ts-paper-refine
Stage 4 of the ts-paper suite. One holistic pass that right-sizes each drafted section toward Traitement du Signal JOURNAL length (not conference page-compression), enforces cross-section coherence and terminology consistency, removes redundancy, and preserves all citations, equations, and markers.
Install / Use
npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-refineInstalls into whichever agent you are using.
SKILL.md
Installable skill definition
Quality Score
Category
Development & EngineeringSupported Platforms
Our assessment of ts-paper-refine
ts-paper-refine scores 84/100 on our quality scale, 2917th of 4,700 Development & Engineering skills we index.
Its SKILL.md is 9.8 KB long, well organised into 8 sections and no code examples: a thorough specification that gives an agent plenty to work with.
With 1,126 GitHub stars, it is one of the more widely adopted skills in the catalogue.
Maintenance, license and trust
- The repository was last updated about 3 months ago, so ts-paper-refine is actively maintained.
- Our last check on 2026-11-06 found the source still online.
- It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
- Its trust signals score 100/100, with no cautions. These come from repository metadata, not a code audit β read the skill file before letting an agent act on it.
Safety scan
No issues foundOur scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.
AI review by kimi-k2.7-code on 2026-10-04. Automated pattern scan on 2026-10-04. It catches known dangerous patterns, not every risk β read a skill before letting an agent act on it.
ts-paper-refine compared with similar skills
All 4 of these similar skills score higher than ts-paper-refine; compare them before choosing.
| Skill | Score | Stars | Updated | Format |
|---|---|---|---|---|
| ts-paper-refine (this skill)by Spark-To-Paper-Skills | 84 | 1.1k | 3mo ago | SKILL.md |
| taste-skillby Leonxlnx | 100 | 116.8k | 8d ago | CLAUDE.md |
| ai-job-searchby ai-job-search-dev | 100 | 72.1k | today | CLAUDE.md |
| claude-howtoby luongnv89 | 100 | 47.3k | 2d ago | CLAUDE.md |
| lobehubby lobehub | 100 | 85.6k | today | MCP Server |
Frequently asked questions
- How do I install ts-paper-refine?
- Run
npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-refine. The install tabs above show the steps for each supported agent. - Which AI agents does ts-paper-refine work with?
- It is written for Universal, as a SKILL.md file. Other agents that read the same format can often use it too.
- Is ts-paper-refine safe to use?
- Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 100/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
- Is ts-paper-refine still maintained?
- The repository was last updated about 3 months ago, so ts-paper-refine is actively maintained.
Skill content
View source on GitHubname: ts-paper-refine description: > Stage 4 of the ts-paper suite. One holistic pass that right-sizes each drafted section toward Traitement du Signal JOURNAL length (not conference page-compression), enforces cross-section coherence and terminology consistency, removes redundancy, and preserves all citations, equations, and markers. Use to polish/tighten a TS paper draft before the review stage. Do NOT over-compress.
ts-paper-refine β right-size & polish (one pass)
Do this as one reasoning pass over the whole draft (you see every section at once, so coherence and de-duplication are free). The original's purpose was conference page-compression β that is wrong for a journal. Traitement du Signal papers are substantial (β8β14 two-column pages); right-size, do not aggressively cut.
What to enforce
-
Length toward the template's targets β the table below is the
ts_iietadefault; for another template, right-size to itstemplate.jsonsections[].words(e.g. NeurIPS bands differ). The spec is authoritative (draft_lintreads bands fromtemplate.json). Expand thin sections, trim only genuine bloat β never compress below the floor: | section (ts_iieta) | final words | |---|---| | method | 2000β3000 | | experiments | 1000β1500 | | introduction | 800β1200 | | related_work | 800β1200 (3 themes) | | analysis | 900β1400 | | conclusion | 200β280 (1 paragraph) | | abstract | 150β220 (1 paragraph) |These bands target a substantial ~10β12 page two-column TS article. Expand by adding real substance (more method detail/derivation, more components, deeper analysis, more ablation discussion, fuller related-work lineage) β never by padding, restating, or filler sentences.
