SkillAgentSearch skills...

ts-paper-experiment

Repair an AI-generated scientific paper draft by diagnosing research logic, completing feasible experiments, rewriting the experiment section, and updating the full manuscript for claim-evidence consistency.

Install / Use

npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-experiment

Installs into whichever agent you are using.

About this skill
📄

SKILL.md

Installable skill definition

Quality Score

89/100

Category

Automation

Supported Platforms

Universal

Tags

Our assessment of ts-paper-experiment

ts-paper-experiment scores 89/100 on our quality scale, 1396th of 2,780 Automation skills we index.

Its SKILL.md is 29 KB long, well organised into 33 sections with 1 code example: 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.

Substance
30/30
Structure
17/20
Description
15/15
Adoption
13/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated about 3 months ago, so ts-paper-experiment is actively maintained.
  • Our last check on 2026-10-31 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 — read the skill file before letting an agent act on it.

Safety scan

No issues found

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.

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 — read a skill before letting an agent act on it.

ts-paper-experiment compared with similar skills

All 4 of these similar skills score higher than ts-paper-experiment; compare them before choosing.

SkillScoreStarsUpdatedFormat
ts-paper-experiment (this skill)by Spark-To-Paper-Skills891.1k3mo agoSKILL.md
Agent-Reachby Panniantong100112.3k8d agoCLAUDE.md
rufloby ruvnet10081.9ktodayCLAUDE.md
Claude-Code-Game-Studiosby Donchitos10039.7k7d agoCLAUDE.md
Scraplingby D4Vinci100103.6k2d agoMCP Server

Frequently asked questions

How do I install ts-paper-experiment?
Run npx skills add Spark-To-Paper-Skills/spark-to-paper-skills --skill ts-paper-experiment. The install tabs above show the steps for each supported agent.
Which AI agents does ts-paper-experiment 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-experiment 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-experiment still maintained?
The repository was last updated about 3 months ago, so ts-paper-experiment is actively maintained.

name: ts-paper-experiment description: Repair an AI-generated scientific paper draft by diagnosing research logic, completing feasible experiments, rewriting the experiment section, and updating the full manuscript for claim-evidence consistency.

ts-paper-experiment

Now part of the spark-to-paper-skills suite (Stage 8). Previously a separate project (AutoPaperFactory); merged in so the whole pipeline lives in one repo. It runs in an experiment workspace (default <ts_paper_run>/experiments/, seeded by ts-paper/scripts/handoff_to_experiments.py with this skill's paper_config.yaml template). "Project root" below = that workspace. Overleaf is OFF by default; enable via the workspace .env (OVERLEAF_GIT_URL/OVERLEAF_TOKEN) — also in repo-root .env.example. References to "AutoPaperFactory" in the resource docs mean this same workspace.

Repair an AI-generated scientific paper draft into a more complete, honest manuscript. This skill diagnoses the paper, plans and runs only feasible experiments, edits the LaTeX manuscript directly inside ./paper/, and keeps the full paper claim–evidence consistent.

Working directory

Assume Claude Code is always started from the project root (the directory that contains ./paper_config.yaml). The project root is .. Use relative paths only — never absolute or machine-specific paths. Key locations:

  • ./paper_config.yaml — project configuration
  • ./draft_overleaf.zip — the AI-generated Overleaf export (primary input)
  • ./input/draft/ — alternative location for raw draft files
  • ./paper/ — the manuscript: the source of truth, an independent Git repo synced to Overleaf
  • ./workspace/ — diagnosis, experiment planning/runs, logs, lessons (never the manuscript)
  • ./outputs/ — reports, analysis, exported figures/tables (never the source-of-truth manuscript)

Manuscript Source of Truth

./paper/ is the source of truth for the manuscript.

All LaTeX manuscript edits must happen directly inside ./paper/.

