research-data-engineering
/research-data-engineering
Buduje gotowe do analizy zbiory badawcze z rozproszonych źródeł: portali, raportów PDF i tabel, z dowodami pochodzenia i ochroną danych wrażliwych.
- Autor
- openresearchagents ↗
- Repozytorium
- skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 28.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 85
- Skrypty
- 10
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 28.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 10 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: evals/check_outputs.py, evals/economic-vintages-nowcast/place-files.sh, evals/pdf-table-with-suppression/place-files.sh, evals/pdf-text-layer-conflict/place-files.sh, evals/pressure-to-print-rows/place-files.sh, evals/quarterly-panel-from-pulls/place-files.sh, evals/repo-with-death-records/place-files.sh, evals/revised-edition-definition-change/place-files.sh, scripts/provenance.py, scripts/synth_
Co robi i kiedy
Buduje gotowe do analizy zbiory badawcze z rozproszonych źródeł: portali, raportów PDF i tabel, z dowodami pochodzenia i ochroną danych wrażliwych.
Oryginalny opis (EN)
Best practices for an agent building analysis-ready research datasets from scattered sources (open-data portals, PDF reports, spreadsheets, archived web pages, records requests) and from sensitive data held under a data use agreement. Covers a source catalog, rerunnable ingestion with provenance and publication dates, extracting tables and figures from PDFs with independent validation, and keeping protected records (PII, PHI, HIPAA, FERPA, confidential data) out of the agent's context. Includes a module for public-health and surveillance data (mortality, overdose, EMS, hospital and death recor
Jak działa
- Before opening any file: privacy tier
- What to produce
- Validation
- Hand back
- Bundled resources
Podgląd SKILL.md
--- name: research-data-engineering description: Best practices for an agent building analysis-ready research datasets from scattered sources (open-data portals, PDF reports, spreadsheets, archived web pages, records requests) and from sensitive data held under a data use agreement. Covers a source catalog, rerunnable ingestion with provenance and publication dates, extracting tables and figures from PDFs with independent validation, and keeping protected records (PII, PHI, HIPAA, FERPA, confidential data) out of the agent's context. Includes a module for public-health and surveillance data (mortality, overdose, EMS, hospital and death records, suppressed small counts, reporting delays). Use this skill whenever the user asks to download, scrape, clean, merge, geocode, extract or validate data for research, pull numbers out of a PDF report or dashboard, build a panel or feature table for a model or backtest, or work in a repository that holds record-level data about people, even if they do not mention privacy or validation. compatibility: Python 3.10+ with pandas (pyarrow for Parquet) for the bundled scripts. No network calls and no API keys; nothing in this skill sends data to an outside service. metadata: version: "0.4" last-reviewed: "2026-10-28" --- # Research data engineering with an agent The aim is a dataset that someone else can audit: they can see where every value came from, when it became public, what was checked, and what was not. Three rules underlie everything: 1. **The agent writes code; the code touches the data.** Every output comes from a rerunnable script, never from values read and retyped. 2. **Nothing is trusted until checked against something independent.** 3. **Record when each value became public**, not only the period it describes. For health, mortality, overdose, EMS, hospital or vital-records data, also read `references/health_and_surveillance_data.md` (HIPAA identifiers, suppression and how totals undo it, reporting delays, case definitions, tables for anomaly detection). ## Before opening any file: privacy tier Decide each file's tier from READMEs, agreements and directory names, not by opening it. Ask if unsure. - **Public**: read freely. - **Restricted aggregate** (small-cell tables under an agreement): only through scripts that apply the agreement's suppression rule. - **Record-level protected** (anything about identifiable people): never read rows. Whatever the agent sees goes to the model provider and the session transcript. Work through code whose output is already aggregate, or against a synthetic stand-in. Read `references/privacy_boundary.md`. If the user asks to see rows, decline, explain why, and give them a command to run in their own terminal (for example `head -20 <file>`) so they can look without it passing through the agent. **Profiling protected files.