메뉴
Record a model training run — log parameters, metrics, and artifact paths, then append a structured row to the experiments ledger. Use when a training run completes, when asked to "log this experiment", or when comparing runs to decide which model to promote.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | experiment-log |
| description | Record a model training run — log parameters, metrics, and artifact paths, then append a structured row to the experiments ledger. Use when a training run completes, when asked to "log this experiment", or when comparing runs to decide which model to promote. |
| metadata | {"forged-by":"cursor-agent","forged-from":"session-2026-03-10","forged-reason":"Logged 3 training runs in three days; each time the ledger entry format was different and comparison was impossible — encoded a consistent schema","usage-count":"26","last-used":"2026-06-10"} |
Record a training run's parameters, metrics, and artifact paths in the experiments ledger.
churn-v3-lgbm){name}-{YYYYMMDD-HHmmss})learning_rate=0.05 n_estimators=500)auc=0.83 f1=0.71 precision=0.79 recall=0.64)Collect run metadata
import datetime, hashlib, json, os
run_id = "{name}-" + datetime.datetime.utcnow().strftime("%Y%m%d-%H%M%S")
git_sha = os.popen("git rev-parse --short HEAD").read().strip()
dataset_version = open("data/.version").read().strip() # semver or hash
Capture parameters and metrics If using a training script that outputs JSON:
python train.py --config config/{name}.yaml --output-metrics /tmp/metrics.json
Otherwise, capture values directly from the trainer object:
params = model.get_params()
metrics = {"auc": roc_auc_score(y_test, y_pred), "f1": f1_score(y_test, y_pred)}
Save artifacts
import joblib
artifact_path = f"artifacts/{run_id}/model.pkl"
os.makedirs(os.path.dirname(artifact_path), exist_ok=True)
joblib.dump(model, artifact_path)
Append to the experiments ledger
scripts/log-experiment.sh \
--name "{name}" \
--run-id "{run_id}" \
--params '{"learning_rate": 0.05}' \
--metrics '{"auc": 0.83, "f1": 0.71}' \
--artifact "{artifact_path}" \
--notes "{notes}"
The ledger is experiments/log.csv — a flat CSV with one row per run.
Verify the entry
tail -n 5 experiments/log.csv
Confirm: run_id is unique, metrics are numeric, artifact path exists.
Compare with previous runs (optional)
python scripts/compare-runs.py --metric auc --top 5
Promote if best (invoke pipeline-stage skill to wire the new model into serving)
Only promote after peer review of the metrics and a sanity-check on the holdout set.
experiments/log.csv — never delete rows; mark superseded runs with status=supersededartifacts/{run_id}/ — one directory per runrun_id, experiment_name, timestamp, git_sha, auc (or primary metric), artifact_pathdataset_version column)-2, -3) rather than overwriting; log a warning.status=failed and the error message in notes; don't leave the ledger entry absent.artifact_size_mb column for reference.artifact_path; the log script accepts both local paths and URIs.Close implementation loops with target-bound independent evidence and a portable proof receipt. Use when an agent changes a real system and must earn a completion claim that survives the session.
First fixture skill for skdd add tests. Use when testing drop installation.
Second fixture skill for skdd add tests. Use when testing multi-skill drops.
Safely backfill historical data for a pipeline stage — scope the date range, run a dry-run, execute in chunks to avoid overwhelming the database, and verify row counts after each chunk. Use when historical data is missing, when a pipeline stage is added retroactively, or when asked to "backfill the {stage} table from {start_date}".
Add or extend data validation checks in the pipeline when data quality issues are detected — null-percentage thresholds, value ranges, referential integrity, and row-count sanity checks. Use when a pipeline breaks on bad data, when onboarding a new dataset, or when asked to "add a quality check for {column}".
Onboard a new raw dataset into the data platform — sniff the schema, generate a profiling notebook, create the ingestion job, and add an entry to the data dictionary. Use when adding a new data source, when a partner delivers a new CSV/Parquet drop, or when asked to "onboard the {name} dataset".