/mstack:draft-section

Stage: write · Voice: writer, anchored to voice.writing_style in .mstack/config.yaml

Draft once results are stable (output/tables/ and output/figures/ no longer changing) and a target journal is set; prose written before results lock gets rewritten.

$ARGUMENTS is the section: intro, theory, data, methods, results, discussion, or abstract. If missing or unrecognized, list these and stop.

Procedure

1. Load .mstack/config.yaml (title, target journals, status), .mstack/learnings.jsonl (conventions, variable names, framing), every existing section in paper/sections/ (the manuscript is the context; consistency across sections is a referee's first check), the citable keys of paper/refs.bib (grep -o '^@[A-Za-z]*{[^,]*' paper/refs.bib), and the upstream memos the section depends on:

   Section	Upstream memos and artifacts
   intro	.mstack/research-question.md, .mstack/lit-map.md, .mstack/hypotheses.md (estimand), the primary table and figure
   theory	.mstack/theory.md (mechanism, DAG, rivals, scope), .mstack/hypotheses.md
   data	data/codebook.md, data/raw/PROVENANCE.md, data/clean/clean-log.md, the variation figure
   methods	.mstack/identification-review-*.md, prereg/osf-prereg.md, code/02-analyze.R
   results	code/02-analyze.R, output/tables/*, output/figures/*, output/analyze-log.md, .mstack/robustness-*.md
   discussion	.mstack/theory.md (scope, rivals), .mstack/identification-review-*.md, .mstack/robustness-*.md
   abstract	intro, results, discussion

   Every number in methods / results prose must match the tables on disk.

2. Read ${CLAUDE_PLUGIN_ROOT}/references/writing-conventions.md for what the section owes the reader and the referee complaints it must pre-empt.
3. Voice. If voice.writing_style names a skill, invoke it for tone, rhythm, and vocabulary. If unset, write clean generic academic prose (short sentences, active verbs, no hedge-stuffing, no thesaurus reaches) and tell the user once that they can set a style skill in .mstack/config.yaml.

4. Draft to the section's bar:

   Section	Bar
   intro	Opens with the question and its stake; a contribution paragraph with what we do, what we find (a number in substantive units), and what changes if believed; the design and estimand in one sentence; positioned against the two or three closest papers; roadmap in one sentence or none. The puzzle survives a "so what?" attack in two sentences.
   theory	Mechanism in one sentence and its steps; assumptions; observable implications that separate it from the named rivals; scope conditions as where it fails; hypotheses numbered with direction and estimand.
   data	Unit and coverage; sources with vintages; each key variable's operationalization with a validity argument; the variation the design uses, shown; missingness and restrictions with counts; every claim cites the codebook.
   methods	Estimand, then the identifying assumption in one sentence, then the equation with every symbol defined, fixed effects, and clustering; why this comparison isolates the effect; the falsification tests and sensitivity analysis; preregistration status and deviations; the estimator cited.
   results	Primary estimate first, then falsification tests, mechanism, heterogeneity, and a one-paragraph robustness summary with the sensitivity sentence; one claim per table or figure; point estimate, interval, and magnitude against a benchmark; causal verbs only under a design that licenses them; nulls read against the SESOI, never as "no effect"; no "highly significant".
   discussion	The contribution restated with its evidence; what the result is not (the estimand's limits); the most plausible rival explanation and why the evidence weighs against it; specific limitations; implications proportional to evidence; one concrete next study.
   abstract	Question, design with estimand, data, the finding with a number, contribution; within the journal's cap; no citations, no hedging stack.

5. Citations. Only keys already in paper/refs.bib. For a missing one, insert \cite{TODO-author-year-keyword} and append % TODO: add ref — <one-line description>. Never fabricate author-year-title combinations. Cite the methods used (estimators, sensitivity tools) where the reference file names them.
6. Write to paper/sections/<name>.tex (.qmd if .mstack/config.yaml says format: quarto). Overwrite only if the file is empty or a placeholder comment; otherwise show the candidate and ask whether to overwrite, append, or save as paper/sections/<name>.candidate.tex. On the first substantive section, set paper.status: "writing".
7. Self-check before saving: every number in the section appears in a table, figure, or the analyze log; every table and figure referenced exists; causal language matches design.type; the estimand named here is the one in hypotheses.md; no citation key is invented.

Outputs

• paper/sections/<section>.tex.
• Optional .mstack/draft-log.md — one line per draft: date, section, word count, open TODOs.

Anti-patterns

• Fabricated citations. TODO markers, never inventions.
• Drafting results before analysis is stable. Missing tables or figures → stop and run /mstack:analyze.
• Generic voice over a configured style. Defer to the style skill; do not paper over it with hedge phrases.
• A literature review where the theory should be.
• Nulls as zeros and associations as effects.
• Filling space. Short and sharp beats long and hedged.

Next

Before intro, /mstack:lit-map output should exist; before methods, /mstack:identification-review. After abstract, /mstack:abstract-shotgun. After all sections, /mstack:coauthor-review then /mstack:referee-mock.