claudeers.
// Claude Skills

icassp-paper-condense

Agent skill (Claude Code / Codex) for condensing, typesetting and fact-checking a LaTeX paper to ICASSP's 4+1 page limit, verified by measurement

// Claude Skills[ cli ][ claude ]#claude#skills◷ MIT$open-sourceupdated 23 days ago
Actively maintained
97/100
last commit 21 days ago
last release none
releases 0
open issues 0

Install with your AI

Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.

Install and set up icassp-paper-condense (claude-skill project) into my current project.
Found on https://claudeers.com/icassp-paper-condense
Repo: https://github.com/WaIdo/icassp-paper-condense
Homepage/docs: —
Detected install method: claude-skill → # copy this skill into .claude/skills/icassp-paper-condense/
Category: skills. Platforms: cli.
Read the repo's README for exact setup and env vars, then install it and wire it into my project.

Claudeers Health Verdict:
active; community-verified: false. Confirm the source before running anything.
// or install directly (claude-skill)
# copy the skill dir into your project:
# .claude/skills/icassp-paper-condense/   (or ~/.claude/skills/icassp-paper-condense/ for all projects)
// or clone
git clone https://github.com/WaIdo/icassp-paper-condense

// compatibility

Platformscli
Operating systems—
AI compatibilityclaude
LicenseMIT
Pricingopen-source
LanguagePython

Get your FREE $2.50 API credits to access TickAtlas financial data ↗

icassp-paper-condense

English · 简体中文

An agent skill for condensing, typesetting and fact-checking a LaTeX paper to ICASSP's 4 + 1 page limit (spconf template) — by changing typesetting only, never deleting or rewording your text, figures or tables, and with every edit verified by measurement rather than by eye.

Works with Claude Code and OpenAI Codex (both read the same SKILL.md format). The measured venue facts come from seven accepted ICASSP 2026 papers; the process and the pitfalls come from doing this on a real submission, including the mistakes.

What it does

  1. Phase 0 — measures the paper (pages, column bottoms, overflow in mm, short last lines, font sizes, citation order) and the venue corpus, writes a gap ledger, and stops.
  2. Phase A — figures, tables and captions only: table font/rule spacing, float grouping and numbering order, in-figure typography, caption de-duplication and scope qualifiers. Stops for the author's confirmation.
  3. Phase B — body text, typesetting only: microtype, \looseness, hyphenation and float placement to buy lines without changing a word; then filling the last page, and a fact-check. Anything that would need words cut is written up as a proposal for you, not applied.
  4. After every edit: rebuild, and a seven-gate check that must pass.

Three red lines:

  • Your words are never edited. Not a paragraph, not a sentence, not "two redundant words" — for text the skill may change only how the same words are set, using microtype, \looseness, hyphenation and float placement. On one 5-page paper, loading microtype alone was worth 32 lines (about a full page) without touching a single word. When layout runs out, the skill reports the paragraphs and the words it would cut, and stops; you decide.
  • Figure and table content is never cut: no dropping metric columns, no removing panels.
  • No over-correction: a claim made weaker but not more accurate is left alone; a factual problem is reported with evidence and a proposed wording, not silently rewritten.

Back up before you start

Everything the skill does is reversible only if there is something to revert to, and page-limited typesetting is exactly the kind of work where the third attempt is worse than the first. Snapshot the whole source, not just the PDF:

git add -A && git commit -m "pre-condensation snapshot"
# not using git? then:
cp -r my-paper my-paper.backup-$(date +%Y%m%d_%H%M%S)

Phase 0 of the workflow does this first and refuses to continue without it. Keep a dated PDF of each stage in backup/ as well, so diff_numbers.py always has a previous build to compare against.

Install

Claude Code

git clone https://github.com/WaIdo/icassp-paper-condense ~/.claude/skills/icassp-paper-condense

or drop the folder into a project's .claude/skills/. Claude Code picks it up from the name / description in SKILL.md.

