Trueprofit Blog Multilingual Tabs Triggers

TrueProfit multilingual tab trigger repair + link localization (ES / DE / FR)

A blog Google Doc has tabs: English (tab 1, often titled "article"), then Spanish, German and French versions. The English-only trueprofit-blog-triggers skill preps tab 1. This skill does two things to the three translated tabs:

1. Repair / place CMS triggers — so the n8n multilingual publisher picks up correct Image (sentence note): and Content Highlight lines in every locale.
2. Localize internal links (via update_links.py) — rewrites trueprofit.io/blog/<slug> links and Shopify app utm_campaign tags for each language.

Run the English skill first. This skill reads the English tab's image triggers to learn the shared image URLs (in order), so tab 1 must already be triggered.

Two image modes: repair vs. placement

Translated tabs are produced by translating the English markdown, so they may arrive with trigger lines already present but stale (image triggers split across two lines, old/test URLs & alts, Content Highlight: with a colon), and they have no embedded image objects. The skill auto-picks per tab, per the image-trigger-line count vs. the English image count:

• REPAIR (counts match). The tab carries one image-trigger line per English image, in order — repair each in place: rewrite to canonical form, swap in the English URL by order, translate the alt, drop the stale second URL line.
• PLACE (counts differ → tab is out of sync). The tab is missing its trigger lines (or they were reset). The skill removes whatever stale lines exist and mirrors the English tab: every English image — and every English Content Highlight — is placed at the same paragraph position as English. For each English image it (1) maps the image's English heading → the translated heading (by order, confirmed with cognate keywords that survive translation — competitioncompetencia, organicorganica, tofutofu, USBUSB), then (2) inserts the trigger after the translated paragraph at the same ordinal — i.e. after the translated equivalent of the English paragraph it follows, not grouped at the section end. Pure paragraph counting drifts when a language merges or splits a paragraph, so within a small window around the expected ordinal it picks the translated paragraph whose content best matches the English anchor paragraph (cognate-token overlap), falling back to the ordinal. An image whose English heading only matches a closing-region heading (Final Thoughts / FAQ), by ordinal or a coincidental keyword, is redirected to the end of the body (e.g. a closing CTA lands before Final Thoughts, never inside the FAQ). The dry-run prints, per image, the translated heading and the paragraph it lands after, with keyword / ordinal / fallback confidence, plus STRUCTURE WARNINGS for any anchor not confirmed by a shared heading keyword — review them.

What it does per language tab

Trigger labels stay in English in every tab (Content Highlight, Image (sentence note):, Alt is) because the n8n parser keys on them — only the alt text is translated.

How to run it

Run everything from this skill's scripts/ directory. Auth is identical to the English skill (GOOGLE_TOKEN_JSON env var → token.json → first-time OAuth); see references/setup-google-api.md. The English skill's token works here.

Step 1 — see the English images (what to translate)

python gdocs_ml_triggers.py --doc "<DOC>" --dump-en

Lists the English tab's image URLs and English alts, in document order, plus which tab mapped to each language (check that mapping).

Step 2 — translate the alts (you, the agent)

Translate each English alt into ES / DE / FR following the TrueProfit localization conventions (see trueprofit-blog-localization: faithful translation, brand names kept in English, no em dashes). Avoid ", (, ) in alt text — the n8n parser stops alt at those characters. Save alts.json:

{
  "es": ["alt es para imagen 1", "alt es para imagen 2"],
  "de": ["alt de für Bild 1", "alt de für Bild 2"],
  "fr": ["alt fr pour image 1", "alt fr pour image 2"]
}

One entry per image in document/English order. A language with fewer entries falls back to the English alt for the rest (the run prints a note). If there are no images, skip alts.json — the Content Highlight repairs still run.

Step 3 — dry-run, review, apply

python gdocs_ml_triggers.py --doc "<DOC>" --alts alts.json --dry-run   # preview all 3 tabs
python gdocs_ml_triggers.py --doc "<DOC>" --alts alts.json             # apply

Limit to one language with --lang es (repeatable: --lang es --lang fr).

The dry-run prints, per tab: how many image trigger lines and Content Highlight lines it found (vs the English image count), Quick Recap / FAQ presence, and every planned change (replace / insert) with its reason.

Always dry-run first. The header shows each tab's image mode (REPAIR or PLACE) and, in PLACE mode, the translated heading and the paragraph each image lands after, with a keyword/ordinal/fallback confidence tag — review that mapping before applying, especially any ordinal rows (matched by position, not a shared word) and the STRUCTURE WARNINGS block (any anchor not confirmed by a shared heading keyword — the tell-tale of an English/translated heading mismatch). PLACE mode assumes the translated article has the same sections, in the same order, with the same paragraph breaks as English (true of a faithful translation); paragraph drift from a merge/split is auto-corrected by content match, but the dry-run is where you catch a genuinely missing or reordered section.