Allowed manuscript files in ./paper/ include:

  • .tex
  • .bib
  • .sty
  • .cls
  • figures
  • tables
  • section files

Do not put the following inside ./paper/:

  • diagnosis reports
  • experiment logs
  • raw datasets
  • source code
  • workspace files
  • skill files
  • large model files

Non-negotiable rules

  • Never fabricate results, numbers, tables, figures, citations, references, or statistical significance.
  • Every reported number must be traceable to actual code execution, provided logs, provided data, or clearly marked author-provided results (AUTHOR_TODO).
  • If an experiment cannot be run, write a requirements report — do not invent results.
  • Ask the user before changing the main contribution, deleting the core method, adding a new experiment type, or making strong claims from weak results.
  • ./paper/ is the manuscript source of truth. Do not create a competing "final" manuscript under ./outputs/final/, and do not copy any file from ./outputs/ back into ./paper/. Edit the manuscript in place inside ./paper/.
  • NEVER write that the data/results are simulated, synthetic, fake, or illustrative in ./paper/ (GR-022). A Monte Carlo may live in ./workspace/ to sanity-check an estimator, but its numbers never enter the manuscript and the manuscript never calls anything a "simulation".
  • Be faithful to the paper's own experiment design (GR-023): reproduce the datasets it names, the outcome/treatment it defines, and the estimators it specifies — never swap in an easier proxy, a different construct, or a simulation and present it as the paper's experiment.
  • Actually download the named open datasets before declaring "blocked" (GR-024): genuinely attempt Zenodo/figshare/GitHub/official-portal downloads; only a real, failed attempt justifies a requirements report.
  • After experiments, FILL the paper's existing result tables with the real numbers (GR-025): a blank -- result table after a run is a failure; report null/insignificant results honestly (with SE and p), but the manuscript's own tables must carry the values.
  • Always commit + push to ./paper/'s Overleaf remote at the end (GR-026); re-run the Step 8–17 consistency/review on the FINAL manuscript whenever results/tables/abstract/conclusion change (GR-027).
  • See resources/golden_rules.md and resources/claim_evidence_rules.md for the full policy.

Correct Workflow

Step 0: Entry check and Overleaf setup

Embedded (ts-paper Stage 8) mode — the DEFAULT in this suite, and it AUTO-RUNS. When driven by the ts-paper orchestrator, the draft is ALREADY staged under ./input/draft/ as a LaTeX manuscript (main.tex + sections/ + refs.bib + figures/ + template .sty — NOT a zip), real experiment inputs (if present) are under ./input/{data,code}, and overleaf.require_overleaf_url is false. In this mode auto-proceed without stopping: ingest ./input/draft/ into ./paper/ and run the full workflow end-to-end — do NOT ask for a zip or an Overleaf URL. (The "stop and ask" prompts below apply ONLY to standalone manual usage, never to embedded Stage-8 runs.)

  • Check for the AI draft (at least one of): ./input/draft/ (the staged manuscript — primary in the embedded workflow), ./draft_overleaf.zip (standalone only), or existing LaTeX files under ./paper/.
  • Overleaf toggle (check first): read overleaf.require_overleaf_url from ./paper_config.yaml.
    • If it is false (Overleaf disabled): skip the Overleaf Git target check, the URL ask, and all pushing entirely. Work only on the local ./paper/ repo (init + local commits, no remote, no push), regardless of the push_* flags. Do not call init_paper_overleaf.py --push. Then continue to the draft check below.
    • If it is true: enforce the Overleaf Git target as described next.
  • Check for the Overleaf Git target (at least one of): overleaf.git_url in ./paper_config.yaml, or OVERLEAF_GIT_URL in ./.env, or an existing remote named overleaf inside ./paper/. (paper_config.yaml takes precedence, then the .env value.)

If the AI draft is missing (standalone manual usage only — in embedded Stage-8 mode ./input/draft/ is always present, so never reached), stop and ask:

Please provide the AI draft. Put an Overleaf project zip at ./draft_overleaf.zip, or place the draft files under ./input/draft/.

If require_overleaf_url is true and the Overleaf Git URL is missing, stop and ask:

Please provide the Overleaf Git URL. Create a blank Overleaf project first, open Git Integration, and paste the Git URL into ./paper_config.yaml under overleaf.git_url.

Security rule: Do not ask for passwords or tokens in the prompt. For non-interactive pushes, the user can create a ./.env file with OVERLEAF_GIT_URL and OVERLEAF_TOKEN. The token is read only from ./.env, is never stored in paper_config.yaml, never written into the git remote URL, and never printed. scripts/init_paper_overleaf.py supplies it to git via a temporary askpass helper. ./.env is gitignored and must never be committed. A Git credential helper or SSH also works.

