Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Create, revise, and format thesis or dissertation Word documents with strict academic formatting control. Use when an AI agent needs to generate or revise thesis content, normalize Word styles, follow a school template, fix captions or page numbers or section levels, or produce evidence-based Mermaid figures and LaTeX-formatted code listings for a thesis document.
.claude/skills/catlog22-thesis-docx/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 66% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 86% | 0% |
| case-04 | ✓→✗ | ▼ Worse | 63% | 0% |
| case-14 | ✓→✗ | ▼ Worse | 7% | 0% |
Use this skill for thesis-oriented .docx work where content quality and format fidelity both matter. Prefer Microsoft Word desktop automation over WPS-like alternatives whenever the task involves batch formatting, styles, captions, pagination, tables of contents, or cross-references.
This skill is designed to avoid the most common thesis-editing failure mode: the agent makes broad formatting changes too early, introduces new layout problems, and then forces the user to catch them one by one. The default behavior should therefore be:
.workflow/codebase/ARCHITECTURE.md exists, read for project contextmaestro load --type spec --category coding — load coding conventionsmaestro search "thesis academic writing docx formatting" --json — top 5 entries as prior contextpowershell powershell -ExecutionPolicy Bypass -File scripts/check_word_com.ps1 -Json
fidelity for thesis formatting.
existing document, sample pages, and any explicit chapter rules.
standards. Use conservative academic defaults and say that they are defaults.
must enforce: school guide, template, user-confirmed style rulespreserve current state: anything the school guide does not definescripts/audit_docx_ooxml.py before large-scale formatting.text, and style-ID mismatches before deciding on a repair strategy.
abstract, and appendix styles separate and consistent.
direct paragraph formatting or OOXML overrides that will survive the style change.
evidence, codebase, notes, or source material.
diagrams, and similar thesis figures.
the user, such as code, SQL schema, API docs, project docs, or the thesis draft itself.
for the missing source information.
page number scheme, cross-references, hidden paragraph overrides.
you clearly state the remaining manual visual checks.
Body TextHeading 1 / Heading 2 / Heading 3Figure CaptionTable CaptionReferencesAbstractKeywordsAppendix Titletheir font, spacing, indentation, and numbering behavior.
bring the document back under style control.
Examples of high-risk overreach:
state such as firstLineChars, numbering indentation, titlePg, differentFirstPageHeaderFooter, REF field display text, and direct run formatting.
Before bulk editing, produce and internally follow a checklist like this:
Do not say "finished" merely because a structural audit looks good. For thesis work, pagination and page-level rendering are part of correctness.
Read references/failure-patterns-and-quality-gates.md before large-scale formatting work. In particular, guard against:
firstLineChars or numbering indentation creating invisible extra indentsbox title clipping
English references
For school-format-sensitive thesis work, final verification should prefer this order:
If page rendering cannot be verified, say so explicitly instead of implying the formatting is fully validated.
real source material.
references/figure-and-code-rules.md before generating diagrams.listings.
scripts/check_word_com.ps1scripts/audit_docx_ooxml.pyfield behavior before making risky formatting changes.
scripts/normalize_word_styles.ps1captions through Word COM automation.
scripts/export_word_pdf.ps1review.
scripts/render_mermaid_figure.ps1references/paper-format-workflow.mdreferences/figure-and-code-rules.mdreferences/failure-patterns-and-quality-gates.mdchecked.
references/script-usage.mdtext.
from first pages.
punctuation normalization did not damage DOI or English references.
layout fidelity was not guaranteed.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | fail→fail | 12,681 | 10,210 | -19% | 1 | 1 | 0% | 1,804 | 3,764 | +109% | 0 | 0 | — |
case-01 | fail→fail | 6,629 | 5,598 | -16% | 1 | 1 | 0% | 977 | 2,488 | +155% | 0 | 0 | — |
case-02 | fail→fail | 6,994 | 6,932 | -1% | 1 | 1 | 0% | 391 | 2,527 | +546% | 0 | 0 | — |
case-03 | fail→fail | 6,321 | 6,509 | +3% | 1 | 1 | 0% | 341 | 2,598 | +662% | 0 | 0 | — |
case-04 | pass→fail | 12,685 | 8,578 | -32% | 1 | 1 | 0% | 2,212 | 3,611 | +63% | 0 | 0 | — |
case-05 | fail→fail | 15,465 | 9,437 | -39% | 1 | 1 | 0% | 2,392 | 3,715 | +55% | 0 | 0 | — |
case-07 | fail→pass | 14,344 | 8,230 | -43% | 1 | 1 | 0% | 2,116 | 3,502 | +66% | 0 | 0 | — |
case-08 | fail→fail | 11,243 | 7,867 | -30% | 1 | 1 | 0% | 1,676 | 3,369 | +101% | 0 | 0 | — |
case-09 | fail→fail | 10,335 | 11,082 | +7% | 1 | 1 | 0% | 1,729 | 3,918 | +127% | 0 | 0 | — |
case-10 | pass→pass | 11,327 | 7,134 | -37% | 1 | 1 | 0% | 1,638 | 3,166 | +93% | 0 | 0 | — |
case-11 | fail→pass | 12,749 | 4,831 | -62% | 1 | 1 | 0% | 2,320 | 2,986 | +29% | 0 | 0 | — |
case-12 | pass→pass | 12,220 | 9,808 | -20% | 1 | 1 | 0% | 1,717 | 3,690 | +115% | 0 | 0 | — |
case-13 | pass→pass | 16,679 | 10,351 | -38% | 1 | 1 | 0% | 2,637 | 3,854 | +46% | 0 | 0 | — |
case-14 | pass→fail | 17,320 | 5,509 | -68% | 1 | 1 | 0% | 2,822 | 3,025 | +7% | 0 | 0 | — |
case-15 | pass→pass | 11,391 | 8,246 | -28% | 1 | 1 | 0% | 1,788 | 3,413 | +91% | 0 | 0 | — |
case-16 | pass→pass | 10,151 | 10,589 | +4% | 1 | 1 | 0% | 1,573 | 3,790 | +141% | 0 | 0 | — |
case-17 | pass→pass | 14,899 | 10,722 | -28% | 1 | 1 | 0% | 2,235 | 3,862 | +73% | 0 | 0 | — |
case-18 | fail→pass | 11,954 | 6,534 | -45% | 1 | 1 | 0% | 1,684 | 3,138 | +86% | 0 | 0 | — |
case-19 | pass→pass | 15,342 | 7,534 | -51% | 1 | 1 | 0% | 1,995 | 3,279 | +64% | 0 | 0 | — |
case-20 | pass→fail | 14,443 | 6,083 | -58% | 1 | 1 | 0% | 2,580 | 2,950 | +14% | 0 | 0 | — |
case-21 | fail→fail | 9,668 | 6,100 | -37% | 1 | 1 | 0% | 1,483 | 3,150 | +112% | 0 | 0 | — |
case-22 | pass→pass | 8,328 | 8,006 | -4% | 1 | 1 | 0% | 1,315 | 3,495 | +166% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of 0 percentage points is the difference between those two pass rates over the 19 comparable cases. 4 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.