Refactor structure

Refactor Structure

Fix the architecture of the argument, not its words. A paper can be grammatical, on-template, and within the page limit and still fail because the story is assembled in the wrong order, a section doesn't do its job, the introduction promises things the body never delivers, or the conclusion outruns the evidence. This skill diagnoses those problems and produces a restructuring plan — move / merge / split / cut, with the reasoning for each — that the author approves before a single section is rewritten.

It plans and (on approval) executes the moves. It does not reword sentences (that is polish-prose) and it does not fit the paper to a venue's rules (that is tailor-to-venue).

When to use

• "The structure feels off" / "it doesn't flow" / "reorganize my paper".
• Reviewers said: hard to follow, disorganized, the story is unclear, the contribution is buried, related work is in the wrong place, a section doesn't belong.
• The intro over-promises (claims a proof / user study / ablation the body doesn't contain), or the conclusion claims something never evidenced.
• Redundant sections (two Background sections), or one section doing two jobs (Results carrying Threats-to-Validity).

Not for: sentence wording / de-AI-ifying (polish-prose), page limits / templates / track fit (tailor-to-venue), the Related Work section's content (draft-related-work), or whether the citations are real (verify-citations).

Inputs

• The draft: main .tex file (the script follows \input/\include).
• Optional: the paper's intended contribution list and target paper type, from .paper-memory/profile.yml if present (a theory paper and an empirical paper have different expected arcs). If absent, ask the author for the one-sentence claim and the contribution bullets.

Process

1. Read the structure, don't re-read the paper

Run the bundled extractor to get the skeleton on one screen:

python3 skills/refactor-structure/scripts/outline_extract.py <main.tex>

It prints the section/subsection tree, each unit's first sentence (the topic sentence — where a well-built section announces its job), and per-unit signals (word count, citations, whether it carries floats/equations/lists, units that open with no topic sentence). Use --md to drop the skeleton into the plan, --json to drive your own analysis, --no-inputs to run on a single section file. Exit 2 on bad input. This is a structure map, not a verdict — it tells you what is where; the diagnosis below is the judgment.

2. Establish the intended arc

Before judging order, know what the paper is trying to argue. Read the contribution bullets and the abstract's claim. Write down, in one line each: the problem, the claimed contribution(s), the evidence that should back each contribution, and the conclusion each contribution licenses. This is the spine every later check measures against.

3. Diagnose against the architecture checklist

Work through references/architecture-checklist.md, which covers the seven failure modes (wrong order, a section not doing its job, gaps, redundancy, misplaced content, intro/body mismatch, a broken contribution -> evidence -> conclusion arc) and the standard section arcs by paper type. The two highest-value, most objective checks:

• Promise-vs-delivery. For every contribution bullet and every "we show / we prove / we evaluate" in the intro, find the section that delivers it. A promised proof, user study, or ablation with no home section is a gap the reader (and reviewer) will hold against the paper. Conversely, list any major result section the intro never advertised.
• Conclusion-vs-evidence. Every claim in the conclusion must trace back to a result in the body. A conclusion that introduces a new claim (e.g. "scales to any number of nodes" with no scaling experiment) is an arc break — flag it as either "remove the claim" or "the body is missing the experiment that would support it."

See references/diagnosis-patterns.md for how each symptom maps to a concrete remedy (move / merge / split / cut) and the ripple effects to check (forward references, "as we show in Section~X" pointers, a cut section's surviving citations).

4. Write the restructuring plan (the author approves it before any edit)

Produce, do not apply. Write paper-workspace/writing/restructuring-plan.md with:

1. Current skeleton — the --md outline from step 1.
2. Intended arc — the spine from step 2.
3. Diagnosis — findings grouped by failure mode, each with: what's wrong, why it hurts the argument, severity (blocks comprehension / weakens it / minor), and the proposed remedy.
4. Proposed skeleton — the section tree after the plan, so the author sees the new order at a glance.
5. Ordered move list — each move/merge/split/cut as a discrete, reversible step with its ripple effects, sequenced so the document stays coherent between steps.

