Check Metadata Style

Audit a dataset's user-facing text against OWID's Writing and Style Guide. Flags fields that break the rules and offers to rewrite them.

The text itself is usually authored in the garden .meta.yml (description_key/WYSK, description_short, title, display.name) and inherited by grapher; a smaller share is set or overridden in the grapher step. This skill reads the resolved metadata off the built grapher dataset, so it covers both — you point it at the grapher step, and any fix it proposes goes back to whichever layer authored the field.

Rules live in STYLE_GUIDE.md next to this file — a committed snapshot of the OWID Notion page. The file records a Last synced from Notion date in its header; the skill checks that date on every run and refreshes the snapshot from Notion when it is more than two weeks old (see step 1). Refreshes are committed via a PR.

When to use

• After editing a .meta.yml under etl/steps/data/ — garden or grapher.
• When the user asks to check chart copy / titles / subtitles / descriptions / WYSK against the style guide.
• As part of pre-PR QA for a dataset update.

Scope

Current dataset only. Ask for the step path if it's not obvious from context. Do not walk all active steps — keep the skill focused on the one dataset the user is working on.

Read the grapher step even when the edit was to a garden .meta.yml (e.g. you edited etl/steps/data/garden/un/2026-04-08/child_labor_report.meta.yml → read etl/steps/data/grapher/un/2026-04-08/child_labor_report). Only the grapher dataset carries the resolved, inherited text a reader actually sees. If the grapher build is stale relative to your garden edit, rebuild it first (step 3) or you will be auditing the old copy.

---

Implementation

1. Check the guide is fresh

Read the > Last synced from Notion: YYYY-MM-DD line in the header of STYLE_GUIDE.md.

• If the date is less than two weeks before today, skip to step 2 — no fetch needed.
• If the line is missing or the date is two weeks old or more, refresh the snapshot, trying in order:

1. Notion MCP. Load the Notion tools with ToolSearch (query +notion fetch) and fetch the Notion page. Use the result only if it looks complete: all sections present (Branding, Capitalization, Grammar and syntax, Short citations for charts, Writing in GDoc, Descriptions, Dates, Punctuation, Spelling) with the ❌/✅ examples intact. Note: authorizing the connector mid-session doesn't help — MCP tool lists are fixed at session start (subagents inherit the same registry, and resuming the session doesn't refresh it either), so a newly authorized connector is only reachable from a brand-new session.
2. Manual export. If the Notion MCP is unavailable or unauthenticated, or the fetched content is incomplete or mangled, ask the user to download a markdown export of the page (open the Notion link → ••• menu → Export → format Markdown & CSV) and provide the file path.
3. Skip. If neither works, warn the user that the committed guide may be stale (state its Last synced date) and continue the audit with the committed copy. Never block the audit on the refresh.

On a successful fetch/export:

• Convert the content to the existing file's structure (same heading hierarchy, ❌/✅ example formatting) and keep the 3-line header block.
• Preserve sections marked NOTE: local addition — rules agreed in PRs that aren't on the Notion page yet. Their absence from the fetched content is not a removal; drop one only when Notion covers the same rule (then remove the note too).
• Diff against the committed copy and summarize any rule changes to the user in-conversation.
• Rewrite the body only if the content changed; stamp today's date on the Last synced from Notion: line either way.
• Remind the user that the modified STYLE_GUIDE.md should be committed via a PR. Do not commit or push as part of this skill.

Plain WebFetch/curl does not work for the refresh: app.notion.com serves a JavaScript shell with no page content (verified 2026-07-07). Don't waste time trying it.

2. Read the Writing and Style Guide

Read STYLE_GUIDE.md in this skill's folder. That file is the source of truth the skill evaluates against — do not fall back to memory, and do not invent rules that aren't in the file.

If the file is missing, rebuild it from Notion via the refresh flow in step 1 (treat it as stale).

3. Collect user-facing strings from the step

Prefer the rendered (post-Jinja) metadata from the built catalog, so template artifacts are included in the check.

