered and indexed into a separate code index.
  3. 03Extract facts: each claim is categorised and labelled code-verified, needs a literature reference, or needs a validation study. Version and commit are detected automatically.
  4. 04Every claim links to the real snippet — file, line range and an "Open on GitHub" deep link.
  5. 05Draft the Methods or System description section; markers [C#], [REF NEEDED] and [VALIDATION REQUIRED] show exactly what still needs your input. Insert it as a new section or append it to an existing one.
  6. 06After you push new code, "Re-sync" flags the claims that may no longer hold.

TipNever let a performance number reach the manuscript without a validation study behind it — the module deliberately refuses to invent metrics.

Step 15

Rules of good practice (read this once)

InkPhD is infrastructure, not an author. Journals and universities hold you responsible for every sentence.

  1. 01Read every generated paragraph against the cited source. If you cannot defend the sentence in a viva, rewrite or delete it.
  2. 02AI is not an author (ICMJE). Disclose AI assistance the way your journal requires — InkPhD keeps the per-section metadata you need for that statement.
  3. 03No patient-identifying data in prompts or uploads; the platform is research-use-only and is not a medical device.
  4. 04Your manuscripts and library are never used to train models. You can export or erase everything at any time from Profile → Data & privacy.
  5. 05Keep your own backup: export a DOCX or Markdown snapshot at every milestone.

TipThe strongest workflow is boring: small sections, good sources, verify, approve, export. Repeat.

InkPhD is research-use-only writing infrastructure. Every AI-generated passage must be read, verified and approved by you before it enters a submission.

Back to my projects

InkPhD stores only what is strictly necessary to run your session and remember your workspace preferences. No advertising, no analytics, no cross-site tracking. Cookie Policy · Privacy