# Ai4science > Guide stage-aware Python AI and AI4Science experiment work: clarify the experiment stage, evaluation protocol, representation, reproducibility, artifacts, and framework reuse while keeping the current workflow sufficient, clear, maintainable, and repo-context driven. - Skill: `yunchengliu/ai4science` (Agent Skill, multi-file: 9 files) - Install (CLI): `npx skillmds@latest add yunchengliu/ai4science` - Raw SKILL.md: https://api.skillmd.com/api/skills/yunchengliu/ai4science/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: YunchengLiu (https://skillmd.com/u/yunchengliu) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/yunchengliu/ai4science --- # AI4Science ## Purpose Use this skill to land and iterate on Python-based AI and AI4Science experiment code while keeping small experiments clean, explicit, and easy to evolve. The skill keeps the current repo and current scientific question in charge: inspect the local context, identify the smallest useful model of the problem, then add only the structure that carries clear meaning for the current question. ## Hard Rules - Use the current experiment stage as the primary design input. - Inspect relevant repo code, data flow, and existing conventions before proposing structure, fields, artifacts, or dependency changes. - Prefer the simplest working research structure that is clear, explicit, rerunnable, and easy to maintain. Engineering formality, architectural purity, and document structure are secondary unless the current workflow benefits from them. - Add modules, classes, configuration objects, artifact types, and workflow machinery only when they make the current experiment materially easier to understand, change, reuse, verify, or rerun. - Model the common structure of the problem before naming fields. Treat a low-information flag as a signal to search for a better shared representation. - Add core dimensions when they express stable, meaningful properties or relations of the problem. Keep values derived when they are cheaper and clearer to compute at the use site. - Treat data loading, preprocessing, modeling, evaluation, and reporting as reasoning stages, not mandatory modules or classes. Split them only when the separation makes the current code or workflow clearer and easier to maintain; keeping related steps together is valid when the main path remains readable. - Keep logging, checkpoints, metrics, and artifacts low entropy: record values that are needed to understand, compare, rerun, or debug the current workflow. Add richer provenance only when the experiment stage requires it. - Let documentation and artifact layout follow actual research use. Persist information needed to interpret, reproduce, compare, or resume work; omit routine activity, transient attempts, and format-only ceremony. - Confirm metric, split, and leakage assumptions before changing train/eval logic. - Keep holdout test data for final reporting, not model selection or iterative feature tuning. - Ask before installing packages, creating environments, changing dependency ownership, or relying on compatibility-sensitive framework behavior. ## Workflow 1. Establish the stage: early setup, local debugging, active iteration, mature comparison, or repeatable workflow. 2. Inspect the relevant local context: entry points, data paths, transforms, model/evaluation code, outputs, and dependency files. 3. Run the representation checkpoint when the task involves feature extraction, classification, preprocessing, schema design, or new fields. 4. Identify the main research path and choose the smallest organization that keeps it clear. Use direct code, local functions, or modules according to the current amount of logic and reuse; do not map each reasoning stage to a code entity. 5. Confirm scientific-integrity basics before train/eval changes: primary metric, validation protocol, split ownership, and leakage risks. 6. Reuse mature framework-native infrastructure before inventing local experiment machinery. 7. Add logging, checkpoints, and artifacts only to the depth needed for the current stage. 8. Surface only assumptions or structural choices that materially affect scientific validity, interpretation, or maintenance of the current task. ## Discuss Before Coding When Needed Pause for a short, natural design discussion only when: - the user asks for feature extraction, classification, analysis variables, or a new data schema but has not stated the modeling assumptions; - the implementation would add several dimensions, flags, configuration keys, log fields, artifact files, checkpoint entries, metrics, or artifact types; - a sample has unusual properties that could become fixture-only fields; - the current implementation has become difficult to understand or change and several materially different organizations remain plausible; - the metric, split, or leakage assumptions would change the implementation. Explain only the relevant entities, representation, downstream use, recommended smallest design, and unresolved choice. Use prose or a small list according to the content; do not require a fixed format. ## Reference Routing - `references/stage-principles.md`: read when the experiment stage controls how much structure to add. - `references/representation-and-decomposition.md`: read before adding feature fields, schemas, stage modules, or pipeline structure. - `references/environment-and-compatibility.md`: read when runtime, dependency, or portability choices matter. - `references/ecosystem-and-reuse.md`: read before building local framework machinery. - `references/structure-principles.md`: read when balancing clear decomposition against needless architecture. - `references/reproducibility-and-artifacts.md`: read when outputs, checkpoints, logs, or reruns are part of the task. - `references/boundaries.md`: read when the task may be PyTorch mechanics, scientific validity, or another specialized domain instead of experiment workflow. ## Self-check Before finalizing a recommendation or patch, verify: - the chosen structure matches the current experiment stage; - every new field, dimension, log entry, artifact, checkpoint entry, or metric has a current purpose for understanding, comparing, rerunning, or debugging; - low-information flags have been replaced by a common representation or kept outside the core model with a stated reason; - IO, preprocessing, modeling/classification, and reporting boundaries are clear; - metric, split, and leakage assumptions are stated or explicitly blocked; - the design leaves repo-specific implementation details to the current code context instead of imposing a generic framework. - every added module, class, configuration layer, artifact type, or document has a current research or maintenance payoff rather than only a conceptual label. ## Response Expectations Explain why the chosen level of structure is sufficient now. Mention the experiment stage only when it materially changes the recommendation. Include representation or decomposition only when it affects the current task. Flag dependency or environment changes for approval. When modeling assumptions are missing, ask a small number of concrete questions or state the unresolved points instead of inventing fields.