Present the diagnosis with the cheaper path called out: often a reorder plus two merges fixes more than a rewrite. Where a symptom has more than one reading — is the user study missing, or just unlabeled? — show both and ask; never silently assume a section is absent when it might live in an \input file.

5. Execute on approval, one step at a time

Only after the author signs off, apply the moves in order. After each structural change, re-run outline_extract.py to confirm the skeleton matches the proposed one, and fix the ripple effects the plan listed (cross-references, "see Section~X" pointers, a moved section's label). Moving prose is mechanical; do not silently reword it — flag any sentence that needs new connective tissue for the author (or hand it to polish-prose). A reference orphaned by a cut (a now-dangling \cite) routes through verify-citations; a claim left without evidence — e.g. the conclusion-vs-evidence break in step 3 — routes through verify-claims for a proper claim→evidence audit. Never invent a citation or a result to patch either.

Output

• A reviewable paper-workspace/writing/restructuring-plan.md: current skeleton -> intended arc -> diagnosis -> proposed skeleton -> ordered moves.
• On approval, the applied moves with a re-extracted skeleton confirming the result and a ripple-effect checklist (cross-refs, pointers, labels).
• The author decides every move; nothing is rewritten at sentence level here.

Adapt to your discipline

The failure modes are field-agnostic; the expected arc is not. references/architecture-checklist.md gives CS-paper arcs (systems / ML-empirical / theory). Fork it for your field: IMRaD for experimental science, a different placement of Related Work, a Methods-before-Results vs. interleaved convention, or essay-style argument structure in the humanities.

Guardrails

• Plan before you touch the paper. Diagnosis and the move list come first; the author approves before any section is moved, merged, split, or cut.
• It restructures; it does not reword. Sentence-level edits are polish-prose; venue/track/page fit is tailor-to-venue. Do not silently rewrite moved prose — flag seams for the author.
• Never fabricate a result, section, or citation to "fill a gap." A missing user study is reported as missing; the fix is to add the work or drop the promise, never to invent it.
• Never delete content the author may want; a cut is a proposal, and cut text is preserved (note where it went) so it is recoverable.
• Diagnosis is from reading the structure, not from a venue rule or memory; the extractor's output and the draft are ground truth. It never claims the restructure will get the paper accepted.
• Never submit anything anywhere on the author's behalf.

Relationship to other skills

• tailor-to-venue — venue/track fit, page budget, template, blind level. Run after the argument is sound; structure-then-fit, not the reverse.
• polish-prose — sentence-level wording, hedging, AI-tells. Run after the structure is settled, so you polish prose you're keeping.
• draft-related-work — the content of the Related Work section; this skill only decides where that section sits and whether it does its job.
• verify-claims — when a conclusion claim outruns the evidence (an arc break), this skill flags the structural break; verify-claims audits whether the claim is actually supported. verify-citations is the separate check that a .bib entry resolves (e.g. a reference orphaned by a cut).
• simulate-reviewers / assess-paper — if a mock review flagged "hard to follow", this skill is the fix.

Memory

Uses the shared .paper-memory/ convention in the user's paper directory (full spec: paper-memory-convention.md).

• At start: read profile.yml (paper type sets the expected arc) and lessons.md — lead with any recurring structural habit already recorded (e.g. "you habitually bury Threats to Validity in Results") so you check for it first.
• At end: append durable findings via reflect-and-improve's reflect_log.py append in the shared format - [YYYY-MM-DD] (refactor-structure | <scope>) pattern -> recommendation. A habit seen across drafts is recurring; a one-off is this-paper. Don't log routine single moves.
• Create .paper-memory/ on demand and offer to add it to the project .gitignore; it is local-only and never uploaded or copied into this repo.