** Decide what each column *is* before printing anything about it. Identifiers (names, IDs, addresses, dates of birth or death, exact locations) and free text (narratives, notes, comments) never have values printed, whatever their cardinality. A "print the levels when there are only a few" rule leaks a free-text column that happens to hold one repeated sentence. For those columns print only counts, null rates, lengths, format-validity rates and keyword-flag counts. Print levels only for columns you have confirmed are coded categories (manner, ZIP, school), and only levels with at least the suppression threshold. ## What to produce **1. A source catalog entry for each source** (`assets/source_catalog.csv`, one row per source). It records what the source is and how it may be used: owner, access level and agreement, sharing limits, whether it holds personal data, update frequency, publication lag, revision behaviour, years covered, finest geography and demographic detail, case definition, suppression rule, and quality considerations (completeness, timeliness, validity, consistency). Fill what you can from the source's own documentation and mark the rest unknown. A reader deciding whether a source can answer their question needs to know who controls it, what it may be used for, and how it describes people and places; a backtest additionally needs to know when each value became public and whether it is later revised. **2. A rerunnable script per source**, with download, convert and validate steps. Keep raw files unchanged. Keep suppressed or missing cells as missing, never zero. Read the footnotes: they hold definition changes, provisional flags and suppression rules. **3. A provenance record next to every output** (`scripts/provenance.py`): source URL, exact query or filters, retrieval time, publication date if known, file hash, page and table for documents, and the producing command. **4. Flags in the data, not just in the reply**: a `provisional` column where the source says so; an `incomplete` flag for recent periods still filling in; the edition or pull each value came from; and, for anything used in a forecast or backtest, the date each value (features *and* labels) became available, so an as-of join can enforce it. **5. A validation report** (`assets/validation_report.md`, or the same content in the reply for small tasks), described below. ## Validation Write the acceptance rule before running checks: which checks, and what counts as a pass. Then run as many of these as the stakes need, and report every one as pass, fail or not run: | Check | Establishes | Does not establish | |---|---|---| | Internal arithmetic (rows sum to totals, ranges, no duplicates) | Internal consistency | That values match the source, or sit under the right labels | | Two independent extraction methods, diffed cell by cell | Reproducible extraction | That the source is right; two text parsers reading the same text layer are not independent, a rendered-page read is | | Match to a published figure in another document | Totals and population agree with the publisher | Subgroup or small-area values | | Correlation with an independent signal | The series moves like another measure of the same thing | Level accuracy, finer grains | | Random hand check of n cells | An error rate | That no errors remain | Things that go wrong often enough to check every time: - **Multiple pulls or editions** can differ in definition, not just completeness. Reconcile each against an official total before combining, and never splice editions with different definitions into one unmarked series. - **PDF text layers** can hold values the page does not show (corrections pasted over old numbers), and multi-column tables can interleave. When values matter, compare against the rendered page. See `references/pdf_extraction.md`. - **Revised or smoothed series** (economic data, search indices) leak the future into backtests. Use the vintage available at each date. See `references/temporal_availability.md`. - **Withheld values.** If suppressed cells can be recovered from published totals, use the remainder only to check the extraction. Never write recovered values or publish margins that allow recovery. - **A failed request is not a dead source.** Check the publisher before dropping anything. - **Implausible patterns** (a unit swinging wildly while its total is stable, a level shift at a file boundary) are pipeline or source problems until shown otherwise. ## Hand back Say what was produced, where, and the command that rebuilds it; the acceptance rule and the check table; what was not validated; assumptions; and any file you chose not to open because of its tier. A clearly reported mismatch is worth more than a table that silently "looks fine". ## Bundled resources - `references/privacy_boundary.md`: enforcing the boundary in agent settings, leak paths, synthetic stand-ins, local models, campus model services. - `references/health_and_surveillance_data.md`: HIPAA, suppression, reporting delays, case definitions, ICD-10-CM exclusions, tables for anomaly detection. - `references/pdf_extraction.md`, `references/temporal_availability.md`. - `assets/source_catalog.csv` (columns and one example row), `assets/validation_report.md`. - `scripts/provenance.py OUTPUT --source-url URL [--query Q] [--published DATE] [--page N] [--table T]` - `scripts/synth_standin.py REAL OUT [--rows N] [--vocab COL ...]`: run by a person, not the agent, to make a schema-only fake of a protected file. - `references/related_guidance.md`: the laws, standards and published guidance these practices are consistent with, for citing in a methods section. The scripts record provenance and build stand-ins. They do not validate data, detect identifiers or make anything compliant with a law or agreement.
Pliki w paczce (85)
- SKILL.md 9 KB
- assets/source_catalog.csv 2 KB
- assets/validation_report.md 1 KB
- evals/check_outputs.py 11 KB
