The manubot lineage: manuscript-as-software

Last updated: 2026-10-20

Synthesis

Manubot (Himmelstein et al., 2019) is the conceptual anchor for scriptorium. It articulated the manuscript-as-software pattern: a scholarly manuscript should be a git repository whose source files are plain text, whose builds are automated, whose changes are reviewable as pull requests, and whose citations are identifier-driven rather than copy-pasted bibliographies. Manubot’s seven-year track record demonstrates that the pattern works in practice — dozens of papers have been authored in manubot, including the manubot paper itself.

Yet the pattern has limits. The git-based manuscript culture remains a niche of computational biology and the open-science movement. Most scholars still write in Word or Overleaf, with manual citation management. And critically — for the scriptorium thesis — manubot’s coordination model is a build pipeline, not an editorial pipeline. It tracks who edited what, but it does not have a shared notion of “what this manuscript is arguing,” “what the core claims are,” “what terminology should be preserved.” Those concerns live in author heads, not in the repo.

Scriptorium extends manubot in exactly this way. It adopts the manubot inheritance — git-tracked manuscripts, plain-text source, identifier-driven citations, inspectable transformations — and adds the missing primitive: shared editorial state (MANUSCRIPT_STATE.yaml) that AI agents and human authors both read.

Landscape

Manubot (the foundational paper)

Himmelstein DS, Rubinetti V, Slochower DR, Hu D, Malladi VS, Greene CS, Gitter A. “Open collaborative writing with Manubot.” PLOS Computational Biology 15(6): e1007128 (2019). DOI: 10.1371/journal.pcbi.1007128. PMID: 31233491. PMC6611653. [1]

Key contributions of the paper:

1. The pattern. Manuscripts written in Markdown, stored in git, hosted on GitHub, built continuously into HTML/PDF. Multiple authors propose changes via pull requests; a cloud service evaluates them.
2. Citation-by-identifier. Authors write [@doi:10.1371/journal.pcbi.1007128] or [@pubmed:31233491] and manubot retrieves the metadata and converts to a target style. This eliminates the “wrong bibliography entry” failure mode that plagues Word workflows.
3. Continuous publication. When source changes, rendered outputs rebuild and republish to a web page. There is no “submission” event; there is a continuously-current version.
4. Versioned, immutable history. The git log is the manuscript’s history; nothing is lost.

The software is at manubot/manubot (476 stars, last pushed 2026-08-02) [2] and the user-facing template is manubot/rootstock (480 stars, last pushed 2026-01-18) [3]. The paper itself was authored in manubot at greenelab/meta-review (51 stars, last pushed 2020-05-25) [4] — that repo has not been updated since 2020, a useful datapoint on post-publication maintenance.

The Greene Lab lineage

The manubot pattern was used for multiple Greene Lab papers and collaborations, most notably the deep-review project on deep learning in biology and medicine. The pattern proved itself at scale (40+ contributors on a single manuscript). The Greene Lab’s broader contribution is that peer-review-grade manuscripts can be authored using software-engineering conventions: branching, merging, CI, PRs, issue tracking, public version history.

Quarto Manuscripts (Posit, 2023–)

Quarto is Posit’s open-source successor to R Markdown and the most active current development in the manuscript-as-software space.

• quarto-dev/quarto-cli has 6,104 stars, last pushed 2026-10-20 [5] — actively maintained.
• “Reproducible Manuscripts with Quarto” (Çetinkaya-Rundel, 2023) at posit::conf [6] introduced the Quarto Manuscript project template: a self-contained directory whose source produces HTML, PDF, JATS XML, and a connected index.qmd with executable code blocks.
• USGS scientists published a Quarto manuscript template, and demonstrated it for the Journal of Fish and Wildlife Management, in a 2024 article in that journal (Erickson, Archer & Fienen; doi:10.3996/jfwm-24-003) [7][8] — early evidence of journal-side adoption.

Quarto Manuscripts extends manubot in two ways: (a) executable code embedded in the manuscript (R, Python, Julia, Observable JS via Jupyter/Knitr); (b) multi-format build with rich cross-references (sub-figures, layout panels, hoverable citations, callouts). Where manubot is “manuscript-as-prose-software,” Quarto Manuscripts is “manuscript-as-executable-document.”

quartobot and Quarto + manubot

quartobot is now a maintained codebase rather than a loose community pattern. It lives at seandavi/quartobot (MIT-licensed; created 2026-05-13; last pushed 2026-10-01) [12] and is published on PyPI (first release 0.1.0 on 2026-05-14; 0.6.0 on 2026-05-28) [13]; the earlier quartobot/quartobot repository is an archived redirect to it. Its README describes it as “Citation resolution and manuscript-as-software CI for Quarto”: a Quarto pre-render hook resolves persistent-identifier cite keys (@doi:, @pmid:, @arxiv:, @isbn:, @url:, @wikidata:, @pmc:) through manubot’s resolver into a committed bibliography, adds render and PR-preview CI scaffolding, and exposes the resolver to agents through an MCP server [12]. It describes itself as the work to close manubot’s open “Quarto integration” issue, opened in April 2022 [12][14]. (Disclosure: quartobot is maintained by scriptorium’s author, and this knowledge layer’s .qmd notes build with it.) The conceptual point stands: Quarto and manubot are converging, with Quarto absorbing most of the manuscript-as-software energy circa 2025–2026 — but the identifier-driven citation step still comes from manubot, now via a third-party Quarto extension rather than Quarto itself.

