ts-paper-review
Adversarial peer-review HARDENING pass for a ts-paper draft â the thing a forward-only drafter/refiner can't do: argue the other side of the paper's claims.
Install / Use
npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-reviewInstalls into whichever agent you are using.
SKILL.md
Installable skill definition
Quality Score
Category
Development & EngineeringSupported Platforms
Our assessment of ts-paper-review
ts-paper-review scores 84/100 on our quality scale, 2918th of 4,700 Development & Engineering skills we index.
Its SKILL.md is 12 KB long, split into 7 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-review 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 â 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 â read a skill before letting an agent act on it.
ts-paper-review compared with similar skills
All 4 of these similar skills score higher than ts-paper-review; compare them before choosing.
| Skill | Score | Stars | Updated | Format |
|---|---|---|---|---|
| ts-paper-review (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-review?
- Run
npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-review. The install tabs above show the steps for each supported agent. - Which AI agents does ts-paper-review work with?
- It is written for Claude Code and Zed, as a SKILL.md file. Other agents that read the same format can often use it too.
- Is ts-paper-review 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-review still maintained?
- The repository was last updated about 3 months ago, so ts-paper-review is actively maintained.
Skill content
View source on GitHubname: ts-paper-review description: > Adversarial peer-review HARDENING pass for a ts-paper draft â the thing a forward-only drafter/refiner can't do: argue the other side of the paper's claims. Distilled from PaperJury into a lean Claude-native engine: N isolated domain reviewers read the whole paper (verbatim-quote anti-skim) â merge/dedupe â perspective-diverse skeptics try to refute each issue â loop-until-dry â a prioritized, verified issue list, which is fixed through the refine stage (each fix bound to a close_criterion) and re-linted. Run after refine (before figure/latex), or whenever the user asks to "review / critique / 审皿 / harden" the paper. Adapts to proposal vs data-aware mode. Stage 5 of the suite: runs by default (engine-agnostic â via the Workflow tool, parallel subagents, or fully in-context; identical algorithm and output, so it is NEVER skipped merely because a Workflow tool is absent); skip only on explicit user request for a quick/no-review draft. Cost-tiered (it spends extra passes), so tune how heavy it is â not whether it runs.
ts-paper-review â adversarial peer-review (harden the draft before finalizing)
Stage 5 of the ts-paper suite â runs by default after refine, before figure.
The rest of the suite is forward-only: it drafts, then polishes. Nothing argues against the paper the
way a reviewer will. This stage does â a lean version of PaperJury's courtroom engine. It runs by
default as stage 5; skip only on explicit user request for a quick/no-review draft, recorded in
logs/0_route.io.md. It is cost-tiered (it costs extra passes), so tune how heavy it runs, not whether.
When to use / not
- Use by default after
ts-paper-refine(beforets-paper-figure/ts-paper-latex) as the pre-finalization hardening pass, or whenever the user says review / critique / 审皿 / è¯å®¡ / harden / "pick it apart". - Skip only on explicit user request for a quick/no-review draft (record the opt-out in
logs/0_route.io.md), or when the user just wants a single asked-for edit (that is the refine / write stage). This is not an official-venue rebuttal and not a from-scratch drafter.
THE REVIEW ALGORITHM (one spec, three execution tiers)
The algorithm is engine-independent â it runs identically on any of the three execution tiers in "How to run it". Describe it once; the tier changes only HOW issues are produced, never the algorithm or the output. It implements, in order:
- N isolated reviewers (default 3 full-surface lenses â Theory/Foundations, Empirical/Benchmark,
Applied/Systems â each with the "unflinching academic gatekeeper" core + a two-pass critique). Each
reads the WHOLE paper; isolation = the paper text is the only thing in its prompt (no peers, no
prior passes). Each weakness MUST carry an exact verbatim quote (can't quote = didn't read =
anti-skim) + severity + a
close_criterion. - Merge â semantic dedupe within the pass and against everything already seen: collapse same-issue
rows (same section anchor AND overlapping summary; when unsure, keep separate) into one whose
raised_bylists every source, carrying itsevidence_quotethrough unchanged; and drop any candidate that lacks a usableclose_criterionintodropped_no_criterion(it is not actionable). - Adversarial verify â each new issue faces 3 perspective-diverse skeptics (misreading /
already-addressed / scope-or-severity); it survives unless a MAJORITY refute it â i.e. with 3
angles it is dropped on â¥2 refutes (
refutedCount < ceil(N_angles/2)survives). Bias to keep â the human gate catches residual noise. This filters plausible-but-wrong issues. - loop-until-dry â a pass counts as dry when it yields no NEW issues OR when every new
candidate is refuted (zero survivors); any surviving issue resets the dry counter to 0. Re-run fresh
panels until
dryStopconsecutive dry passes, capped bymaxRounds(budget-aware).
