Skip to content

write

Category: drafting
Field: —
License: none declared
Updated: 2026-05-11
Stages: paper-drafting

Write

Draft paper sections, apply a cleanup pass, or extract a personal style guide from prior papers by dispatching the Writer agent.

Input: $ARGUMENTS — section name or mode, optionally followed by file path.

Modes

/write [section] — Draft Paper Section

Draft a specific section: intro, strategy, results, conclusion, abstract, or full.

Agent: Writer Output: LaTeX section file in paper/sections/

Workflow:

1. Context Gathering

Before drafting, read all available context: 1. Read existing paper draft in paper/ (if it exists) 2. Read master_supporting_docs/ for notes, outlines, research specs 3. Read most recent quality_reports/research_spec_*.md or quality_reports/lit_review_*.md 4. Read .claude/references/domain-profile.md for field conventions 5. Check Bibliography_base.bib for available citations 6. Scan paper/tables/ and paper/figures/ for generated output 7. Read quality_reports/results_summary.md if it exists (from Coder)

2. Paper Type Detection

Before routing, identify the paper type from the strategy memo or existing draft: - Reduced-form — DiD, IV, RDD, event study - Structural — Model estimation, counterfactual simulations - Theory + empirics — Propositions tested with data - Descriptive / measurement — New data, new measure, stylized facts

This determines which section templates the Writer uses.

3. Section Routing

Based on $ARGUMENTS: - full: Draft all sections in sequence, pausing between major sections for user feedback - intro: Draft introduction (most common request) - strategy: Draft empirical strategy (reduced-form), model + estimation (structural), or model + tests (theory+empirics) - results: Draft results — narration style depends on paper type and output type (regression tables, event study figures, counterfactual simulations, etc.) - conclusion: Draft conclusion with type-appropriate ending (policy implications, counterfactual implications, or research agenda) - abstract: Draft abstract (must have other sections first) - data: Draft data section — expanded for descriptive/measurement papers - model: Draft model section (structural or theory+empirics papers only) - No argument: Ask user which section to draft

4. Dispatch Writer

Dispatch Writer with paper type and argument-move templates for the target section. The writer drafts using paragraph types (motivation, result statement, mechanism, etc.), applies design-specific moves, then runs the cleanup pass. Save to paper/sections/[section].tex.

5. Quality Self-Check

Before presenting the draft: - [ ] Paper type identified and correct template used - [ ] Every paragraph has an identifiable purpose (argument move type) - [ ] Findings lead sentences — not buried after setup - [ ] Design-specific elements present (see writer.md for checklists per design) - [ ] Every displayed equation is numbered (\label{eq:...}) - [ ] All \cite{} keys exist in Bibliography_base.bib - [ ] Introduction contribution paragraph names specific papers - [ ] Effect sizes stated with units - [ ] No banned hedging phrases - [ ] Notation consistent throughout - [ ] All tables/figures referenced actually exist in paper/tables/ or paper/figures/ - [ ] Results narrated correctly for output type (tables, event study figures, counterfactuals) - [ ] Personal style guide loaded (not template) — or user prompted to run /write style-guide - [ ] Claim-source map produced for all numerical claims (quality_reports/claim_source_map_{project}.md) - [ ] Results/Conclusion only drafted after verifying actual output files exist

6. Present to User

Present sections through drafting gates, pausing for approval at each:

GATE 1: Introduction + Literature positioning → present, wait for approval GATE 2: Data + Empirical Strategy (or Model) → present, wait for approval GATE 3: Results + Robustness + Conclusion → present, wait for approval

For single-section drafts, present the section directly. For full, use all three gates.

Flag items that need attention: - BLOCKED items: Results/Conclusion cannot be drafted without output files - VERIFY items: Citations that need user confirmation - VOICE items: Style guide not yet extracted (drafting blocked until resolved)

/write style-guide [paper-dir] — Extract Personal Voice

One-shot extraction of the user's writing voice from their published or drafted papers. Produces .claude/references/personal-style-guide.md, which the writer auto-loads on every subsequent invocation.

When to run: - Once at the start of a project, after pointing at a directory of the user's prior papers - After publishing a new paper that shifts voice (re-run to refresh the profile)

Input: $ARGUMENTS — path to a directory containing prior papers (.tex or .pdf). If omitted, defaults to master_supporting_docs/ and scans for .tex/.pdf files.