Accept the step path in any of these forms and normalize it before loading:

• etl/steps/data/grapher/un/2026-04-08/child_labor_report (matches the Scope example)
• data://grapher/un/2026-04-08/child_labor_report
• grapher/un/2026-04-08/child_labor_report

.venv/bin/python -c "
from etl.paths import DATA_DIR
from owid.catalog import Dataset

raw = '<step path as the user gave it>'
step_path = raw.removeprefix('data://').removeprefix('etl/steps/data/').strip('/')
# step_path is now '<channel>/<namespace>/<version>/<dataset>', e.g. grapher/un/2026-04-08/child_labor_report
ds = Dataset(DATA_DIR / step_path)

import json
rows = []
for table_name in ds.table_names:
    tb = ds[table_name]
    for col in tb.columns:
        m = tb[col].metadata
        entry = {'table': table_name, 'variable': col, 'fields': {}}

        def put(key, val):
            if val:
                entry['fields'][key] = val

        put('title', getattr(m, 'title', None))
        put('description_short', getattr(m, 'description_short', None))

        dk = getattr(m, 'description_key', None) or []
        if isinstance(dk, str):  # modern format: one free-form markdown string
            dk = [dk]
        for i, v in enumerate(dk):
            put(f'description_key[{i}]', v)

        display = getattr(m, 'display', None) or {}
        put('display.name', display.get('name'))

        pres = getattr(m, 'presentation', None)
        if pres is not None:
            put('presentation.title_public', getattr(pres, 'title_public', None))
            put('presentation.title_variant', getattr(pres, 'title_variant', None))
            put('presentation.attribution', getattr(pres, 'attribution', None))
            put('presentation.attribution_short', getattr(pres, 'attribution_short', None))
            gc = getattr(pres, 'grapher_config', None) or {}
            put('presentation.grapher_config.title', gc.get('title'))
            put('presentation.grapher_config.subtitle', gc.get('subtitle'))
            put('presentation.grapher_config.note', gc.get('note'))

        # The chart footer resolves presentation.attribution > origin.attribution >
        # origin.producer (year), so the visible credit usually comes from the origin.
        for i, o in enumerate(getattr(m, 'origins', None) or []):
            put(f'origins[{i}].attribution', getattr(o, 'attribution', None))
            put(f'origins[{i}].producer', getattr(o, 'producer', None))

        if entry['fields']:
            rows.append(entry)

print(json.dumps(rows, indent=2, ensure_ascii=False))
"

Fields checked:

Field	Why it matters
title	Variable short title; shows in the catalog and some chart views
description_short	One-liner rendered under chart titles
description_key[i]	Key information on chart data pages (markdown string, or legacy bullet list)
display.name	Series label in chart legends
presentation.title_public	Public-facing chart title
presentation.title_variant	Disambiguator ("Historical", "WHO estimate", …)
presentation.attribution	Full source credit under the chart (producer – data product (year))
origins[i].attribution	Where that credit usually comes from — the footer falls back to it when presentation.attribution is unset
origins[i].producer	The last fallback: with no attribution set, the footer renders producer (year)
presentation.attribution_short	Short source credit under the chart
presentation.grapher_config.title	Overrides chart title when set
presentation.grapher_config.subtitle	Chart subtitle
presentation.grapher_config.note	Chart footnote

Fields deliberately skipped:

• description_from_producer — verbatim text from the source, not OWID copy.
• citation_full — follows the producer's requested citation, so its wording and capitalization are theirs, not ours to restyle. (Dash typography is the one exception: a spaced hyphen separating producer from data product is an en dash everywhere, per the style guide.)
• unit, short_unit, processing_level, internal names — not user-facing prose.
• description_long, description_processing — technical, de-prioritized (re-enable later if needed).

Fallback if the dataset isn't built:

If DATA_DIR / step_path does not exist, parse the .meta.yml directly with etl.files.ruamel_load and pull the same field names from the tables → <name> → variables → <var> tree.

Origin fields (origins[i].attribution, origins[i].producer) live in the snapshot .dvc, not the .meta.yml, so this fallback does not cover them — say so in the report.

