Reviving Praderas (Day 9) — translation with method: plan, batch 1, and numbers

Today we formally started English content migration with one principle: coherence before raw speed.

What we shipped today

1. Batch migration plan in .agents/translation-migration-tracker.md: 8 batches (themed work packages) ordered for a coherent site.
2. Batch 1 completed: full Reviving Praderas series paired in English from Day 1 through Day 9 (including this wrap-up), so the series is not half-translated.
3. Glossary updates for recurring naming choices to keep navigation and tone consistent.
4. Translation metadata hardening (Lang + Translation_Key) for Day 1 to Day 7 so language switcher and hreflang behave on complete pairs.

What people mean by AI “context” (without assuming you already know)

Lots of readers hear “context” in AI discussions and reasonably wonder what it is. In practice:

• A model does not magically remember your whole project. It only “sees” what fits in the working window for that chat or task: instructions, pasted files, repo snippets the assistant reads, prior messages, and so on.
• That window has a hard size limit (measured in tokens, chunks of text). If you cram too much in at once—many long articles in one go—the oldest or least central material falls out or gets heavily compressed. That is when detail fidelity drops and consistency errors rise (names, links, tone, cross-chapter alignment).
• So “context overflow” is not just engineer slang: it means quality loss because the model literally no longer has the full text in view that you thought was “there.”

This is different from everyday “context” (situation, intent). Here, context = the material that fits in the model’s window for that pass.

Why we split the migration into batches

The reasons line up with the above and with how we want the blog to read:

1. Context limits — translating the entire archive in one giant prompt would force too many .md files into one window; early chapters fall out as the model works late ones. Batches keep each PR bounded and reviewable.
2. Whole series per batch — avoids an English reader hitting Day 4 in English and Day 5 only in Spanish; narrative order stays intact.
3. Glossary and tone — each batch can lock recurring choices before the next topic cluster.
4. Human review — one enormous PR is hard to audit; several batch-sized PRs preserve traceability.

Ad-hoc page-by-page translation often breaks reader flow: half-finished series, drifting names, unstable voice. Thematic batches plus a living glossary address that deliberately.

How many batches and what timeline (plan as of April 2026)

We defined 8 batches. Full detail lives in the repo tracker; here is the planned timeline (close dates are planning targets; we tick them off as each batch merges):

Order is intentional: series that define the rebuild first, then topic clusters so terminology and tone do not fight each other.

Time note (real wall clock vs traditional localization)

For this batch (plan + batch 1 translations + wrap-up post + PR), human wall-clock time on the author side was on the order of ~20 minutes (review, opening the PR, steering the assistant). That is not “how long the model worked inside the server”—models do not experience calendar time the way you do. The useful number is how much calendar time the workflow took you.

Separately, a localization specialist plus a classic editorial pass for equivalent coverage (terminology consistency and metadata checks) often lands in a rough 8 to 14 hour range depending on review depth and tooling—not a universal law, just an order-of-magnitude anchor.

The lesson for readers who want to understand AI is not “AI translates the whole archive in twenty minutes.” It is that scoping work (batches, documented repo, full series in one batch) keeps context useful and can make your wall-clock time very small when the PR scope is tight and well prepared.

What comes next

The next logical block is Batch 2: the full Control de Tiempo Desacoplado series (13 chapters), following the same rule of not splitting a series in half.