agentleFS
flonatflonat/flonat-research/skills/latex-polish/SKILL.md

latex-polish

Skill150 starsChanged 11 days agoMITView on GitHub

Inspect a cleanly compiling LaTeX document for source pathologies and rendered visual defects by linting and viewing selected PDF pages. Use when compilation succeeds but title pages, floats, tables, figures, or layout still need publication-quality review. Not for basic compilation health; use $latex-health-check.

  • Installs packages

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

12 rules. About 0 are common, 2 are distinctive, and 10 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 10

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    After major restructuring (new sections, table re-layouts, figure swaps) when you want eyes on the rendered output

    line 16
  2. Very specific:

    When latex reports a clean build but something still "looks weird"

    line 17
  3. Distinctive:

    When preparing for a panel / supervisor read where presentation matters

    line 18
  4. Very specific:

    Verify pdftoppm is on PATH.

    line 51
  5. Very specific:

    Create render dir: mktemp -d /tmp/latex-polish.XXXXXX.

    line 52
  6. Very specific:

    Run the 9 grep-based source-pathology detectors from <skills-root>/latex/references/source-pathologies.md, plus chktex and latexindent -k …

    line 56
  7. Very specific:

    If overfull-box locations are not known (no log file or out/main.log missing), skip those entries.

    line 74
  8. Very specific:

    If --no-vision was passed, skip Phases 3–5 entirely.

    line 113
  9. Very specific:

    Write the polish report to <project>/reviews/<paper-slug>/latex-polish/<YYYY-MM-DD-HHMM>.md (per rules/review-artefact-routing.md), where …

    line 117
  10. Very specific:

    Read references/quality-rubric.md.

    line 152
  11. Very specific:

    If Phase 2 detects a Pattern-1 (spacing-hack) issue, the report includes an Offer: block:

    line 160
  12. Very specific:

    Wait for explicit user approval.

    line 166

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

LaTeX Polish — Source Lint + Vision-On-Rendered-PDF

Heavier-weight visual-quality review than latex. Combines the Phase 4 source-pathology lint with vision-model inspection of selected rendered PDF pages. Designed to catch the failure class that latex Phase 4 source-grep misses: blank pages, clipped tables, illegible shrunken figures, weird title-page line spacing that the regex couldn't fingerprint, bad float placement, overlap.

When to Use

  • Pre-submission visual check on a paper, proposal, thesis, or deck
  • After major restructuring (new sections, table re-layouts, figure swaps) when you want eyes on the rendered output
  • When latex reports a clean build but something still "looks weird"
  • When preparing for a panel / supervisor read where presentation matters

When NOT to Use

NeedUse instead
Compile-fix looplatex
Citation auditlatex (Phase 5) or bib-validate
Prose / academic-argument reviewproofread
Single-figure analysisfigure-feedback
Continuous integrationThis skill renders + vision-inspects pages; too expensive to run on every save

Inputs

  • .tex path — auto-discovered or user-provided. Prefer the user's argument.
  • PDF path — assumed at <project>/<basename>.pdf (per latex convention). If absent or older than .tex, the skill aborts and suggests latex first.
  • --no-vision — skip the vision-rendering phases (Page rendering onward), run the source-pathology lint only.
  • --pages auto|N,N,N — page selection override. Default auto.

Hard Rules

  1. Do not compile. This skill presupposes a clean PDF. If the PDF is missing or stale, abort and suggest latex first.
  2. Read-only with respect to source files. Exception: Pattern-1 (spacing-hack) opt-in fix, which is offered with a diff and only applied on explicit user yes.
  3. No deploy actions. This skill produces a report; it does not commit, push, or modify build infrastructure.
  4. Cap render cost. Max 8 pages rendered per invocation. If page selection would exceed 8, prioritise: title page, worst-overfull page, first table/figure page, then arbitrary sample.
  5. Cap vision-prompt cost. One Read per rendered page. No per-page back-and-forth.

Protocol — 7 Phases

Phase 1: Pre-flight

  1. Resolve .tex path. If a directory was given, find main.tex or single .tex inside.
  2. Resolve PDF path: <dirname>/<basename without .tex>.pdf. If missing or older than .tex, abort with: "Stale or missing PDF — run latex first."
  3. Verify pdftoppm is on PATH. If not, abort with install hint (brew install poppler on macOS).
  4. Create render dir: mktemp -d /tmp/latex-polish.XXXXXX. Track this as $RENDER_DIR.

Phase 2: Source-pathology re-run (Phase 4 of latex)

Run the 9 grep-based source-pathology detectors from <skills-root>/latex/references/source-pathologies.md, plus chktex and latexindent -k if installed. Collect findings into a structured list.

This phase is identical to latex Phase 4 and is rerun here for self-containment (so latex-polish can be invoked without a prior latex run).

