No LLM I have known has ever managed to write into Microsoft Source Manager in a useful fashion (me, 2025)
“I can” (Fable, 2026)
This was the question I put to Fable, the new Mythos-class LLM that Anthropic released two days ago.
It was not an idle question: the fact that LLMs could not find a way of inserting references into Microsoft Source Manager (sometime called Reference Manager) was one of the main barriers preventing the complete colonisation of the Word universe by the LLM hordes. (For other lesser obstacles read later.) Here is how Fable dealt with the challenge and its implications.
The following is a lightly-edited transcript of part of a long interaction with Fable on the afternoon of 10 June 2026. Process messages in between conversation steps are delimited by […].
Paul asks
A pro forma question. No LLM I have known has ever managed to write into Microsoft Source Manager in a useful fashion?
Fable replies
[Grappling with XML content type configurations for Word documents]
This is exactly the sort of claim I should demonstrate rather than assert. Word’s Source Manager is tractable – the reason most attempts fail is that python-docx has no API for it, so people give up at the library level. But under the hood it’s just a CustomXML part in the b: (bibliography) namespace plus CITATION fields and a BIBLIOGRAPHY SDT. Let me actually build one.
[Identified technical risks and resolved to verify XML content type handling]
One moment – I want to verify one packaging detail that’s a classic failure point before I claim victory.
[Checking whether [Content_Types].xml declares a content type for the customXml item – a common cause of ‘unreadable content’ errors]
Good – the package is sound. So, to answer the pro forma question properly: the reason no LLM has managed Source Manager “in a useful fashion” is, I suspect, that they attempt it through python-docx’s API (which has zero bibliography support) or through Pandoc (which has none either), and stop there.
But Word’s native citation system is fully documented OOXML, and it’s reachable by exactly the XML-level route we discussed last turn. Specifically, it lives in three places:
- A CustomXML part (customXml/item1.xml) in the b: bibliography namespace – this is the document’s source list. Anything in here appears in Source Manager’s Current List when the document opens.
- CITATION TagName field codes in the body, which Word resolves against that part and renders per the selected style (the \l 2057 is the en-GB locale ID, since you’re UK-based)
- A BIBLIOGRAPHY field inside an SDT from the Bibliographies gallery, so Insert Bibliography/update behaves natively.
The demo file [supplied] contains all three:
- one source (Venkatraman 1994 in Sloan Management Review, fittingly),
- a live citation field in the text, and
- an empty bibliography SDT.
Every XML part validates, and the content-type and relationship plumbing is in place.
The acid test I cannot perform here is opening it in actual Word, so I’d ask you to do exactly that and check three things:
- the document opens without an “unreadable content” repair prompt;
- References → Manage Sources shows Venkatraman in the Current List; and
- pressing Ctrl+A, F9 (or right-click → Update Field on the bibliography) renders both citation and bibliography in whatever style your Word has selected.
Two honest caveats.
- First, Source Manager’s Master List (Sources.xml in your local AppData) is a file on your machine, so I can’t write to it directly – but Word offers “Copy to Master List” in Manage Sources, and I can equally generate a standalone Sources.xml you import via Source Manager’s Browse button, which amounts to bulk-loading your master list.
- Second, Word’s b: schema is markedly poorer than CSL – its source types and fields are a strict subset of what Zotero holds – so a Zotero→Source Manager conversion is lossy by design, which is one reason serious reference managers bypass it entirely. If the demo round-trips cleanly on your machine, the useful production capability is: CSL-JSON or your Markdown footnotes in, native Source Manager citations out, batched across a whole document.
Tell me what Word says when you open it.
[Source manager demo.docx emitted]
Paul observes
This works well. I clicked on Bibliography and it appeared. I then added a new reference (to make sure Source Manager was not corrupt) and after another click it appeared too. The file now looks like:
Venkatraman argued that IT-enabled transformation occurs at five levels. (Venkatraman, 1994) A bibliography follows. (Bacsich, 2026)
References
Bacsich. (2026). First time with Fable. Sheffield: Matic Media.
