Write Survey

Produce or maintain survey/survey.tex: the single, complete, detailed reading reference for the whole SOTA. After reading the survey, nobody should need to reopen syntheses or PDFs for the large majority of questions.

Read First

• survey/README.md — the survey contract
• sota/index.md — what the SOTA currently contains
• survey/coverage.md — what the survey currently integrates
• references/survey-content.md — content checklist and style

The survey renders citations author-year (biblatex style=authoryear, per survey/README.md) — a discursive review reads better than numbered [1].

Mode Selection

• survey/coverage.md lists no citekeys → create mode.
• Otherwise → update mode. The user can force a full rewrite, which is

create mode.

Create Mode

1. Hard gate: read every synthesis. List all digested citekeys from

sota/index.md, read every sota/papers/<citekey>/synthesis.md, and confirm the count matches before writing a single word of survey prose.

1. Design the grouping from what was read: themes, concepts, methodologies

— whatever organization fits this SOTA best. Write the outline as LaTeX section comments first.

1. Write section by section: per group, discuss each paper's contributions

and notable aspects in depth, compare approaches, name tensions and contradictions. Respect each synthesis's "Safe claims / do-not-claim" section. Cite with \cite{<citekey>} only; every key resolves in the root references.bib.

1. Write the mandatory final content section, **Gaps and Research

Directions**, grounded in the full picture per references/survey-content.md.

1. Fill survey/coverage.md with every integrated citekey.
2. Build (make survey) and validate (make check). Fix all errors.

Update Mode

1. Diff: to_add = digested keys in sota/index.md missing from

coverage.md; to_remove = covered keys now excluded or gone.

1. Read the syntheses of to_add and the survey sections they belong to.
2. Integrate additions into the right groups — extend or restructure a

group when the new papers change its internal logic; never just append.

1. Excise removals: delete their discussion, repair comparisons and

transitions that referenced them, re-check the gaps section.

1. Update coverage.md, rebuild, run make check.

Rules

• Every statement about a paper is supported by its synthesis (or its PDF

when more depth is needed) — never by model memory.

• The survey is self-contained: detailed enough to replace re-reading the

SOTA for most purposes. No length limit; completeness wins.

• Never edit references.bib from this skill: a missing entry means the

SOTA work is incomplete — stop and report.

Done When

• coverage.md equals the set of digested citekeys (or the user-approved

subset), survey.pdf is rebuilt, make check passes