Distill.pub

Distill was a peer-reviewed ML journal (founded 2017, supported by Google, OpenAI, DeepMind, Y Combinator Research) [9] that demonstrated interactive scientific publishing: articles could include explorable visualizations, live model interactions, and code-driven figures. Articles included Olah’s “Building Blocks of Interpretability,” Carter & Nielsen’s “Using Artificial Intelligence to Augment Human Intelligence,” and others that defined a new aesthetic for ML communication.

Distill announced a one-year hiatus in 2021 “which may be extended indefinitely”; its about page now describes it as on indefinite hiatus. The published explanation cited volunteer editor burnout, the structural friction of running a journal, and the editorial cost of interactive articles — with early articles, volunteer editors often spent 50 or more hours improving a single submission [9]. The lesson for scriptorium: the editorial labor cost of high-quality publication does not vanish when you change the technology — it shifts. Distill failed not because the tooling was bad but because the editorial pipeline was expensive.

Idyll

Idyll (Conlen & Heer, ACM UIST 2018, DOI: 10.1145/3242587.3242600) [10] is a markup language for authoring interactive narratives. Repo idyll-lang/idyll has 2,042 stars, last pushed 2023-02-04 — effectively dormant. Idyll’s intellectual contribution (a structured author interface to JavaScript components, allowing rich interactivity without becoming a web developer) has been largely absorbed by Observable, Quarto, and Svelte-based article systems.

Observable / Observable Notebooks

Mike Bostock’s (d3.js) Observable platform [11] is the most active interactive-publishing environment as of 2026. Observable Notebooks are reactive JavaScript documents with executable cells and inline visualization. The Observable Framework is a static-site generator for data apps. Observable is conceptually intertwined with the manuscript-as-software pattern.

What the pattern has and hasn’t achieved

Has:

• Reproducibility of computational figures (Quarto, Observable).
• Citation accuracy via identifiers (manubot).
• Git-tracked, multi-author manuscript history (manubot, Quarto).
• Open-by-default authoring (manubot, Distill while it lived).
• Continuous publication / no “submission” event (manubot).

Hasn’t:

• Mainstream adoption. Word and Overleaf still dominate.
• A shared model of what the manuscript is doing (claims, terminology, target venue, constraints). Manubot and Quarto track what is written, not what is meant.
• Editorial workflows beyond the build. There is no “review this PR for argument structure” in the manubot pipeline — only “does it build?” and “do humans approve?”
• AI integration. Manubot pre-dates the LLM era; Quarto post-dates it but does not integrate AI capabilities natively. (quartobot’s MCP server, 2026, exposes citation resolution to agents — resolution only, not editorial judgment [12].)
• Sustainability of high-end interactive publication (Distill is the cautionary tale).

How this informs scriptorium

• Scriptorium is an editorial layer, not a manuscript-rendering layer. Manubot and Quarto handle source → output rendering. Scriptorium handles source → improved-source. The two are complementary: a Quarto manuscript can be the host for MANUSCRIPT_STATE.yaml, and scriptorium skills can operate on the Markdown/Qmd source.
• The MANUSCRIPT_STATE.yaml file is the missing primitive in the lineage. Manubot’s content/ directory and metadata.yaml track structure. Quarto’s _quarto.yml tracks build config. Neither tracks editorial intent. This is exactly where scriptorium contributes.
• Identifier-driven citations are non-negotiable, but the syntax is not. The right unit is the persistent identifier itself (DOI / PMID / arXiv ID), not any particular cite-key convention. Manubot’s [@doi:...] form is one resolver’s encoding; CSL-JSON entries, BibTeX entries with verified DOIs, or even plain markdown DOI links ([DOI: 10.x/y](https://doi.org/10.x/y)) all satisfy the requirement. Scriptorium’s citation-audit assumes an identifier-resolved bibliography and refuses to operate on bare-text citations, but is deliberately agnostic to which resolver (manubot, quartobot, biblatex, none) the author uses — see the “Skill citation discipline” section in CONTRIBUTING.md.
• Distill’s failure mode is a design lesson. The editorial labor of producing a great paper does not go away when you add AI; it gets redistributed. Scriptorium should be designed so that AI capabilities reduce the load on editorial labor, not just add a new ingestion task (“now you also have to maintain this YAML file”).
• The greenelab/meta-review repo’s 2020-05-25 last-commit date is a sustainability signal. The manubot paper itself stopped being maintained five years ago. Scriptorium will face the same pull. The defense is: keep the surface area small, document the pattern such that others can carry it, and make MANUSCRIPT_STATE.yaml valuable enough that authors maintain it whether or not the toolchain is alive.