The issue object (identical across all tiers): { severity: blocker|major|minor|nit, section, summary, evidence_quote (EXACT verbatim â "cannot quote â do not file"), close_criterion, raised_by[] };
after the loop each is sorted by severity (blocker<major<minor<nit) and assigned a final id I-NN.
Returns { issues, refuted, dropped_no_criterion, rounds_run, per_round }.
Invariant: the chosen execution tier changes only HOW issues are produced, never WHICH algorithm or
WHAT output. Every tier emits this exact structure and the same logs/5_review.io.md.
Mode adaptation (passed as resultsMode)
- proposal: reviewers are told the result tables are intentionally blank and prose is forward-looking â so they do NOT file "results missing" as a flaw; they judge method soundness/completeness, novelty/ positioning, clarity, whether the evaluation PLAN would validate the claims, internal consistency, unsupported claims.
- data_aware: full claims-vs-evidence scrutiny (numbers support claims, baseline fairness, ablation coverage, over-claiming, abstract numbers match tables).
How to run it
-
Assemble the paper text. Concatenate
sections/*.tex(+ abstract) into one readable blob (strip pure-LaTeX noise; keep section headings + prose + table/figure captions). This ispaperText. -
Select the execution tier, then run the algorithm on it. Look at the tools you actually have and use the highest tier whose tool is available; Tier 3 is the universal floor and always works. The knobs are the same algorithm params on every tier: lean default
maxRounds:2, dryStop:1, verify:true; cheapestmaxRounds:1, verify:false; thoroughmaxRounds:4, dryStop:2.venueProfileandresultsModecome fromtemplate.json(e.g. "Traitement du Signal journal â rigor, well-read, signal processing";proposal|data_aware).- Tier 1 â Workflow (use if the
Workflowtool is available): true parallel reviewer isolation â the gold reference.Workflow({ scriptPath: "scripts/review_panel.workflow.js", args: { paperText: "<assembled text>", venueProfile: "<âŠ>", resultsMode: "<proposal|data_aware>", maxRounds: 2, dryStop: 1, verify: true }}) - Tier 2 â subagents (use if there is NO Workflow tool but you CAN spawn subagents â the
Task/Agenttool; this is the normal path at "max tier", and it has the SAME quality as Tier 1 because each subagent is a separate context = true isolation). Per round: (a) spawn the N reviewer subagents in one message (parallel) â each prompt = the gatekeeper core + its one lens +venueProfile+ the mode note + the isolation rule + the inlinedpaperText+ "return your ISSUE_TABLE as JSON"; a fresh subagent literally cannot see peers or prior rounds. (b) You merge the tables in-context (a single reasoning op â no isolation needed) â the round's candidates. (c) For each candidate, spawn the 3 angle-verifier subagents in parallel (each prompt = one angle + the issue fields + the inlinedpaperText+ "return VERIFY JSON: {refuted, angle, reason}"); apply the same majority-refute rule. (d) Same dry/round bookkeeping, same sort +I-NNids, same return object. - Tier 3 â in-context sequential (universal floor; use when nothing is spawnable). You role-play
each lens in sequence under strict isolation discipline that simulates separate contexts: for
each lens, (i) re-read ONLY the paper text + that lens's mandate, deliberately treating any issue/
verdict from an earlier lens as NON-evidence (no contamination); (ii) enforce the verbatim-quote
anti-skim â every issue carries an exact
evidence_quotecopied from the paper, no quote â not filed; (iii) write that lens's full ISSUE_TABLE before starting the next. ONLY after all N lenses: merge/dedupe, then adversarial-verify each candidate by genuinely arguing each of the 3 angles as a skeptic trying to REFUTE (not confirm) and applying the majority rule. loop-until-dry by consciously restarting each round from the paper (use the runningseenset only to suppress duplicates at merge). Emit the identical issue structure +5_review.io.md. Honest note: Tier 3's simulated isolation is a real floor but not equal to Tier 1/2's separate contexts â prefer Tier 2 whenever a subagent tool exists (it almost always does, including "max tier").