Phase 3: Page selection

Determine which pages to vision-inspect. Default auto:

PageWhyMandatory if available
1Title / cover — high-stakes visual real estate; title-page failure mode is commonYes
Page containing any overfull vbox > 50ptVertical overflow is almost always visually brokenYes
Page containing the largest overfull hboxThe "worst" horizontal overflow seenYes
First page containing a tableTables produce the most common visual pathologiesYes
First page containing a figureSameIf figures exist
Last body page before bibliographyCatches widow/orphan and end-of-content layout issuesIf pages > 5
Bibliography page 1Citation formatting + hanging indent regressionsIf bib exists

If overfull-box locations are not known (no log file or out/main.log missing), skip those entries. Cap total at 8 pages; if the selection would exceed 8, drop the lowest-priority entries.

Pull overfull-box page locations by reading <project>/out/<basename>.log:

awk '/Overfull \\(v|h)box/ { for(i=1;i<=NF;i++) if($i ~ /\[[0-9]+\]/) print substr($i,2,length($i)-2) }' out/main.log | sort -u

For each candidate page, also record WHY it was selected (used to pick the right vision-prompt later).

Phase 4: Page rendering

for P in $SELECTED_PAGES; do
    pdftoppm -r 150 -f $P -l $P -png "$PDF" "$RENDER_DIR/page-$P"
done

150 DPI is enough for layout analysis without ballooning file sizes. Each PNG lands at $RENDER_DIR/page-<N>-<P>.png (pdftoppm pads the page number).

Phase 5: Vision analysis

For each rendered page, Read the PNG and apply the per-page-type prompt from references/vision-prompts.md:

  • Title page prompt: focuses on alignment, line-spacing consistency, weight hierarchy, vertical block separation, font-size jumps
  • Body page prompt: focuses on paragraph spacing, hyphenation/justification artefacts, ragged margins, widow/orphan lines, float positioning
  • Table/figure page prompt: focuses on caption position, alignment with surrounding text, legibility at this rendered size, clipping, awkward whitespace
  • Bibliography page prompt: focuses on hanging indent, entry alignment, consistent inter-entry spacing
  • General prompt: catches anything unmatched

Each Read produces a structured finding list per page. Output schema per page:

Page N (selection reason: "title page" | "overfull at L<line>" | …):
  - Finding 1: <one-sentence description> [Tier: Major/Moderate/Minor]
  - Finding 2: …
  Verdict: PASS / NOTES / REVISE

If --no-vision was passed, skip Phases 3–5 entirely.

Phase 6: Report consolidation

Write the polish report to <project>/reviews/<paper-slug>/latex-polish/<YYYY-MM-DD-HHMM>.md (per rules/review-artefact-routing.md), where <paper-slug> is the paper being reviewed (e.g. paper-jtp, paper-philtech):

# LaTeX Polish Report — YYYY-MM-DD HH:MM

**Document:** <path>
**PDF age:** <minutes/hours since last build>
**Pages rendered for vision:** N
**Source detectors run:** 9 grep + chktex + latexindent

## Source pathologies (Phase 2)

<table from Phase 4 spec>

## Per-page vision findings (Phase 5)

### Page 1 — title
<verdict and findings>

### Page <N> — <reason>
<verdict and findings>

...

## Consolidated verdict

PASS / NOTES / REVISE — based on the worst finding tier across both phases.

## Suggested remediations

<prioritised list, max 10 items>

Phase 7: Quality score

Read references/quality-rubric.md. Compute deductions across:

  • Source-pathology findings (mirrors latex rubric)
  • Vision findings (this skill's distinct contribution)

Verdict thresholds per skills/shared/quality-scoring.md. Append the Score Block to the report.

Pattern-1 opt-in auto-fix

If Phase 2 detects a Pattern-1 (spacing-hack) issue, the report includes an **Offer:** block:

**Offer:** Wrap `<env-name>` in `\begin{singlespace}…\end{singlespace}` and delete the negative kerns at L<N1>, L<N2>. Apply? (Y/N)

Wait for explicit user approval. If yes, apply via narrow Edit calls and recompile via latexmk. If no, leave the report as-is. Do not silently apply.

This is the only auto-fix offered. All other findings are report-only.

What latex-polish does NOT do

  • Compile or fix compile errors (use latex first)
  • Modify prose or argument (proofread covers that)
  • Audit citations (latex Phase 5 or bib-validate)
  • Replace figure-feedback for standalone figure-file analysis
  • Auto-commit, auto-push, or modify settings.json / hooks / rules

Cross-references

Skill / FileRelationship
latexCompile + Phase 4 grep lint + citation audit. latex-polish is the deeper sibling for visual-quality review.
figure-feedbackVision pattern this skill reuses for per-page analysis
proofreadProse / argument review; non-overlapping concern
<skills-root>/latex/references/source-pathologies.mdShared detector catalogue
rules/review-artefact-routing.mdWhere the polish report is filed
rules/manuscript-edit-budget.mdWhy this skill is report-only by default

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.