Step 4 — localize internal links (optional but recommended)

After the triggers are applied, run update_links.py to rewrite internal links in the language tabs:

python update_links.py --doc "<DOC>" --dry-run   # preview all changes
python update_links.py --doc "<DOC>"             # apply

What it rewrites:

Both operations are idempotent — already-localized links (/es/blog/… or utm_campaign=es-…) are silently skipped.

Tab matching: update_links.py identifies language tabs by exact title match ("Spanish Version", "German Version", "French Version"). Tabs not carrying one of those exact titles are ignored. If the doc uses different tab names, the script will report 0 updates — check that the tab titles match exactly.

Updating the slug list: MULTILINGUAL_SLUGS is a Python set inside update_links.py. To add a new localized blog post, add its slug (the part after /blog/) to that set and commit the file.

Idempotent & plain-text

• Idempotent. A line already in canonical form is detected and skipped — re-runs make no change.
• Plain text. Every repaired/added trigger line is forced to NORMAL_TEXT with bold/italic/underline cleared, per tab, after editing.
• Edits are applied descending by position (replace = delete range + insert), so indices stay valid within each tab.

Resetting

python gdocs_ml_triggers.py --doc "<DOC>" --reset --dry-run   # preview removals
python gdocs_ml_triggers.py --doc "<DOC>" --reset             # remove canonical triggers from ES/DE/FR tabs

--reset only touches the language tabs, never the English tab, and removes canonical trigger lines (single-line Image (sentence note): http… and Content Highlight). Run a normal pass first if the tab still has the stale two-line form, so they become canonical and removable.

How tabs are identified

Tabs are matched by full language name in the title (case-insensitive contains): english, spanish/español, german/deutsch, french/français. Bare 2-letter codes are deliberately NOT used (e.g. "en" hides inside "french"). If a title doesn't match (the English tab is often titled "article"), it falls back to position (tab 0 = en, 1 = es, 2 = de, 3 = fr). The dump/dry-run header prints the final mapping and the full tab list — check it before applying.

How the code is organized

• scripts/detect_ml.py — image repair planner (plan_repairs, parse_existing_triggers), the Content-Highlight planner (plan_ch: keep backed, delete orphans) and the localized recognisers (LANG_TERMS / compile_recognisers, used by detect) — callout now also covers Your Takeaway-style labels, and a final_thoughts recogniser feeds the end-of-body fallback.
• scripts/place_ml.py — the paragraph-anchored, English-mirroring placement engine used in PLACE mode. english_image_anchors / english_ch_anchors record, per image / Content Highlight, the heading it sits under, the body-paragraph ordinal it follows/labels, and that paragraph's text. align_headings maps English → translated headings by cognate keywords (tokens lowercased, accent- stripped, matched on a shared 4-char prefix; monotonic ordinal fallback). plan_placements / plan_ch_placements then insert each trigger at the same paragraph ordinal, using choose_anchor (cognate-token overlap in a ±2 window) to correct translation drift; CH placement snaps to the translated callout label, and — when the English Content Highlight labels a formula (= present) — to the in-window translated line that also carries = and is not a URL. This = snap keeps a formula highlight anchored correctly in languages whose formula terms don't cognate-match English (e.g. German compounds like Nettogewinn), where token scoring would otherwise drift onto a nearby further-reading line whose English URL slug (…/gross-profit-vs-net-profit) coincidentally repeats the formula's words. body_end_index is the closing-region landmark for the CTA fallback.
• scripts/gdocs_ml_triggers.py — the orchestrator (tabs, modes, structure-warning guard, batchUpdate).
• scripts/update_links.py — standalone link-localization script. Walks every text run in the ES / DE / FR tabs, rewrites trueprofit.io/blog/<slug> links to their localized equivalent for slugs in MULTILINGUAL_SLUGS, and prefixes utm_campaign values on apps.shopify.com/trueprofit links. Matches tabs by exact title ("Spanish Version", "German Version", "French Version"). Applies changes in batches of 400 (Docs API limit is 500). Run independently after the trigger step; supports --dry-run.
• scripts/test_detect_ml.py — python test_detect_ml.py runs the whole suite (detection, repair, placement, and request-builder ordering regressions).

Add-missing-Content-Highlight detection rules:

• Formula: the localized formula word (fórmula/Formel/formule) anywhere in a line whose next line contains = (the = is the real discriminator, so culinary "fórmulas de tofu" or a footnote legend like "* Estimado = …" don't fire).
• Callouts: localized Pro tip / Note keyword + colon (French space-before-colon allowed), so prose never triggers.

Repairing a carried-over Content Highlight: line doesn't need these — it just strips the colon. Detection only adds a Content Highlight that's missing entirely.

Scope and assumptions

• Operates on the ES / DE / FR tabs; the English tab is read-only (source of the shared image URLs and their heading anchors, by order).
• PLACE mode assumes the translated article keeps the same sections in the same order as English (true of trueprofit-blog-localization output and of any faithful translation). It anchors to headings, so the products must be real headings in both tabs (they are, in TrueProfit listicles). Review the dry-run's placement table.
• The signed-in Google account needs edit access to the doc.