Agent skill · magnus919
data-cleaning
Clean, profile, validate, reshape, and document messy tabular, text, JSON, and relational data through an evidence-first, reproducible workflow. Use when preparing data for analysis, reporting, modeling, ingestion, migration, or matching. Do not use for statistical modeling, dashboard design, or operating a named data platform; route those tasks to data-scientist, data-engineering, or the relevant tool skill.
What it needs
About 3k tokens when loaded.
What this skill does
Data cleaning Treat cleaning as a controlled transformation of an observed dataset, not cosmetic editing. Preserve raw input, state the target use and grain, make every lossy decision explicit, and prove that the cleaned output satisfies a contract. Route by task Need Read next --- --- End-to-end method, scope, and stopping rules references/methodology.md Choose a library or platform references/tool-selection.md Missingness, duplicates, types, ranges, categories, dates, joins references/operations.md Text, identifiers, Unicode, and entity resolution references/text-and-entity.md Schemas, contracts, validation, drift, scale references/validation-and-scale.md CLI, OpenRefine, monitoring, and interactive remediation references/cli-and-interactive-tools.md Source claims and version-sensitive caveats references/sources.md Plan, logs, exceptions, contracts, or reports templates/cleaning-plan.md, templates/transformation-log.jsonl, templates/exception-register.csv, templates/schema-contract.yml, templates/quality-report.md Lightweight profile or reconciliation Run python3 scripts/profiledataset.py --help or python3 scripts/reconciledataset.py --help Default workflow 1. Frame: identify the decision, owner, source, privacy constraints, unit of observation, keys, expected grain, time window, and acceptance threshold. Do not silently infer a business rule from a suspicious value. 2. Freeze evidence: record source path/URI, retrieval time, file size/hash where feasible, encoding, delimiter, schema, row/column counts, and software versions. Keep raw data read-only and write to a new output. 3. Profile before changing: inspect missingness, sentinel values, duplicates, cardinality, type candidates, ranges, invalid dates, whitespace/Unicode anomalies, cross-field relationships, and drift. Use the bundled profiler for a dependency-free first pass. 4. Design decisions: classify each finding as preserve, standardize, repair, impute, quarantine, reject, or escalate. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills magnus919/data-cleaning