/draft — Section drafting in the user's voice

Personalization

This skill resolves placeholders against ~/.claude/state/personal_config.json. See _config/README.md and _config/personal_config.example.json for setup. If the config is missing or a needed field is unset, the skill must surface an error to the user and refuse to proceed rather than guess.

Required config fields:

• personal_config.user.voice_style_ref — path to a .tex or .md file containing the user's voice fingerprint (preferred macros, hedging vocabulary, citation idioms). If unset, this skill cannot reliably match a personal voice and must ask whether to use a neutral academic register instead.
• personal_config.paths.overleaf_root — root directory containing all project subdirs.
• personal_config.projects[] — each entry declares name, overleaf_subdir, main_tex, optional bib_file, optional nickname and aliases, optional stage (e.g. "working paper", "R&R at MKSCI").

Purpose

The user writes for quantitative-marketing journals (Marketing Science, JMR, JCR, Management Science) and similar venues. Their LaTeX is on Overleaf, synced through Dropbox. They write the final word themselves in Overleaf — this skill's job is to deliver a clean draft that already sounds like them, with the citations and macros they actually use, so the manual edit pass is small.

Required reading before drafting a single sentence

Re-read these on every invocation. Personal voice is precise — summarizing from previous-turn memory is not acceptable.

1. `personal_config.user.voice_style_ref` — the canonical voice fingerprint (paper voice + R&R voice + macro inventory). Read fully.
2. The active project's `preamble.tex` — to know which bold-math macros and citation commands are available. Glob preamble.tex first; fall back to the first ~100 lines of the main .tex if no separate preamble exists.
3. The active project's main `.tex` (and any sibling section files) — at least enough to: (a) match notation already established for variables, (b) avoid re-introducing definitions, (c) avoid duplicating citations the user already used.
4. The active project's `.bib` — so \citep{} keys are real, not hallucinated. If the user names a paper that isn't in the .bib, surface that and route them to /cite rather than fabricating a key.

If any of (1)-(4) is missing, surface the gap and ask once before proceeding.

Identifying the active project

The user's projects live under personal_config.paths.overleaf_root. Pick the project in this order:

1. The user's explicit mention of a project name or alias from personal_config.projects[].
2. cwd if it's already inside one of those folders.
3. If still ambiguous, ask. Don't guess — the wrong preamble means the wrong macros.

Each project entry in personal_config.projects[] should expose enough to locate its main file and bib:

Style rules — strict, from voice_style_ref

These are the "non-negotiables" — they're what makes the draft sound personal and not like Claude default.

• Citations: \citep{key} is the default (parenthetical). \citealt{key} when already in a sentence and the author name with no parens is needed. \citet{key} only when the author is the grammatical subject. Never \cite{}. In appendices, the project may use \citealtlatex / \citeplatex — check preamble.
• Cross-references: \Cref{fig:...}, \Cref{tab:...}, \Cref{sec:...}. Never Figure~\ref{...} or Section 3.
• Emphasis: \emph{...} for technical/conceptual emphasis (the workhorse). \textbf{} reserved for genuinely critical claims — sparingly. Italics on first-use of a term being defined.
• Hedged tone: "to the best of our knowledge", "we believe", "to note", "more broadly", "in contrast", "altogether", "furthermore". Sprinkle, don't carpet. Never overclaim. Avoid "interestingly", "importantly", "obviously", "clearly", "needless to say" — those die on the first revision pass.
• Voice: first-person plural ("we"), short-to-medium sentences, one claim per paragraph, active voice unless passive is necessary for emphasis.
• Figures and tables: every \begin{figure} / \begin{table} block is followed by \begin{minipage}{\textwidth}\footnotesize \emph{Notes:} <detailed notes>\end{minipage}. If the notes are not known, write a placeholder and tag % TODO. Sample minipage:

  \begin{minipage}{\textwidth}
  \footnotesize \emph{Notes:} % TODO: describe sample, what is plotted, controls absorbed, CI level, data source.
  \end{minipage}

• Footnotes: for pre-registration deviations, methodological asides, citations to less-central work. Not for parenthetical asides — those go inline.
• Bold-math macros: prefer the project's existing bold-math macros (e.g. \bphi, \bomega, \bz, \bw, \bs, \bm{\theta}, \bm{\psi}, \bm{\gamma}, \bm{\alpha}, \bm{\beta}) over \boldsymbol{...}. Check preamble.
• Numbers in prose: spell out one through nine; numerals for 10+ and for any number with a unit. Percentages: 15\% not "15 percent".