-
Coherence: the same method/component/dataset/baseline names everywhere (use the blueprint
terminology); no contradictions between sections; intro/abstract accurately summarize what Method/Experiments actually say. -
No redundancy: each mechanism explained once (Method), not re-derived in Analysis; remove repeated motivation across intro/related/method.
-
Preserve exactly: every
\cite{}, every equation, every\ref{}, every table/figure label and the--result placeholders. Refinement changes prose, never citations/structure/math semantics. -
Honesty intact: still zero fabricated numbers in any sentence; results stay forward-looking; tables stay blank.
-
TS tone: no bold in body, hedged conclusion language, no hype, no filler phrases.
De-AI pass (this is AI-written prose β scrub the tells)
Every section was drafted by an LLM, so it is prone to AI tells. Read each section and rewrite where it improves (don't churn natural sentences). Scan for and replace:
- Prose flow (the #1 fix). Turn any bare-comma comma soup into connected prose: punctuate an appositive/enumeration with a colon or em-dash, not a bare comma β "a pipeline: retrieval, planning" or "a pipeline β retrieval, planning β in which β¦", NOT "a pipeline, retrieval, planning, β¦ in which β¦" (that comma soup is what reads as fragmented "η’ε₯ε"). A legitimate appositive em-dash is GOOD and encouraged; only remove em-dash abuse (dramatic-pause dashes). Vary sentence length; each paragraph must read as a flowing whole, not a clause list. Avoid the recipe rhythm of "X does A [cite], Y does B [cite]" in related work β synthesize across papers.
- Tell words/phrases: leverage, delve, utilize, showcase, underscore, intricate, pivotal, seamless, holistic, nuanced, realm, tapestry, testament, "it is worth noting", "plays a crucial role", "in order to" (β "to"), rule-of-three triplets, connective stacks (firstly/moreover/furthermore β a single appropriate transition is fine), empty "-ing" wind-ups, vague attribution ("studies show"), "not only β¦ but also" overuse.
- Translationese (these drafts trend that way): stacked attributive chains ("the β¦ of the β¦ of the
β¦"), gratuitous passive voice, hollow rhetoric (paradigm shift, disruptive, profound, in essence).
Prefer plain, precise words; keep
\cite/\ref/math intact; add no new emphasis.draft_lintcode-enforces (hard-fails,ai_tell) ONLY the CONTEXT-FREE canned phrases β tapestry, testament, "plays a crucial role", "it is worth noting", "delve into", "realm of", "paradigm shift", "navigating the landscape", "ever-evolving", β¦ plusin order toβ those MUST be gone. The CONTEXT-DEPENDENT cohesion tells β em-dashes, sentence-initialfirstly/moreover/furthermore/additionally,,not only β¦ but also, rule-of-three triplets,leverage/utilizeβ are deliberately NOT gated (a regex would kill the good appositive em-dash along with the abuse β precisely what produced the comma soup): you remove the abuse by judgment while keeping legitimate uses. (Source: PaperJury writing-toolkitde-ai.)
Logic self-check (after each edit, narrow self-gate β not a reviewer)
After you change a passage, re-read ONLY that passage for a show-stopper introduced by the edit: a logic contradiction, an undeclared terminology switch (a term silently renamed), or severe Chinglish. Fix it; do not raise style nits here (the de-AI pass owns those). This is the in-pass honesty check; the adversarial peer-review (ts-paper-review, stage 5, runs by default) is the heavier, separate hardening loop that runs after this stage and routes its fix-now issues back here in Review-fix mode.
DATA-AWARE branch (only when template.results_mode == "data_aware")
When the paper reports real results (data-aware β see ts-paper-data), the "remove unbacked
numbers" instinct inverts for numbers that trace to results.facts.json:
- Do NOT convert real numbers to forward-looking language; do NOT remove real numbers from running text; do NOT flag real (traceable) numbers as hallucinations.
- Verify cross-section number consistency β the same metric formatted identically everywhere; the abstract/conclusion headline numbers equal the filled result-table values.
