lllllllama/ai-paper-reproduction-skill
11 skills563k installs5.6k starsGitHub
Install
npx skills add https://github.com/lllllllama/ai-paper-reproduction-skillSkills in this repo
1Paper Context ResolverPaper-context-resolver is a helper skill for resolving specific reproduction-critical gaps from primary research papers. It supplements README-first reproduction by narrowing down reproduction questions about dataset versions, splits, preprocessing, evaluation protocols, and checkpoint mappings. Use when the README leaves gaps on these specific details.141kinstalls2Repo Intake And PlanRepo-intake-and-plan is a helper skill that scans deep learning repositories to extract structure, documented commands, and candidate classification. It returns a minimum trustworthy reproduction plan by mapping repo structure and identifying inference, evaluation, and training targets. Use at the beginning of reproduction work to understand what commands are available and which are safest to run.140kinstalls3Minimal Run And AuditMinimal-run-and-audit is a rigor skill that executes documented smoke tests and evaluation commands, then captures evidence in standardized formats. It normalizes repro_outputs/ files, documents scientific changes via SCIENTIFIC_CHANGELOG, and provides a COMPARABILITY_REPORT. Use when running reproduction commands and needing auditable, comparable evidence.140kinstalls4Env And Assets BootstrapThe env-and-assets-bootstrap skill is the Rigor Setup step in a README-first deep learning reproduction pipeline. It runs after repo intake selects a credible target and before any training or evaluation commands execute. The skill produces conservative environment setup notes, candidate conda commands, asset path plans, checkpoint and dataset source hints, and explicit unresolved dependency or asset risks. It deliberately does not own target selection, full orchestration, paper interpretation, final run reporting, or generic package management outside a specific reproduction goal. Inputs include the target repo path, chosen reproduction objective, relevant README setup steps, and known OS or package constraints. Supporting references env-policy.md and assets-policy.md plus scripts bootstrap_env.py, plan_setup.py, and prepare_assets.py ground setup planning. Developers skip it when the repository already ships a ready-to-run environment that needs no translation from README instructions.140kinstalls5Ai Research Reproductionai-research-reproduction is a Rigor Reproduce compatible agent skill from lllllllama/ai-paper-reproduction-skill for README-first deep learning repository reproduction with auditable evidence and conservative patching. Listed on skills.sh with 507 installs, it guides agents to read the repository first, pick the smallest documented inference or evaluation target, coordinate intake and setup, run trusted execution, optionally attempt trusted training, and record assumptions, deviations, and human decision points. The standardized repro_outputs/ bundle captures evidence so reproduction means following documented commands and environments—not arbitrary changes to force a green run. The skill loads shared references including agent-operating-principles.md, research-rigor-principles.md, and deep-learning-experiment-principles.md when scientific comparability matters. Developers reach for ai-research-reproduction when validating paper claims, benchmarking model repos, or producing audit trails for ML experiments. It explicitly excludes paper summaries, generic environment setup alone, or isolated repo scanning without a reproduction target.573installs6Analyze Projectanalyze-project is a research-engineering skill from lllllllama/ai-paper-reproduction-skill that scopes AI paper reproduction efforts against real repositories. It inspects existing code, configs, datasets, and experiment scripts to list what already matches a target paper and what gaps remain—missing modules, untrained checkpoints, or divergent hyperparameters. ML engineers and research reproducibility teams invoke analyze-project at the start of a reproduction sprint when a partial implementation exists and they need a structured gap analysis before writing training jobs or benchmarking scripts. Output feeds the broader ai-paper-reproduction workflow with prioritized tasks for datasets, model code, evaluation metrics, and compute requirements.422installs7Explore Codeexplore-code is a skill from lllllllama/ai-paper-reproduction-skill that helps agents orient inside third-party ML research repositories. Before writing a reproduction plan, it systematically finds training scripts, configuration files, dataset loaders, and main entrypoints that authors used to produce published numbers. Developers reach for explore-code when cloning a paper's GitHub repo and facing an unfamiliar layout—multiple experiment folders, nested configs, or undocumented launch commands. The skill reduces time lost grep-wandering and prevents wrong-script execution that wastes GPU hours. It pairs with downstream reproduction skills by delivering a navigable map of the codebase's execution paths and data dependencies.417installs8Safe Debugsafe-debug is a Claude Code skill for ML researchers and engineers reproducing academic papers who need iterative debugging without destroying expensive training artifacts. The skill guides agent-assisted fixes to failing reproduction scripts, dependency mismatches, and runtime errors while enforcing safeguards around datasets, saved checkpoints, and in-flight GPU jobs. Developers reach for safe-debug when a paper reproduction stalls mid-pipeline and blind file edits or reckless reruns risk overwriting weeks of checkpoint data or relaunching costly training from scratch. It fits agent-driven debugging sessions on PyTorch or TensorFlow reproduction codebases where one wrong command can cascade into data loss.401installs9Run Trainrun-train is the Rigor Train agent skill in lllllllama/ai-paper-reproduction-skill that executes a pre-selected training command conservatively and normalizes resulting evidence into train_outputs/. It supports startup verification, short-run verification, full training kickoff, and resume modes while recording commands, configs, seeds, logs, checkpoints, status, and metrics. The skill explicitly excludes environment setup, exploratory sweeps, speculative code changes, and end-to-end orchestration—it executes one bounded training run with reproducibility context preserved. Outputs include SUMMARY.md, COMMANDS.md, LOG.md, SCIENTIFIC_CHANGELOG.md, COMPARABILITY_REPORT.md, and status.json. Developers reach for run-train when a paper-reproduction repo already has a chosen training command and needs structured, auditable training evidence rather than ad-hoc log files.395installs10Explore Runexplore-run is a paper-reproduction scouting skill for ML engineers evaluating whether an AI research repository is worth fully replicating. It guides exploring repository layout, installing dependencies, and executing baseline experiments to verify stated results before investing in a complete reproduction pipeline. The workflow catches broken setups, missing assets, and overstated claims early. Use explore-run when onboarding to a new paper codebase, benchmarking reproducibility risk, or deciding if full replication belongs on the roadmap.393installs11Ai Research Exploreai-research-explore is a Claude Code skill for ai & agent building. It helps solo builders move faster with AI-assisted development.180installs