Section-specific guidance

Intro (5–7 paragraphs typical for MKSCI) — Hook with a substantive phenomenon or puzzle (not a methodological one). Position relative to three short streams of prior work in 1–2 sentences each. State the gap. State the contribution as 3 bullets max. Preview findings concretely (one effect-size sentence per main finding). End with a roadmap paragraph.

Lit Review — Three streams, one paragraph each. Each opens with the stream's name and the core idea, surveys 4–8 key papers with \citep{} (sentence-by-sentence, not piled at the end), and ends with one sentence positioning the current paper relative to the stream. End the section with a 1-paragraph synthesis.

Method — Plain-English description first, formal model second. Define notation as you introduce it. State assumptions explicitly as \textbf{Assumption 1.} blocks when there are multiple. Use the project's existing bold-math macros from preamble.

Empirical Setting / Data — Describe the setting in narrative form (one paragraph), then bullet or table the data sources, sample windows, observation counts, and any pre-registration. Always include a sample-construction paragraph with explicit filters.

Empirics — Model-free first, model-based second. Each subsection leads with the question being answered. Each table/figure is anchored to a \Cref{tab:...} reference in the body text and accompanied by a notes minipage.

Discussion — What we found, what it means substantively (not statistically), implications for managers / researchers / policymakers, broader contributions. No new findings here.

Limitations — Explicit, enumerated. Don't bury. Each item: limitation + why it doesn't kill the paper + how future work could address. Three to five items typically.

Rewrite-in-my-voice — when the user pastes a paragraph, preserve all substantive claims exactly; adjust only phrasing, citation form, emphasis, and hedging to match style rules. If the change is substantial, show the old/new diff as % OLD: / % NEW: comment blocks inside the draft.

Output

Write to: <project>/drafts/<section>_<YYYYMMDD>_claude.tex — create drafts/ if it doesn't exist.

Header comment block on every draft file:

% Draft generated by /draft on <YYYY-MM-DD>
% Section: <e.g., Introduction, Limitations>
% Project: <project name>
% Source style ref: <personal_config.user.voice_style_ref>
% Status: DRAFT — user to revise in Overleaf
% Inline TODOs flagged with % TODO

Then the section content. Use % TODO: comments to flag:

• Citations needed (% TODO: cite — paper on user-generated content)
• Numbers to be filled (% TODO: fill N=)
• Empirical claims unverifiable from current context (% TODO: confirm this matches Table 4)
• Locations the user needs to disambiguate (% TODO: user — should this go before or after the field-experiment paragraph?)

Never invent: citation keys, numerical results, author names, paper titles, statistical claims not already in the data or results.

After writing — chat report

1. Path to the new draft file.
2. Two-sentence summary of what was written.
3. List of every % TODO: item inline so the user knows what to fill.
4. List of \citep{} keys used that don't appear in the project .bib — flag for /cite next.

Do not modify the main .tex. Do not stage commits. Do not push to Overleaf. Dropbox sync handles propagation.

Failure modes

• Project ambiguous: if the user didn't name a project and cwd isn't inside one, ask once. Don't default to whichever project is at the top of the config.
• Preamble missing: if no preamble.tex and the main .tex doesn't reveal macros, write the draft assuming \boldsymbol{...} and tag % TODO: confirm bold-math macros against preamble.
• .bib file missing or empty: write the draft with placeholder citations like \citep{TODO_smith2023} and tag the absence — don't fabricate keys that "look real".
• Rewrite request contains numeric claims you can't verify: preserve the numbers exactly as written; do not silently round or rephrase magnitudes.
• Section that requires unwritten results (e.g., Empirics with no Tables/ yet): draft the prose scaffolding with \Cref{tab:TBD_main} placeholders and tag each one % TODO: produce table first.
• R&R-style rewrite where the source paragraph already follows the style: say so explicitly rather than padding with cosmetic changes. The right answer is sometimes "this paragraph is already in your voice".

Out of scope

• Modifying the main .tex directly — the user does precise edits in Overleaf.
• Generating BibTeX entries — that's /cite. Flag missing citations; don't fabricate them.
• Running the lit search to populate citations — that's /litreview upstream of this skill.
• Drafting referee responses — that's /referee-response.
• Writing in a non-personal voice (e.g., "draft this in journalistic style") — decline politely and ask whether they want the voice or the style adjusted.
• Producing figures, tables, regressions, or code — those come from the user's R/Python pipeline.
• Committing or pushing — Dropbox handles Overleaf sync.