Codex

git clone https://github.com/WaIdo/icassp-paper-condense ~/.codex/skills/icassp-paper-condense

Dependencies

pip install pymupdf

plus poppler (pdftotext, pdfinfo, pdffonts) and a TeX distribution with latexmk. On macOS: brew install poppler.

Scripts

Run from the paper's directory.

scriptpurpose
scripts/measure_layout.py main.pdffull layout / compliance report (--json available)
scripts/find_runts.py main.pdf --main main.texparagraphs and captions ending in ≤2 words
scripts/check_citations.py --main main.texcite ↔ bib ↔ bbl audit, contexts, IEEE et-al rule
scripts/diff_numbers.py old.pdf new.pdfnumbers lost/gained between two builds
scripts/corpus_baseline.py papers/ICASSP2026/ --also main.pdfmeasure accepted papers, place yours beside them
scripts/verify_all.sh main.tex --backup old.pdfrebuild + every gate; non-zero exit on failure

Layout

SKILL.md                  workflow, red lines, gates, report template
references/
  rules.md                hard vs. soft ICASSP rules; what spconf.sty does
  measurements.md         venue baselines (title block, abstract, tables, refs, punctuation)
  compression.md          levers with measured yields; gap ledger; runt rule; \flushbottom
  author-block.md         add an author / second affiliation at zero body cost
  fact-check.md           which sentences break; evidence order; over-correction
  pitfalls.md             silent failures, each one real
scripts/                  see above

Scope

Written for spconf-based IEEE conferences (ICASSP, ICIP, …) with a hard page limit and a references-only extra page. The measurement scripts work on any two-column PDF; the venue numbers in references/measurements.md are ICASSP 2026 and should be re-measured for other venues with corpus_baseline.py.

License

MIT — see LICENSE. Copyright (c) 2026 WaIdo (github.com/WaIdo).

Every script carries an SPDX-License-Identifier: MIT header and every reference file a one-line footer, so an individual file copied out of the skill still carries its origin and terms. Please keep them.

Not covered by this licence: the ICASSP paper kit and spconf.sty, which are IEEE/ICASSP's; and any papers you measure with corpus_baseline.py. The venue numbers in references/measurements.md are measurements of published papers — factual observations about page geometry, font sizes and punctuation counts.

If you use this in published work, a link to the repository is appreciated but not required.

// faq

What is icassp-paper-condense?

Agent skill (Claude Code / Codex) for condensing, typesetting and fact-checking a LaTeX paper to ICASSP's 4+1 page limit, verified by measurement. It is open-source on GitHub.

Is icassp-paper-condense free to use?

icassp-paper-condense is open-source under the MIT license, so it is free to use.

What category does icassp-paper-condense belong to?

icassp-paper-condense is listed under skills in the Claudeers registry of Claude-compatible tools.

8 views
★ 19 stars
unclaimed
updated 23 days ago

// embed badge

icassp-paper-condense on Claudeers
[![Claudeers](https://claudeers.com/api/badge/icassp-paper-condense.svg)](https://claudeers.com/icassp-paper-condense)

// retro hit counter

icassp-paper-condense hit counter
[![Hits](https://claudeers.com/api/counter/icassp-paper-condense.svg)](https://claudeers.com/icassp-paper-condense)

// reviews

// guestbook

0/500

// related in Claude Skills

🔓

An agentic skills framework & software development methodology that works.

// skillsobra/⟨Shell⟩★ 292,190◷ MIT[ claude ]
🔓

Public repository for Agent Skills

// skillsanthropics/⟨Python⟩★ 178,324[ claude ]
🔓

💫 Toolkit to help you get started with Spec-Driven Development

// skillsgithub/⟨Python⟩★ 138,919◷ MIT[ claude ]
🔓

AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs,…

// skillsGraphify-Labs/⟨Python⟩★ 123,800◷ MIT[ claude ]
→ see how icassp-paper-condense connects across the ecosystem