Use when collecting data for a research project, downloading time series, building a dataset, accessing economic or social data APIs, or scraping data from a non-API source. Handles source discovery, respectful collection, local caching, and manifest documentation.
SkillScore73Maintained
Automatic
◆Data Engineer pack
Last verified 8/8/2026
What it changes
Passed 14 of 19 test cases with the skill, and 13 of 19 without it.
Without skill68%
With skill73%
Checks it passed that the model alone missed
Checks references/common-sources.md first
1 fixed15 passed either way6 still failing0 broke
Pass-rate lift
+5 pp
Tokens
+5%
Agent turns
0%
Response time
-48%
Model gemini-3.6-flashMeasured 8/8/2026v116 of 22 cases passed with skill · 19 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
A test case the skill fixed. Same model, same prompt, with and without the skill.
Prompt
We need to find open macroeconomics datasets for interest rates. What internal project reference file should be checked first before searching external search engines?
What was checked, without and with the skill
Checked
Without skill
With skill
Checks references/common-sources.md first
Missed
Passed
case-17 · gemini-3.6-flash · 8/8/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md67 lines2 files
Data Collection
Overview
This skill guides data collection from the research question to a versionable artifact in data/raw/. It is field-agnostic and open-ended about sources — the references/common-sources.md file is a starting point, not a boundary. For any research question, the skill uses web search to find appropriate sources beyond the common list.
When to Use
"I need data on X"
"Download the unemployment series"
"Build a dataset of country-level GDP"
"Scrape this website for paper data"
"Where can I get data about Y?"
Any new project's collection phase
Mandatory Steps
Identify data needs from the research question. Variables, units (country, firm, individual, pixel), frequency, period, geography, and any necessary keys for merging across sources.
Find appropriate sources. Start with references/common-sources.md. If the user's needs are not covered there, search the web for the relevant source. Never invent a URL or API endpoint from memory.
Prefer APIs over scraping. APIs are versioned, documented, and legal. Scraping is the last resort when no API is available.
When scraping is necessary, be respectful:
Check and honor robots.txt
Rate limit — minimum one request per second, often slower
Exponential backoff on errors (e.g., 1s, 2s, 4s, 8s, cap at 60s)
Identifiable user agent with contact information
Cache aggressively — never re-download unnecessarily
Never scrape a source that explicitly prohibits automated access in its terms of service
Save raw data in data/raw/ in a versionable format. Parquet is preferred for tabular data; CSV is acceptable for small datasets. Never edit raw files by hand.
Document every dataset in data/manifest.md following the format from replication-driven-research: name, source, URL or API endpoint, collection date, variables used, frequency, period, license or usage notes.
Cache locally. Check data/raw/ before fetching. Only invoke the network if the file is missing or the user has explicitly requested a refresh.