Agent: Writer (style-extraction mode) Output: .claude/references/personal-style-guide.md

Workflow: 1. Discover corpus. List .tex and .pdf files in the target directory. If fewer than 2 papers found, flag and ask before proceeding (style extraction on a single paper overfits). 2. Sample strategically. For each paper, extract: - The full introduction - The first two paragraphs of each major section - The abstract and conclusion - A random sample of 5–10 results-section paragraphs This keeps context usage bounded while capturing voice variation across sections. 3. Extract patterns. The Writer (in style-extraction mode) produces quantitative and qualitative patterns: - Sentence-length distribution (median, 10th–90th pct) - Passive-voice frequency, first-person-plural frequency, em dash rate - Paragraph opening and closing moves - Section-architecture patterns (how introductions open, how results lead) - Lexicon: words used repeatedly, words demonstrably avoided - Hedging and comparison patterns - Citation conventions (textual vs. parenthetical split; papers-per-claim) - Tone markers and anti-patterns already stripped 4. Write to .claude/references/personal-style-guide.md. Fill every template section with quoted examples from the corpus. Never invent patterns — if a section has no evidence, write "[insufficient corpus evidence]". 5. Present summary. One-paragraph recap of the voice profile: sentence length, passive rate, signature lexicon, distinguishing tone markers. User confirms before the guide takes effect on subsequent /write calls.

Principles for the extraction: - Ground every claim in the corpus. Each pattern must have at least one quoted example. - Extract, don't prescribe. The guide records the author's observed behavior, not what the Writer thinks is good style. - Don't duplicate domain-profile.md. The style guide is about voice; the domain profile is about field conventions. - Don't override working-paper-format invariants. Voice doesn't trump INV-1..21.

/write humanize [file] — Cleanup Pass Only

Strip AI writing patterns from existing text without rewriting content.

Agent: Writer (cleanup mode) Output: Edited file with AI patterns removed

Strips 24 patterns across 4 categories: - Structural: forced narrative arcs, artificial progression - Lexical: "delve, leverage, nuanced, robust" - Rhetorical: rule-of-three, negative parallelisms, em dash overuse - Formatting: excessive bullet points, promotional language

Section Standards

All paper types share the same backbone. Moves diverge by type — see writer.md for full templates.

Section Length Reduced-Form Structural Theory+Empirics Descriptive
Introduction 1000-1500 ...preview → result → contribution ...model preview → counterfactual → contribution ...theory preview → test result → contribution ...data innovation → key fact → contribution
Data 800-1200 Treatment, outcome, controls Moments that identify parameters Standard 1200-1800 (core contribution)
Strategy/Model 800-1500 Design-specific (DiD/IV/RDD/ES) Environment → decisions → equilibrium → estimation Model → propositions → tests N/A (merged into Data)
Results 800-1500 Main spec → robustness → heterogeneity Estimates → model fit → counterfactuals → welfare Prediction-by-prediction evidence Key facts → decompositions → implications
Conclusion 500-700 Policy implications Counterfactual implications + model limitations What model gets right/wrong Research agenda enabled by new data
Abstract 100-150 Question, design, finding with magnitude Question, model, counterfactual finding Question, prediction, test result Question, measurement, key fact

LaTeX Conventions

  • \citet{} for textual citations ("Smith (2024) shows...")
  • \citep{} for parenthetical citations ("...is well documented (Smith, 2024)")
  • booktabs rules (\toprule, \midrule, \bottomrule) — never \hline
  • Notation protocol: Y_{it}, D_{it}, \gamma_i, \delta_t, \varepsilon_{it}

Bundled Resources (Level 3)

Loaded on demand by the writer agent:

Resource Path When
Section templates templates/section-templates.md Always -- defines section structure
Paragraph moves templates/paragraph-moves.md Always -- defines argument types
Cleanup patterns templates/cleanup-patterns.md After drafting -- cleanup pass
Style extraction templates/style-extraction-protocol.md /write style-guide mode
Drafting gates templates/drafting-gates.md Full draft mode
Claim-source map templates/claim-source-map.md After results section
Notation protocol references/notation-protocol.md Strategy + results sections

See also: gotchas.md for known failure points and edge cases.

Principles

  • This is the user's paper, not Claude's. Match their voice and style.
  • Never fabricate results. Use TBD placeholders.
  • Citations must be verifiable. Only cite confirmed papers.
  • Argument moves first, cleanup second. Draft with structure, then strip AI patterns.