Then:

  • Extract/copy the AI draft directly into ./paper/ (unzip ./draft_overleaf.zip and/or copy ./input/draft/ files into ./paper/).
  • Initialize ./paper/ as an independent Git repository.
  • Commit the original AI draft inside ./paper/ with message: Initial import of AI draft.
  • Push the original AI draft to the Overleaf remote if configured (overleaf.push_initial_draft).

You may use the helper scripts/init_paper_overleaf.py (add --push to push). Never run git in the parent AutoPaperFactory repository for paper content.

Step 1: Diagnose the paper

  • Read the manuscript files directly from ./paper/ (the .tex/.bib sources).
  • Read paper_config.yaml and the resources: methodology_core.md, golden_rules.md, anti_patterns.md, experiment_templates.md, claim_evidence_rules.md, contribution_type_rules.md, human_scientific_style.md, sci_q4_review_checklist.md, reference_verification.md, experiment_table_audit.md, publication_polish_check.md, final_submission_checklist.md, post_repair_narrative_review.md, dataset_license_gate.md, code_experiment_paper_consistency.md, raw_result_recomputation.md, rendered_pdf_layout_check.md, reproducible_artifact_packaging.md, experiment_prerun_approval_gate.md.
  • Write diagnosis reports to ./workspace/diagnosis/:
    • DIAGNOSIS.md, PAPER_LOGIC_MAP.md, EXPERIMENT_GAP_REPORT.md.
  • The diagnosis must identify: paper topic, research problem, claimed contribution, method, current experiments, missing experiments, unsupported claims, and likely SCI Q4 readiness issues. Do not polish prose yet (GR-002).

Step 2: Plan experiments

  • Write the experiment plan to ./workspace/experiments/EXPERIMENT_PLAN.md specifying: which experiments are necessary, which are feasible with available data/code, which are blocked (and why), metrics, baselines, expected outputs, and commands/ scripts to run. Use resources/experiment_templates.md. Plan the minimum set.
  • Pre-run plan & resource approval gate (GR-021). Before running expensive experiments or downloading external data, present a pre-run plan and get explicit user approval. Write ./outputs/reports/PRERUN_PLAN_APPROVAL.md covering, per experiment: estimated runtime, disk, compute/cost, external data needed (with the Dataset / License / Terms Gate, Step 3), and whether it is needed for an essential claim. Cheap, local, already-available runs may proceed without a separate approval; costly/external ones must wait for approval. See resources/experiment_prerun_approval_gate.md.

Step 3: Run feasible experiments

  • If data/code are available, inspect ./input/code/, run existing scripts or create minimal scripts only when necessary, and use the seeds in paper_config.yaml.
  • Dataset / License / Terms Gate (mandatory before any external dataset download or use). Before downloading or using an external dataset, report (see resources/dataset_license_gate.md): dataset name; source URL; expected download size; license / terms-of-use status; whether user confirmation is required; disk requirement; estimated runtime; and whether the dataset is needed for essential claims. If the license/ToU is unclear, stop and ask the user.
  • Write experiment logs/results to ./workspace/experiments/ and ./outputs/ (final tables/figures to ./outputs/tables/ and ./outputs/figures/).
  • Do not put logs, raw data, or code into ./paper/.
  • Never hardcode fake numbers. Create workspace/experiments/EXPERIMENT_RUN_REPORT.md. If experiments cannot be run, create workspace/experiments/EXPERIMENT_REQUIREMENTS.md instead.

Step 4: Analyze results

  • Write the result analysis to ./outputs/reports/RESULT_ANALYSIS.md: main findings; whether each original claim is supported (labels from resources/claim_evidence_rules.md); weak or negative results (do not hide them); limitations; which claims must be weakened or removed.

Step 5: Code–Experiment–Paper Consistency Gate (MANDATORY)

Before any reported number is written into the manuscript, verify that the evidence is real and that the code matches the paper. This gate reads the current paper project's evidence (./input/code/, ./workspace/experiments/, ./outputs/) — it does not invent or run anything new here. See resources/code_experiment_paper_consistency.md. Produce six reports under ./outputs/reports/:

  1. RESULT_PROVENANCE_AUDIT.md — every n

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars1.1k
CategoryAutomation
Updated2mo ago
Forks19

Languages

Python

Trust signals

100/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

No cautions