- Preserve the filled result tables and the evidence sentences; preserve ablation deltas; enforce
definitive past tense; apply the TBD silence rule (no claim about a
"TBD"metric). - Keep every number consistent with
results.facts.json(the number-audit re-runs after refine). Everything else (length right-sizing, coherence, de-duplication, tone) is as in proposal mode.
Match the template's venue type
If the active template is a journal (e.g. ts_iieta), strip any leaked conference framing ("to fit
the page limit / 8-page / camera-ready") and don't over-compress β journals are substantial. If the
template is a conference style (e.g. neurips), respect its tighter conventions instead. Let the
template's word bands β not a fixed venue habit β decide length.
After refining (enforced)
Re-run python ../ts-paper-write/scripts/draft_lint.py <workdir> β it now fails the build on any section outside its template.json word band and on any broken shape contract, so "right-sized" is verified in code, not by eye. Drive every word_band violation to zero by expanding thin sections with real substance (more method detail, more discussion, a deeper failure analysis) β never by padding. Then re-run python ../ts-paper-cite/scripts/citations_lint.py <workdir> to confirm citations survived. Re-run both until they exit 0 (ok:true); do not hand off on a nonzero exit. Then run python ../ts-paper-write/scripts/reflow_tex.py <workdir> so every sections/*.tex is one logical line per paragraph (your de-AI rewrites may have re-wrapped them; idempotent, PDF-neutral). Then write logs/4_refine.io.md (INPUT: pre-refine word counts; DECISIONS: per-section deltas + what was added/cut; OUTPUT: post-refine counts, all in band). Hand off to ts-paper-review (stage 5, the default next stage): it returns a triaged issue list, and for each fix-now issue the orchestrator re-invokes THIS skill in Review-fix mode (below). Only after review is dry/closed does the chain proceed to ts-paper-figure then ts-paper-latex. If the user explicitly opted out of review, hand directly to ts-paper-figure.
Review-fix mode (ONLY when invoked with a single review issue β NOT a full pass)
This mode is triggered only when the orchestrator hands you one ts-paper-review issue
{section, severity, evidence_quote, close_criterion} to close. It is a separate entry path: do not
run it on a normal stage-4 invocation, and do not run the holistic right-size / de-AI pass here. The
default full-pass behavior above is unchanged.
- Make the minimal edit to the named section only that makes that one
close_criteriontrue. Bind strictly to that criterion; do not touch other sections and do not over-rewrite. - Preserve exactly: every
\cite{}, equation,\ref{}, table/figure label, the--placeholders, and the section'stemplate.jsonword band (stay in band β no compression below the floor). - If the fix touches an equation, it must keep notation complete: any symbol the edit introduces into
an equation has to be added to the notation table (or the fix renamed to an already-defined symbol) in
the same edit β
draft_lintnow hard-failsnotation_incomplete(this is exactly the path that let a stray\psiinto the central equation), so an unmatched new symbol will block the re-gate below. - Re-run
python ../ts-paper-write/scripts/draft_lint.py <workdir>ANDpython ../ts-paper-cite/scripts/citations_lint.py <workdir>until both exit 0 (ok:true) to confirm nothing broke (includingnotation_incomplete); do not return a fix on a nonzero exit. - Append the fix (issue id, section, what changed, close_criterion met) to
logs/5_review.io.md(the review log β not4_refine.io.md). - Return control to the orchestrator for the next issue, or to advance to figure once review is dry.
Related Skills
taste-skill
116.8kAgent skills that give coding agents taste: art direction, reference images and a browser check instead of generic AI pages
ai-job-search
72.1kThe job search that runs on your machine. AI job application framework built on Claude Code: evaluate postings, tailor CVs, write cover letters, prep interviews. Fork it and own it.
claude-howto
47.3kA visual, example-driven guide to Claude Code β from basic concepts to advanced agents, with copy-paste templates that bring immediate value.
lobehub
85.6kπ€― LobeHub - Super-Agents on Your Team. Add super-agents to the chats your team already uses.They take on work by themselves and get better every week.
Languages
Trust signals
From repository metadata: license, adoption, age and documentation. Not a code audit β see the Safety scan above for what the skill file itself contains.