Parse the garden .meta.yml, not just the grapher one. Most datasets author their user-facing text in garden and have a thin or absent grapher .meta.yml, so reading only the grapher layer here finds nothing and reports a clean bill of health on unaudited text. Read both (etl/steps/data/garden/<ns>/<version>/<dataset>.meta.yml and the grapher one if it exists) and note that grapher values override garden ones on the same field.

Warn the user that Jinja templates (<<var>>, {definitions.xxx}, <%- ... -%>) and garden→grapher inheritance are not resolved in this fallback path, so template-generated violations will be missed. Suggest building the step first:

.venv/bin/etlr grapher/<namespace>/<version>/<dataset> --grapher

Drop --only here on purpose: when the catalog is missing, upstream meadow/garden outputs are usually missing too, and --only would skip them and fail on missing inputs.

4. Evaluate each string against the guide

Claude reads STYLE_GUIDE.md and checks every collected string. Keep the evaluation rule-driven: cite the specific section/heading of the guide, not a generic "doesn't sound right".

Focus on rules the guide actually states (e.g. sentence case vs. title case, acronym expansion on first use, number/unit formatting, banned phrases, punctuation, tone). Do not invent rules the guide doesn't cover.

Report format, one block per violation:

[<table>.<variable>.<field>] "<offending text>"
  Rule: <section name / short rule>
  Why:  <one line rationale>
  Fix:  "<suggested rewrite>"

Group the blocks by variable for readability. End with a summary count.

Exception: when the producer explicitly asked for a specific short citation, schemas/definitions.json says to follow it. Report a matching origins[i].attribution as producer-prescribed rather than as a violation, and don't offer a fix.

If no violations are found, say so and list the fields that were inspected — the user should know what was checked, not just that nothing came up.

5. Offer to fix

Match the pattern in check-metadata-typos §4–5:

• Fix all — apply every suggested rewrite.
• Review each — step through one at a time, user confirms/rejects/edits each.
• Cancel — exit without changes.

Apply fixes to the .meta.yml file with ruamel_load / ruamel_dump so comments and key ordering are preserved (see CLAUDE.md → YAML Editing):

from etl.files import ruamel_load, ruamel_dump
data = ruamel_load(meta_yml_path)
# ...edit the tables → variables → <var> → <field> tree...
with open(meta_yml_path, 'w') as f:
    f.write(ruamel_dump(data))

Origin fields are set in the snapshot .dvc (meta.origin.*) — or, occasionally, by the step's own code after it loads the snapshot. Check which before editing, and edit that one. Producer-prescribed attributions stay out of Fix all.

If a violation only shows up in the rendered output because of a Jinja definition (e.g. the issue is inside {definitions.foo}), flag it for manual fix — don't auto-rewrite the definition without asking.

6. Verify

After fixes:

1. Rebuild the step if metadata text was changed in a way that affects rendering:

   .venv/bin/etlr grapher/<namespace>/<version>/<dataset> --grapher --force --only
   

2. Re-run steps 3–4 of this skill on the same step. Expect zero violations.
3. Run make check to confirm no lint/format regressions from the YAML edits.

---

Notes

• No persistent output. Analysis results stay in-conversation: no report .md, no scripts under scripts/. The only persistent file tied to this skill is STYLE_GUIDE.md (the committed rulebook).
• Current step only. If the user asks to audit the whole catalog, say the skill is scoped to one step and suggest running it per dataset.
• Keep STYLE_GUIDE.md in sync with Notion. The header's Last synced from Notion date is checked on every run (step 1); when it is more than two weeks old the skill refreshes the snapshot — Notion MCP first, manual markdown export as fallback — and the refresh is committed via a PR. Between syncs the skill stays deterministic and offline-capable by always auditing against the committed file.
• Archive-aware. If the provided step path is under dag/archive/*.yml, point that out and confirm the user still wants to check it.
• Don't hallucinate rules. If STYLE_GUIDE.md doesn't say something, don't flag it. Prefer false negatives over false positives.