import-manuscript

name	import-manuscript
description	Bring a manuscript that already exists — a .docx from a supervisor, a LaTeX chapter, a PDF of last year's submission — into the project so the kit's own machinery applies to it. Converts and splits it into section files, reads its reference list as a list of DOIs (re-fetched from the authoritative record, never transcribed), and reports which references cannot be resolved. Nothing in the text is edited and no citation is rewritten. Use when the paper was NOT written here and the author wants it verified, reviewed, restyled or finished.
disable-model-invocation	true
argument-hint	<file.docx|.tex|.pdf|.md> [--resolve]

/import-manuscript — The Paper That Already Exists

What this command is, and is not

It is the on-ramp: a document becomes section files under drafts/, its reference list becomes register entries with authoritative BibTeX, and from that point /verify-citations, /critique, /revise, /restyle and /submit-check work on it exactly as they would on a paper drafted here.

It is not a rewrite. Not one sentence of the author's text is changed by this command, and no [12] becomes [@key] without the author looking at it. An import that quietly edited the argument would make every later check circular — the kit would be verifying its own edit.

Gate: Abort if Failed

1. Check: PROJECT.md exists. If not: run /new-project first — the import writes into a project's drafts/, sources/ and bib/, and those have to exist and belong to something.
2. Check: the file exists and is one of .md .docx .tex .pdf .odt .rtf .html.
   • .docx / .tex / .odt / .rtf need pandoc; .pdf needs pymupdf4llm. Both exit BLOCKED with the install line — the manuscript is never hand-parsed as a fallback, because a .docx read with a regex turns a table into a paragraph and a footnote into body text.
3. Prefer the source over the PDF. If the author has both, take the .docx or .tex. A PDF conversion loses the heading structure, and the import then produces one long file the author has to split by hand.

Procedure

Step 1: Plan, and show it before writing anything

python scripts/manuscript_import.py --file <path> --plan

Report to the author, in their language:

• the sections it found and the files they would become (this is where a bad conversion is visible: three sections from a 40-page thesis means the headings did not survive);
• the word count — against PROJECT.md's budget if one is set;
• the reference list: how many entries, how many carry a DOI, how many do not;
• the citation style it detected (numeric / author-year / pandoc) — this decides what Step 3 can offer;
• any warning, verbatim. "No headings" is the important one.

If a target file already exists the command refuses and names it. Do not reach for --force on the author's behalf: a section file that exists is a section someone has been editing.

Step 2: Import, then resolve the references

python scripts/manuscript_import.py --file <path>            # sections
python scripts/manuscript_import.py --file <path> --resolve  # + bibliography

--resolve sends every DOI in the reference list through doi2bib.py and registers each entry as a candidate stamped metadata_source: "manuscript reference list (unverified)". Say plainly what that means: the bibliography now holds the publisher's record, not the author's typing, and the two are allowed to differ — that difference is a finding, not an error to smooth over.

Entries without a DOI are listed in the report and are the author's next piece of work. For each one:

python scripts/doi2bib.py --title-search "<title>" --year <year>

Never fill one in from memory. A reference that cannot be resolved stays unresolved and /verify-citations will say so — which is the honest outcome and occasionally the true one (a real book chapter with no DOI).

Step 3: The citation rewrite is a proposal, not an edit

When the manuscript uses numeric citations, drafts/citation-map.json holds {"12": {"doi": …, "bib_key": …}}. Offer the conversion to pandoc syntax section by section, with the author reading each mapping, and only if they want it. State the trade honestly:

• converting means /verify-citations, claim_check.py and cite_audit.py can see the citations at all — until then, most of the kit's citation machinery is blind to this manuscript;
• a mis-mapped [12] attributes a claim to the wrong paper, and once the number is gone the error is unrecoverable by inspection.

If the author declines, that is a legitimate end state. Record it and continue — the register and the bibliography still improve every later step.

Step 4: Bind the numbers, if this is an empirical paper

An imported manuscript's results have no analysis/ behind them. Ask whether the data and the scripts exist:

• data exists → /analyze-data or /experiment re-establishes the chain, and analysis/counts.json lets consistency_check.py reconcile prose against it;
• they do not → say so explicitly in the report. failure_scan.py's M3 check and experiment_db.py trace will both come back empty, and the author should know that this is a property of the import, not a verdict on the paper.

Step 5: Hand off to the normal pipeline

python scripts/cite_audit.py          # in-text ↔ bibliography, both directions

Then, in order: /verify-citations → /critique → /revise → /submit-check. Nothing about the rest of the pipeline changes for an imported manuscript.

Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
📥 Makale içe aktarıldı — tez-taslak.docx
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Bölümler:  7 dosya · 8.412 kelime (bütçe: 10.000)
Atıf stili: numaralı · metinde 46 farklı numara
Kaynakça:  52 kayıt · 41 DOI'li (çekildi) · 11 DOI'siz (aşağıda)
Sicil:     SRC-0001…SRC-0041 (aday, "unverified" damgalı)
Eşleme:    drafts/citation-map.json — öneri, uygulanmadı

DOI'siz 11 kayıt: <ilk 5, ad ve yıl>
Sonraki: doi2bib --title-search ile çöz → /verify-citations

Never report an import as complete while entries are unresolved without naming them. The count that matters to the author is not "41 fetched" — it is "11 that nothing has checked".

Notes

• This is not /import-library. That one takes a Zotero/Mendeley export — references only. This one takes the manuscript, and its reference list is a by-product. Running both is normal; the register dedups.
• A PDF-only manuscript is a degraded import and should be reported as one: no heading structure, a reference list mangled by column breaks, and often hyphenated line-ends inside the DOIs. The .docx behind it is almost always worth asking for.
• Existing figures and tables come across as pandoc's media folder; the import does not renumber them. figure_check.py will report numbering and caption problems once the sections are in place.
• This writes drafts/ before an outline exists, and that is not a gate bypass. The outline gate stops the kit from generating prose with nothing behind it; an imported manuscript is prose that already exists and that the author wrote. What the gate still governs is unchanged: /draft on top of this project remains blocked until an outline is approved, so anything the kit adds to the imported text goes through the same door as always. If the author wants to keep working on it section by section, build the outline from the imported structure first — /outline will bind it to the evidence cards, which is the step the imported manuscript never had.

Reference

• Scripts: python scripts/manuscript_import.py (--plan, --resolve, --force) · python scripts/doi2bib.py --title-search (unresolved entries) · python scripts/cite_audit.py
• Writes: drafts/<nn>-<slug>.md, drafts/citation-map.json, and — through their own sanctioned writers — sources/registry.jsonl and bib/references.bib
• Related: /import-library (references only), /verify-citations (the point of doing this), /translate (if the target venue's language differs), /retarget, /critique

---

An imported manuscript's reference list is a claim about 52 papers. Re-fetching them is how it becomes 52 records.