- Tier 1 â Workflow (use if the
-
Triage the returned
issues(you, the orchestrator) into three buckets â this is the three-way verdict, done in-context rather than via a heavy trial:- fix now (
blocker/major/cheapminorwith a clearclose_criterion): hand to ts-paper-refine to edit the named section so it satisfies theclose_criterionâ a minimal patch, binding strictly to that one criterion; do not over-rewrite. - author-required (needs new data, a design decision, or a claim the author must own): list it for the user, do not silently edit.
- drop (you judge it invalid even though it survived verify): record why.
- fix now (
-
Re-gate â AFTER the fixes actually land on disk (derive "green", never forecast it). Once every fix-now edit is written, re-run
../ts-paper-write/scripts/draft_lint.py <workdir>+../ts-paper-cite/scripts/citations_lint.py <workdir>and capture each linter's JSON (ok,n) plus a real wall-clock stamp (date -u). A fix must not break a shape contract, a word band, a citation, or notation completeness âdraft_lintnow flagsnotation_incomplete(a symbol added to an equation by a fix, like a new\psi, must get a notation row or be renamed to a defined symbol). Do not proceed on a nonzero exit. Indata_awaremode the number-audit re-runs too. The 'OUTPUT / re-gated green' block of the log may be written ONLY now â quoting the captured JSON + timestamp as evidence; a re-gate whose timestamp pre-dates the editedsections/*.texis invalid by construction (this is the lean, ledger-free analogue of PaperJury computing "closed" from applied+stamped state, never narrating it ahead of the work). -
(Optional) re-review for a thorough pass: re-run the workflow on the edited paper until a clean round (the workflow's own loop already does this within a run; a fresh invocation re-reviews the post-fix draft). Stop when no surviving
majorremains.
What it deliberately does NOT do (kept lean)
No durable cross-session ledger, no full trial/jury/12-juror escalation, no recall-audit, no clerk, no docx intake, no auto-mode. Those are PaperJury's heavyweight machinery for iterating a real manuscript over many human rounds; for a single generation run the lean panel + verify is enough. (If you ever want the full courtroom, PaperJury itself is the tool.)
Trace
Write logs/5_review.io.md (linked from logs/index.md). The log must mirror real state, not forecast
it â DECISIONS may be written as the panel returns, but OUTPUT only after step 4 (above).
- INPUT â paper text size, resultsMode, knobs, and which execution tier ran (Workflow / subagents / in-context).
- DECISIONS â the issue rows projecting the issue object (its
severity/section/evidence_quote/ close_criterion) plus the orchestrator'sidandverdict-bucket:id | severity | section | evidence_quote(the EXACT verbatim quote each reviewer filed â this is the anti-skim proof; an issue with no quote should never have been filed)| close_criterion | verdict-bucket(fix-now / author- required / drop)| what the fix changed; plus a header line withrounds_runandrefuted/dropped_no_criterioncounts (the workflow returns all of these). Each fix-now row is closed only by citing the landed post-fix snippet that satisfies itsclose_criterion. - OUTPUT â written ONLY
Truncated for display â read the full file on GitHub.
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 â from basic concepts to advanced agents, with copy-paste templates that bring immediate value.
lobehub
85.6kð€¯ 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 â see the Safety scan above for what the skill file itself contains.