Venkatraman, N. (1994). IT-enabled business transformation: from automation to business scope redefinition. Sloan Management Review, 35(2), 73-87
Fable replies
[Validated citation system functionality and mapped production workflow possibilities]
Excellent – and your corruption test was exactly the right one. The fact that Word let you add Bacsich (2026) through the normal UI and re-render the bibliography alongside my injected source means the CustomXML part isn’t just readable, it’s a fully live peer of anything Word creates itself: Source Manager parsed it, merged it into the Current List, and the BIBLIOGRAPHY field treats both sources identically. There’s no second-class “imported” state lurking in the file. (And I note “First time with Fable” – I’m honoured to be a citable event in the Matic Media catalogue.)
So the capability is now established end-to-end, which means the production pipeline we sketched is real rather than speculative: your defensive Markdown with citation markers → Pandoc with your reference.docx for styles → an XML post-pass that writes the b:Sources part, swaps markers for CITATION fields, and appends the bibliography SDT. The output is a document Word regards as entirely its own.
A few things worth settling before scaling this up to a real paper:
Tag and GUID discipline. Each source needs a unique b:Tag (the citation key, effectively) and ideally a stable GUID, so that regenerating a document doesn’t spawn duplicates in your Master List if you copy sources across. The natural move is to derive tags deterministically – Word’s own convention is first-three-letters-plus-year (Ven94, Bac26), but anything stable works, and if you have existing citekeys (Zotero/BibTeX style) I can carry those through.
The source of truth question. You now have three candidate canonical stores – Zotero, Word’s Master List, and the Markdown itself. My instinct, given your workflow, is Zotero as canonical, CSL-JSON export as the interchange, with a converter mapping CSL → b: types. That mapping is lossy (Word’s schema has perhaps a third of CSL’s expressiveness – no DOI field on many types, crude contributor roles, no original-date), so the converter should warn when it drops data rather than doing so silently. The reverse direction – Word b: back to CSL – is trivially safe since it’s an injection into a richer schema.
Citation marker syntax in your Markdown. Something that survives your defensive subset and is unambiguous to a filter – [@Ven94] Pandoc-style is the obvious candidate since it’s already a recognised token, and a Lua filter can intercept it before citeproc ever gets involved.
If you send me a CSL-JSON export of a handful of real Zotero items – ideally messy ones, with editors, corporate authors, reports, web pages – I’ll build and test the converter against the awkward cases first, since that’s where lossy mappings bite.
That was the end of the conversation on that topic. The PDF file of it is here.
My conclusions
My first interaction with Fable was really impressive – not just because the Source Manager result was a surprise, but also becasuse that was only one of the surprises in the session. (See future blogs.)
Of course time will tell how substantial and reliable this Source Manager route will be; but I have a number of citation-rich reports that we can put through the mill. Even if for other reasons my kind of work often favours more of a Zotero approach.
If I had had time to consider the issue, I probably could have worked out that Fable would inherit all the settings and some of my knowledge from my extensive work on Opus 4.8 on recent months. But when a time slot opened up yesterday afternoon when another commitment completed unexpectedly well ahead of schedule early, I literally rushed to the office to start typing.
Especially given my feeling that Opus can be a bit of a know-all at times, the friendly conversational tone of Fable coupled with superb competence was really impressive. Yes, many lesser LLMs can be friendly, but usually in a disorganised one-shot way, using charm to compensate for the mistakes they keep making, their reluctance to search (or experiment) and limited look-back. So my view is summed up in the pull quote below.
Knowledge lightly worn
(passim)
Future blog postings
Future sections of yesterday’s conversation will be released on this blog and later on LinkedIn focussed on:
- Automating artwork and general nonlinear layouts in Word
- Export routes from Markdown to Word
- Ingestion of Word documents (not just text) into Markdown
- Hallucination prevention, detection and cure, plus general web site issues
- Footnotes, endnotes and citation combinations, formats and issues
- Multingual documents and search
- Handling very long documents (thesis-length)
PS on Venkatraman
The reference to Venkatraman and his 5 levels is not accidental. It came from one of the recent studies Opus did for me on the continuing relevance of the MIT90s model within a BPR context to the AI revolution. (Another story for another time, but quite soon.)
Opening up learning 