Automatic execution policy: Read automation guidance first. Creative rules and review findings guide improvements; scores, review/hash/freeze records, CTA choices and intermediate approval steps do not block authorized work. Preserve actual external-action authorization and stop at the artifact the user requested. This policy supersedes older quality-gate wording below.

Public data research — data.go.kr as a content seed

Find and collect open APIs and file datasets related to a topic on the Korea open-data portal (data.go.kr), and record them as sourced research seeds. Source data from government and public bodies is itself a primary source — for the same number, prefer the original dataset over a news article requoting it (it also lightens the cross-check burden: 1 primary source = settled, 2 articles = cross-checked).

Tool map

The auth key is injected as DATA_GO_KR_API_KEY from the shell environment (README §Environment variables). Even with a key, a per-API 활용신청 (usage application on the portal — log in; most are auto-approved) has to come first — see §4 below.

Procedure

1. Search — find the datasets

Search a keyword with datago_search (the arguments match the other search tools — query, limit, page). Tips:

• On the first search, omit type (API and FILE together) — read the totals to see the shape of the data.
• Start with a short keyword ("환율" / exchange rate → 82 hits); if that doesn't narrow, add the agency name or a more specific term ("한국은행 기준금리" / Bank of Korea base rate). Don't churn through query variations that aren't the research subject.
• Selection criteria: recently modified (an old one-off dataset carries staleness risk), and a provider that actually owns the subject.

2. Detail — check the collection path and staleness

Call datago_detail with a candidate dataset's publicDataPk + type. Always read:

• Last-modified date and update cycle — "수시 (1회성 데이터)" ("as needed (one-off data)") means the data's own reference date can be old (e.g. a DB collected in 2017). Unsuitable for time-sensitive topics.
• License scope (이용허락범위) — confirm "제한 없음" ("no restrictions"). If there are restrictions, report to the user and let them decide.
• FILE → publicDataDetailPk (uddi); API → requestUrls / docs (the usage guide).

3. Collection

FILE (the default route) — get the original with datago_file_download:

• saveDir is data/<channel>/episodes/<topic>/storyboard/ when working on a topic, or the absolute path docs/research/<YYYY-MM-DD>-<slug>/ for standalone research.
• If the response's encoding is euc-kr, convert before reading: iconv -f euc-kr -t utf-8 <file> > <file>.utf8.csv
• For tens of MB, or when you need only some rows, paging with datago_file_fetch (needs the usage application) is better.

API (live and query-conditional) — check the parameter contract first, then call:

• datago_detail's requestUrls is the path; the parameters are in docs (the usage guide document, hwp/docx — when WebFetch can't read it, fetch it with curl regardless of the datago_file_download format and convert, or WebFetch the detailUrl).
• Don't call repeatedly while guessing parameters — the daily traffic (usually 1,000 calls) burns out.

4. The 활용신청 gate (manual — guide the user)

When datago_file_fetch / datago_api_call returns "등록되지 않은 인증키" ("unregistered auth key", -4) or "인증 거부" ("auth denied"), it's usually not the key but a missing 활용신청 (usage application). Walk the user through it:

1. Log in to the portal → the dataset's detail page (detailUrl) → the 활용신청 ("apply for use") button
2. Most are auto-approved (instantly). Retry the same call after approval.
3. If it still fails, check that the key under 마이페이지 > 개인 API 인증키 (My Page > personal API auth key) matches DATA_GO_KR_API_KEY.

If the application can't be waited on, route FILE around it with datago_file_download (no auth).

5. Record the seed — the source contract

Record what you collected in the research document (research.md while working on a storyboard, or docs/research/<YYYY-MM-DD>-<slug>/public-data.md for standalone research):

| Claim/number | Source data | Data reference date | Source |
|---|---|---|---|
| 34 marinas nationwide | 해양수산부_전국 마리나 현황 (34 rows) | modified 2025-11-12 | [data.go.kr](https://www.data.go.kr/data/15152090/fileData.do) |

• Cite as agency name_dataset name + data reference date + detail URL — so the content can quote it as "Ministry of Oceans and Fisheries public data (2025)".
• The data reference date may not be the last-modified date — check the actual reference point from the file contents (a year column, the filename) and record that. Describing data collected in 2017 as "current" is a distortion.
• When processing numbers, the principle matches the storyboard skill: a range stays a range, and rounding never changes the meaning.

Traps

• Don't be fooled by the last-modified date — the portal's modified date is when the metadata was updated. A dataset of "5,003 total rows / collected in 2017" can show modified 2025 (a case measured in practice: 해양관광레저정보DB, the marine tourism and leisure information DB).
• EUC-KR files — ignore the response's encodingNote and Read it directly and the Korean comes out mangled in your quote. Convert with iconv, then read.
• Traffic is finite — api_call has a per-API daily limit. Don't repeat the same query; when the response is large, trim it with numOfRows-style parameters.
• Search and detail keep returning 0 hits or failing to parse — a portal redesign may have broken the server's HTML parser. Keep the research going by checking the detailUrl directly with WebFetch, but report to the user that the server (server/src/datago-client.ts) needs fixing.
• License-restricted data — if it isn't "제한 없음" ("no restrictions"), get the user's confirmation before quoting it in commercial content.