| name | claude-latex-paper-skill |
| description | Write and revise a technical report or thesis as a LaTeX document in English, with the discipline a graded or peer-reviewed submission needs — every claim traced to an artifact, no citation from memory, human-sounding prose, and a mechanical verification loop. Use when the user asks to write the paper, technical report, thesis-style writeup, LaTeX report, 写论文, 技术报告, 排版, polish a section, add citations, or reduce AI-sounding prose. |
LaTeX technical-report writing skill
Produce and maintain a LaTeX report or thesis in English, section by section,
with a claim ledger, verified citations, human-sounding prose, and a mechanical
check after every edit. Converse with the user in their language; write the
report in English unless the user sets another target language.
Project metadata (collect once, then never re-ask)
At scaffold time, collect the report's fixed metadata from the user and record
it — never invent it, and never re-ask it later. Typical fields for a title
page and declaration:
| Field | Example |
|---|
| Author | full name (or leave to the user's preference) |
| Institution | university / organisation |
| Faculty / department | as it should appear on the title page |
| Degree / programme | e.g. B.Sc. / M.Sc., or "not applicable" |
| Module / course / venue | module name and credits, or the target venue |
| Supervisor / advisor | name and title, if any |
| Deliverable | what is being submitted (report, software, presentation) |
Confirm these once and treat them as fixed thereafter unless the user changes
them.
Non-negotiable rules
- Claim ledger. Every quantitative or comparative claim must trace to a
project artifact (see references/evidence-map.md).
No artifact → mark
[CLAIM NEEDS EVIDENCE], never polish an unsupported claim.
- Negative results stay. Honestly reported failures — a bug post-mortem, a
result that did not hold up, a hypothesis that failed — are findings, not
embarrassments. Report them with the same care as the positives.
- No citation from memory. Follow
references/citations.md: verify via DOI/CrossRef/
arXiv/Semantic Scholar before any BibTeX entry; unverifiable → explicit
PLACEHOLDER_ key + tell the user.
- Human-sounding English. Apply
references/style.md while drafting, not as a
post-pass. If a passage already reads naturally, leave it alone.
- Compile before judging. Never assess layout from
.tex source; build
the PDF and look at it (references/latex.md).
- Mechanical verification after every section. Run
scripts/verify_paper.py and fix all hard failures before moving on —
see references/verification.md. Trust the
scan, not your impression of what you wrote.
- Context hygiene. Long context is the main hallucination driver in
long-document work: draft one section per sitting with only that section's
evidence artifacts open; re-read the artifact before writing its numbers,
never quote them from conversational memory. If drafting quality degrades or
a session has grown long, hand off to a fresh session (
/handoff).
Workflow
First invocation (no paper/ yet)
- Read the project's source-of-truth artifacts — its results and data files,
experiment logs, design and status notes, README. The report is written
from these, not from memory of the conversation. Build the evidence map
(references/evidence-map.md) from them.
- Propose a section outline (see report skeleton in
references/latex.md) and the evidence each section
will draw on; get the user's verdict before drafting.
- Scaffold
paper/ (main.tex, sections/, refs.bib, a build script) and draft
section by section, citing as you write.
Per-section loop (draft and revision alike)
- Open only the evidence artifacts this section draws on; draft.
- Run
py -3.13 scripts/verify_paper.py paper/ (adjust the path to wherever
the skill lives).
- Fix hard failures (CJK leakage, unresolved markers, cite/bib mismatch),
rerun until clean — two passes is normal; fixes can introduce new errors.
- Compile, scan the log, eyeball changed pages.
- Update
paper/glossary.md with any new term-of-art before the next section
(one term, one rendering — consistency beats variety).
Revision passes
- Section rewrite / polish → load references/style.md;
state which section and what failure mode you are fixing.
- Citation work → load references/citations.md.
- Layout complaints (loose pages, split figures, floats) → load
references/latex.md §Layout, skip the prose rules.
- After every edit batch: recompile, check the log for warnings, eyeball the
changed pages.
Priority order when editing prose
Paper-type architecture → section's job → paragraph logic →
claim/evidence/boundary → sentence polish. Fix the highest broken level first;
a beautiful sentence in a paragraph with no argument is wasted work.
What NOT to do
- Do not invent related work, datasets, or numbers to fill gaps.
- Do not restructure into bullet-heavy "AI slide" prose; the report is flowing
technical prose with tables/figures where the data lives.
- Do not commit anything; leave git operations to the user's verdict.
- Do not re-derive results by rerunning pipelines unless asked — the archived
result files are canonical.