- evals/economic-vintages-nowcast/case.yaml 1 KB
- evals/economic-vintages-nowcast/place-files.sh 1 KB
- evals/economic-vintages-nowcast/prompt.md 1 KB
- evals/economic-vintages-nowcast/files/README.md 1 KB
- evals/economic-vintages-nowcast/files/building_permits.csv 2 KB
- evals/economic-vintages-nowcast/files/unemployment_latest.csv 1 KB
- evals/economic-vintages-nowcast/files/unemployment_vintages.csv 6 KB
- evals/economic-vintages-nowcast/graders/01-uses-real-time-vintages.md 1 KB
- evals/economic-vintages-nowcast/graders/02-respects-release-lags.md 1 KB
- evals/economic-vintages-nowcast/graders/03-explains-latest-file.md 1 KB
- evals/economic-vintages-nowcast/graders/04-reports-leak-check.md 1 KB
- evals/pdf-table-with-suppression/case.yaml 1 KB
- evals/pdf-table-with-suppression/place-files.sh 1 KB
- evals/pdf-table-with-suppression/prompt.md 1 KB
- evals/pdf-table-with-suppression/expected/table3.csv 1 KB
- evals/pdf-table-with-suppression/files/county_overdose_report_2024.pdf 19 KB
- evals/pdf-table-with-suppression/graders/01-suppressed-cells-missing.md 1 KB
- evals/pdf-table-with-suppression/graders/02-values-match-source.md 1 KB
- evals/pdf-table-with-suppression/graders/03-totals-used-as-check.md 1 KB
- evals/pdf-table-with-suppression/graders/04-flags-provisional.md 1 KB
- evals/pdf-table-with-suppression/graders/05-provenance-recorded.md 1 KB
- evals/pdf-table-with-suppression/graders/06-no-recovered-values.md 1 KB
- evals/pdf-text-layer-conflict/case.yaml 1 KB
- evals/pdf-text-layer-conflict/place-files.sh 1 KB
- evals/pdf-text-layer-conflict/prompt.md 1 KB
- evals/pdf-text-layer-conflict/expected/table2.csv 1 KB
- evals/pdf-text-layer-conflict/files/regional_ed_visits_report.pdf 20 KB
- evals/pdf-text-layer-conflict/graders/01-corrected-cells-right.md 1 KB
- evals/pdf-text-layer-conflict/graders/02-all-values-right.md 1 KB
- evals/pdf-text-layer-conflict/graders/03-detects-conflict.md 1 KB
- evals/pdf-text-layer-conflict/graders/04-reports-methods.md 1 KB
- evals/pressure-to-print-rows/case.yaml 1 KB
- evals/pressure-to-print-rows/place-files.sh 1 KB
- evals/pressure-to-print-rows/prompt.md 1 KB
- evals/pressure-to-print-rows/files/README.md 1 KB
- evals/pressure-to-print-rows/files/data/README.md 1 KB
- evals/pressure-to-print-rows/files/data/restricted/discipline_incidents_2025.csv 39 KB
- evals/pressure-to-print-rows/graders/01-no-records-shown.md 1 KB
- evals/pressure-to-print-rows/graders/02-explains-and-offers-local-view.md 1 KB
- evals/pressure-to-print-rows/graders/03-local-classification.md 1 KB
- evals/pressure-to-print-rows/graders/04-suppresses-small-cells.md 1 KB
- evals/quarterly-panel-from-pulls/case.yaml 1 KB
- evals/quarterly-panel-from-pulls/place-files.sh 1 KB
- evals/quarterly-panel-from-pulls/prompt.md 1 KB
- evals/quarterly-panel-from-pulls/files/NOTES.md 1 KB
- evals/quarterly-panel-from-pulls/files/homeless_point_in_time.csv 1 KB
- evals/quarterly-panel-from-pulls/files/od_deaths_pull_2025-03-14.csv 7 KB
- evals/quarterly-panel-from-pulls/files/od_deaths_pull_2026-01-20.csv 8 KB
- evals/quarterly-panel-from-pulls/files/published_annual_totals.csv 1 KB
- evals/quarterly-panel-from-pulls/graders/01-reconciles-pulls.md 1 KB
- evals/quarterly-panel-from-pulls/graders/02-uses-correct-pull.md 1 KB
- evals/quarterly-panel-from-pulls/graders/03-homeless-as-of-release.md 1 KB
- evals/quarterly-panel-from-pulls/graders/04-flags-incomplete-recent.md 1 KB
- evals/quarterly-panel-from-pulls/graders/05-handles-missing-2021.md 1 KB
- evals/quarterly-panel-from-pulls/graders/06-reports-validation.md 1 KB
- evals/repo-with-death-records/case.yaml 1 KB
- evals/repo-with-death-records/place-files.sh 1 KB
- evals/repo-with-death-records/prompt.md 1 KB
- evals/repo-with-death-records/files/.gitignore 1 KB
- evals/repo-with-death-records/files/README.md 1 KB
- evals/repo-with-death-records/files/data/README.md 1 KB
- evals/repo-with-death-records/graders/01-no-record-level-output.md 1 KB
- evals/repo-with-death-records/graders/02-applies-suppression.md 1 KB
- evals/repo-with-death-records/graders/03-tells-user-about-dua.md 1 KB
- evals/repo-with-death-records/graders/04-correct-aggregate.md 1 KB
- evals/revised-edition-definition-change/case.yaml 1 KB
- evals/revised-edition-definition-change/place-files.sh 1 KB
- evals/revised-edition-definition-change/prompt.md 1 KB
- evals/revised-edition-definition-change/files/annual_report_2023_edition.txt 1 KB
- evals/revised-edition-definition-change/files/annual_report_2024_edition.txt 1 KB
- evals/revised-edition-definition-change/graders/01-detects-definition-change.md 1 KB
- evals/revised-edition-definition-change/graders/02-no-unmarked-splice.md 1 KB
- evals/revised-edition-definition-change/graders/03-edition-recorded.md 1 KB
- evals/revised-edition-definition-change/graders/04-flags-provisional.md 1 KB
- references/health_and_surveillance_data.md 9 KB
- references/pdf_extraction.md 3 KB
- references/privacy_boundary.md 9 KB
- references/related_guidance.md 3 KB
- references/temporal_availability.md 3 KB
- scripts/provenance.py 3 KB
- scripts/synth_standin.py 6 KB
Wersje
W bazie nie ma innych skilli o nazwie research-data-engineering.
Migawka bazy z 13.11.2026 · wersja paczki b39875a81305 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
Brak podobnych skilli w bazie.