/review-article-architecture

name: review-article-architecture
description: Use when writing, restructuring, or auditing a Review, survey, or Perspective rather than a research article, especially a commissioned one written across many sessions. Covers the governing plan document, drift detection, single-definition thesis and field judgements, section-owns-figure allocation, display-item budgets, four-state status markers, and authorship-blank discipline. For research articles with Results and Methods, use manuscript-optimizer instead.

Review Article Architecture

Overview

A Review fails differently from a research article. A research article drifts by overclaiming past its own data. A Review drifts by **becoming a different Review**: the argument that felt strongest in session 20 quietly replaces the argument the piece was commissioned to make, and because each individual edit was an improvement, nothing looks wrong until someone reads it against the brief.

That is not hypothetical. One commissioned Review in this lineage drifted over several sessions into a metrology argument about cross-sensor interoperability. It was internally consistent, well sourced, and would have been a reasonable paper. It was the wrong paper, it violated four explicit constraints of its own brief, and it had to be reverted wholesale.

This skill is the machinery that makes that drift detectable and mostly preventable.

When To Use

Use this skill when:

• writing or revising a commissioned, invited, or submitted Review, survey, or

Perspective

• a Review is being written across many sessions, by several people, or by agents
• the piece has an explicit brief, synopsis, or commissioning plan
• structure, section ownership, terminology, or display-item count is unstable
• an outline exists and the question is whether the draft still matches it

Do not use this skill for:

• research articles with Results and Methods. Use `manuscript-optimizer`.
• venue selection among Nature journals. Use `nature-portfolio-playbook`.
• sentence-level prose. Use `scientific-prose-style` after this is stable.
• finding, summarizing, or comparing the literature itself. Use

`academic-researcher` (optional set, installed with `--set all`); this skill governs the article's structure, not source discovery.

• deciding what each figure argues. Use `figure-planner`; this skill decides

which section owns it and how many there may be.

1. The Governing Document Is Authoritative, Not Advisory

Designate exactly one file as the brief, and give it authority in writing:

• it fixes the body sections and their order
• it fixes the display items and their count
• it fixes the terminology
• it records the adjudications that resolved earlier disagreements

Then the rule, stated in the manuscript README so nobody has to infer it:

> Do not reorder, merge, or add a section, a figure, or a term without amending > this file first. **If the plan and a better idea conflict, raise the conflict. > Do not resolve it by editing the manuscript.**

This is the load-bearing rule of the whole skill. A Review written over months accumulates dozens of locally-good decisions, and without an authority to check them against, their sum is a different paper. The plan is not there because it is smarter than you. It is there because it does not change while you are working.

**Record adjudications, not just requirements.** When the plan settles a question, write the settlement into the plan. In the Review above, an early outline had a comparison Table; the final adjudication merged its content into Figure 1 and fixed the count at six figures and no Table. Without that written down, the Table regrows: someone reads the old synopsis, notices a Table is missing, and helpfully adds it back.

**Point the raw materials at the plan.** A commissioning synopsis, a slide deck, and an early figure draft all contain figure numbering and item plans that the plan has since superseded. Say so explicitly next to each raw material, or the next reader will follow the wrong one.

2. Drift Audit

Run this whenever a Review has been drafted across multiple sessions, and always before a compression or polish pass. Polishing a drifted draft makes it harder to detect, not easier.

Four symptoms identified the real case:

1. **The draft is a genre the plan prohibits.** The plan forbade writing a pure device review; the drifted version was one. Read the plan's prohibitions, not just its requirements, and check the draft against each. 2. **A required section is missing.** The plan required a market section. The drifted version had dropped it entirely, and nobody noticed because the remaining sections were good. 3. **The designated centrepiece is flattened.** The plan named one section as the heart of the Review, carrying the author team's accumulated work. The drifted version had it at the same weight as everything else. Check section word counts against the plan's intended emphasis, and require an explanation for any inversion. 4. **The thesis and terminology are different.** Not weaker, not stronger: different. Compare the current thesis sentence to the plan's, word by word.

A fifth check worth adding: **does the draft still answer the commissioned question?** Write the question in one line, then read the abstract. If the abstract answers a neighbouring question, that is drift regardless of quality.

When drift is found, the recovery is not incremental. See `references/drift-audit.md` for what to carry forward from a reverted version and what must not be carried forward.

3. The Thesis Is A Macro, Not Prose

Define the Review's central claim **once**, as a macro, and never retype it:

\newcommand{\thesisline}{...}   % the Review's claim, defined here and nowhere else
\newcommand{\claimA}{...}       % field judgement, owned by section 2 and Figure 1
\newcommand{\claimB}{...}

Call the macro in exactly the places the thesis belongs: the abstract, and