Implementation priority for scriptorium

Verdict: No (framing only).

Why this is useful context anyway: this lineage is scriptorium’s conceptual anchor and positions it as an editorial layer on top of manubot/Quarto rendering, not a competitor to them. It justifies MANUSCRIPT_STATE.yaml as the missing primitive and the resolver-agnostic citation stance citation-audit takes. No skill renders manuscripts or resolves identifiers. The 2026-10 refresh (quartobot now a maintained package; updated repository figures; the Distill and USGS corrections) did not change this verdict.

Open questions / weak evidence

• No published evaluation of manubot’s claimed benefits. The 2019 paper makes the case theoretically; there is no empirical study (then or since) of whether manubot manuscripts are better, more accurate, more reproducible, or more widely cited than Word-authored equivalents. Scriptorium could contribute by being designed for evaluation from the start.
• Quarto Manuscripts is still pre-mainstream. USGS adoption is encouraging, but the workflow is most common among R-centric researchers. JATS XML output is the strongest signal for journal-side integration.
• The manubot/Quarto convergence is partial. Quarto has not absorbed identifier-driven citations as a first-class concept (it supports CSL but citation entries are still local). Third-party bridges now exist (quartobot, 2026 [12]), but manubot’s “Quarto integration” issue remains open [14] and no bridge has been adopted by Quarto itself.
• The “is markdown enough for science?” debate is unsettled. Word-style track changes, semantic comments, and graphical annotation are difficult to replicate in pure-text workflows. The biggest practical adoption blocker is not the tooling but the social workflow of co-author commentary.
• Distill’s dates differ by source. Its about page says it “operated 2016-2021”; Wikipedia gives a March 2017 founding [9]. The difference may reflect pre-launch work versus public launch; it does not affect this note’s argument.

References

1. Himmelstein DS, Rubinetti V, Slochower DR, Hu D, Malladi VS, Greene CS, Gitter A. “Open collaborative writing with Manubot.” PLOS Computational Biology 15(6): e1007128 (2019). DOI: 10.1371/journal.pcbi.1007128. https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1007128
2. manubot/manubot GitHub. https://github.com/manubot/manubot (476 stars; pushedAt 2026-08-02; checked 2026-10-20).
3. manubot/rootstock GitHub. https://github.com/manubot/rootstock (480 stars; pushedAt 2026-01-18; checked 2026-10-20).
4. greenelab/meta-review GitHub. https://github.com/greenelab/meta-review (51 stars; pushedAt 2020-05-25; checked 2026-10-20).
5. quarto-dev/quarto-cli GitHub. https://github.com/quarto-dev/quarto-cli (6,104 stars; pushedAt 2026-10-20; checked 2026-10-20).
6. Çetinkaya-Rundel M. “Reproducible Manuscripts with Quarto.” posit::conf 2023. https://mine.quarto.pub/manuscripts-conf23/
7. USGS. “A Reproducible Manuscript Workflow with a Quarto Template.” https://www.usgs.gov/publications/a-reproducible-manuscript-workflow-a-quarto-template
8. Erickson RA, Archer AA, Fienen MN. “A Reproducible Manuscript Workflow With a Quarto Template.” Journal of Fish and Wildlife Management 15(1):251–258 (2024). DOI: 10.3996/jfwm-24-003. https://meridian.allenpress.com/jfwm/article/15/1/251/501496/A-Reproducible-Manuscript-Workflow-With-a-Quarto
9. Distill journal. https://distill.pub/about/ ; Wikipedia entry on Distill (journal). Hiatus announcement: https://distill.pub/2021/distill-hiatus/
10. Conlen M, Heer J. “Idyll: A Markup Language for Authoring and Publishing Interactive Articles on the Web.” ACM UIST 2018. DOI: 10.1145/3242587.3242600. https://dl.acm.org/doi/10.1145/3242587.3242600 ; repo idyll-lang/idyll (2,042 stars; pushedAt 2023-02-04; checked 2026-10-20).
11. Observable. https://observablehq.com/
12. seandavi/quartobot GitHub. https://github.com/seandavi/quartobot (MIT; created 2026-05-13; pushedAt 2026-10-01; checked 2026-10-20). Documentation: https://seandavi.github.io/quartobot/
13. quartobot on PyPI. https://pypi.org/project/quartobot/ (0.1.0 uploaded 2026-05-14; 0.6.0 uploaded 2026-05-28).
14. manubot/manubot issue #332, “Quarto integration” (opened 2022-04-05 by agitter; open as of 2026-10-20). https://github.com/manubot/manubot/issues/332