| name | manuscript |
| description | Academic manuscript workflow — structure, verified citations, drafting, and rigor critique for papers, theses, proposals, and referee responses. Sequences evidence before prose so every citation is registry-verified before it is written. |
| argument-hint | <manuscript task — draft, revise, respond to referees> |
| disable-model-invocation | true |
Academic Manuscript
Produce or revise an academic manuscript to a submission-credible standard: structure decided before drafting, every citation verified before it is written, claims audited before it is called done.
Primary task:
$ARGUMENTS
Research-Craft Calibration
Full doctrine: ~/.claude/quality-pack/research-craft/citation-integrity.md (references), ~/.claude/quality-pack/research-craft/figure-craft.md §1 (captions carry the argument). Inline baseline:
- Verify-before-write: no reference enters the draft until registry-verified this session (Crossref/OpenAlex/Semantic Scholar/arXiv). DOIs are resolve targets, never generation targets. Recalled-but-unverified sources appear only as
[citation needed: <what>] markers.
- Existence is not faithfulness: load-bearing claims get an anchor (quote or section locator) showing the source actually supports them.
- Numbers come from current analysis outputs — re-check every quantitative claim against the latest results before submission (stale-numeric-claim reconciliation).
- Captions stand alone: what is plotted, method, N, and what the error bars are.
Workflow — evidence before synthesis, synthesis before polish
- Structure — dispatch
writing-architect for the section plan: thesis/contribution, IMRaD skeleton (abstract = context → gap → approach → key result → implication), evidence plan per section, target journal/format constraints (length, reference style, figure count).
- Evidence — dispatch
literature-scout to verify every source the outline needs BEFORE drafting begins. Its verified-record table is the only citation pool the draft may use. Gaps come back as explicit [citation needed] markers, not fabricated fillers.
- Draft — dispatch
draft-writer with the structure + verified records + analysis outputs. LaTeX projects: edit the project's .tex/.bib in place, matching its macros and journal class (revtex, achemso, elsarticle, …); BibTeX entries generated from registry metadata only. Typeset quantities with siunitx (\SI{77.5}{\micro\siemens}, \num{2.5e-5}) when the project uses it — value ± uncertainty with units set consistently is the typographic half of the rigor doctrine's units mandate.
- Critique —
editor-critic for argument/structure/prose AND, whenever the manuscript makes quantitative claims, rigor-reviewer for methods, statistics, figure compliance, and claim-citation support. Act on findings; stricter verdict wins.
- Reconcile and close — re-run the citation audit on the final text (every citation still verified, no
[citation needed] markers remaining unless the user explicitly accepts them), reconcile every number against current analysis outputs, and confirm figures/captions pass figure-craft §7.
Referee-response mode
When the task is a response to reviewers: quote each referee point verbatim, answer point-by-point (what changed, where — page/line/figure), never concede-and-ignore, and keep the tone collegial and factual. New claims added in response to referees get the same verify-before-write treatment as the original draft. Track manuscript deltas explicitly so the response letter and the diff agree.
Execution rules
- Follow all
/autowork operating rules: verify, review, don't stop early.
- The stage order is load-bearing: scout before draft (verification precedes writing), critique before done.
- LaTeX verification: compile when the toolchain is available (
latexmk -pdf, or the project's build command) and fix errors/warnings that affect output; when no toolchain exists, say so and hand the user the exact build command to run.
- A draft with unverified citations is not done — surfacing the unverified list is part of the deliverable.
