Core Idea
The author's design philosophy, what sets this apart from peer projects, and quick notes on the key workflows. Skill names mentioned in the text auto-link to their detail pages.
RigorPilot is a research-first Agent Skills set for deep learning experiments, differentiating from score-chasing coding agents by enforcing comparability, reproducible evidence, and auditable changes.
current_research.repro_outputs/, explore_outputs/, etc. with changelog and comparability reports.Use /ai-research-reproduction or /ai-research-explore as entry; agent routes via lifecycle (Understand→Reproduce→Setup→Run/Train→Debug→Report). Keep researcher in control with auditable artifacts.
Skill Relationships
Skill → Skill usage relationships, with cross-layer references. Same color blocks indicate the same skill form.
Skill Composition
Each skill is designed for a specific kind of caller — some only fire when you name them, others load automatically when the model judges them relevant. The split below summarises both modes at a glance.
These skills only fire when you call them by name. The agent will not load them on its own — they stay inert until invoked.
These skills are loaded automatically by the agent when it judges the context relevant. You usually don't need to call them yourself.
Skill List
Rigor Explore skill for candidate-only deep learning research exploration atop a durable current_research anchor. It uses a two-loop rhythm, idea gating, fair comparison, and governed experiments via orchestration scripts, writing auditable outputs to explore_outputs/. Value: rigorous, comparable, auditable candidate ideation without trusted reproduction claims.
End-to-end README-first skill for trustworthy deep learning repo reproduction. It coordinates intake, setup, execution, optional training, analysis, and paper-gap resolution via sub-skills, enforces conservative patch rules, records evidence and deviations, and writes a standardized `repro_outputs/` bundle for auditable, minimal-trust reruns.
A read-only analysis skill for deep learning research repos. It inspects model structure, training/inference entrypoints, and configs via lightweight static helpers, flags suspicious patterns as heuristics, and outputs SUMMARY.md, RISKS.md, status.json. Value: safe repo understanding before changes without execution.
Prepares conservative conda-first environment and asset path assumptions for README-documented deep learning repo reproduction. Uses policy refs and Python scripts to plan setup, checkpoints, datasets, cache hints. Outputs setup notes and risk list enabling safe downstream runs.
Auditable candidate implementation skill for deep learning repos. On explicit authorization, it plans source-anchored code changes on isolated branches via scripts, producing explore_outputs records with rollback and comparability notes. Value: low-risk exploratory modifications separated from trusted baseline.
Rigor Improve/Explore run leaf skill for bounded exploratory evidence in DL repos. Plans candidate runs (small-subset, sweeps, transfer trials) with fair-comparison caveats and writes no-overclaim summaries to explore_outputs/. Value: auditable, low-risk exploratory ranking separated from trusted baselines.
Rigor Run skill for README-first deep learning repo reproduction. It captures and normalizes evidence from selected smoke tests or documented inference/evaluation commands, then writes standardized repro_outputs/ files with patch notes. Value: auditable, trustworthy execution evidence without hidden scientific changes.
A helper skill for README-first deep learning repo reproduction that resolves narrow reproduction-critical gaps (dataset split, preprocessing, eval protocol, checkpoint mapping, runtime assumptions) from primary paper sources. It records conflicts and distinguishes evidence from inference, providing targeted supplements to the main orchestration without summarizing the whole paper.
A helper skill for README-first deep learning repo reproduction. It scans the repo, reads README and common files, extracts documented commands via scripts, classifies inference/eval/training candidates, and returns the smallest trustworthy reproduction plan to the orchestrator, enabling safe, low-risk repo understanding before execution.
Rigor Train skill for deep learning repos to run a selected training command conservatively for startup, short-run, full, or resume modes. It executes the command, captures config/seed/log/checkpoint/metric evidence, and writes standardized train_outputs/ files. Value: bounded, auditable training evidence with reproducibility context.
A conservative debug/audit skill for deep learning failures (traceback, CUDA OOM, NaN loss). It diagnoses root cause without code changes by default, proposes minimal approved patches, and separates debug fixes from research contributions. Outputs diagnosis, patch plan, and status files for safe, auditable debugging.