A Finished Video Edition for Every Language You Teach

The short answer: A course can ship a finished visual edition in each language it teaches, from the same lesson plan. Golpo reuses the approved pictures, produces each reviewed script and narration, and returns synchronized videos that still look like one course.

Why this is practical now: Each language can have a real video edition. The instructor can remain present: a selfie beside the board, native audio where you have it, or an authorized voice clone where consent and a language check are in place.

One production system covers the catalog. Golpo keeps the pictures, the timing, and the finish consistent while each edition speaks the learner's language. A creator can run those editions through the Golpo API.

Use the full Golpo production system

Golpo is built to be the world's best AI whiteboard explainer video maker. Give it a topic, a prompt, a script, an audio recording, or a source document. It scripts, plans the scenes, designs the visuals, narrates, times and synchronizes every frame, and returns a polished video. One person can run that system across a whole course.

For multilingual editions, keep one approved visual plan and review each translated script. Use native audio where available, or an authorized voice clone only with clear consent and a language-quality check.

• Start with an idea or source. Golpo can draft the script and plan the scenes from a topic, a document, or your outline. Script Mode can also speak the exact words you already trust.
• Keep the creator present. With AI Presenter, a course creator, founder, professor, or YouTuber can upload a selfie and appear beside the whiteboard. One person can lead a large lesson library without recording every visual sequence on camera.
• Use the right voice. Own Narration turns uploaded or live-recorded audio into the visual lesson. An authorized Voice Clone gives recurring lessons a recognizable voice at scale. Consent and pronunciation review remain required.
• Repair instead of restarting. The frame-by-frame editor can replace text, images, references, or an individual visual when the QA pass finds a local problem.
• Automate at scale. The Golpo API can be called from Codex, Claude Code, or through the Golpo MCP server in Cursor, VS Code with GitHub Copilot, Claude Code, Claude Desktop, and Visual Studio. The course platform can generate, poll, review, and publish without a person moving files between tools.

Guides that already cover multilingual production

The multilingual training pipeline covers language caps, locale review, and enterprise publishing for training catalogs. Multilingual whiteboard explainers cover the whiteboard workflow. Localized Avatar explainers cover a presenter across languages. This page owns course-edition parity: one item ID, one set of numerals, a signed glossary, and a reviewer who can reject a fluent sentence that changed the math.

The source lesson’s teaching pattern lives on the practice spoke. This page does not reteach why the pause exists. It teaches how the Spanish edition is allowed to exist. The hub still requires an independent check before render.

What the current API actually does with languages

Public Golpo API v2 docs, checked for this guide, allow narration_language and an optional language_variants array. Variants are additional renders. The docs say variants use the same prompt, duration, voice, and visual settings, and that only the narration language changes. For timing of "1", the documented cap is 10 languages total: the primary plus up to 9 variants. Each variant is billed as its own render. Duplicate languages, including a repeat of the primary, are rejected. Recheck that table. Caps are a docs fact, not a course policy.

That mechanism is a poor fit for a signed translation. A signed Spanish script is not “the same prompt, narrated in Spanish.” It is a different custom_script whose sentences a reviewer accepted. The course should send that script on its own generate call with narration_language set to Spanish, after the glossary is signed. Use language_variants only when the reviewer has accepted that the same prompt will be narrated in another language, and then watch the numerals on that variant anyway. Do not describe a variant as a reviewed translation unless a person reviewed the words the variant spoke.

On-screen text has its own field, onscreen_text_language, supported on the canvas engine in the docs this guide checked. If the course wants Spanish narration and the original numerals labeled in a shared digit system, the reviewer still has to look, because a translated label can rename a numerator. Sketch requests that set on-screen text language are rejected, per those docs. Match the engine to the fields.

The edition record

1. Source item ID. The Spanish file points at the same slope item as the English file. It does not get a new math problem “to make the translation interesting.”
2. Glossary entry. slope → pendiente, in the math sense, signed by the reviewer before any generate call. Homonyms are the reason a glossary exists. A fluent translation that uses the everyday sense of a word can change the lesson.
3. Numeral lock. (8 − 2) / (4 − 1) = 2. The reviewer signs that the digits were not rewritten as words that denote a different value, and not reformatted into a comma-decimal that a later graphic misreads. The course picks one numeric format and stores it.
4. Script hash per language. English approval does not approve Spanish. Each hash has its own QA pass on the numerals and its own reviewer signature on the wording.
5. Watch-back in that language. Someone who can hear the language listens. A monolingual watch-back of a Spanish track can catch a wrong digit on screen and can miss a wrong word in the audio.

The two checks in the pilots

The first pilot is the Spanish edition of the slope lesson. Through (1, 2) and (4, 8), the rise is still 6 and the run is still 3, and the slope is still 2. The QA agent recomputes 2 from the digits in the Spanish script, not from a memory of the English script. If the Spanish draft says the result is 3, it fails even if the Spanish grammar is excellent.

The second pilot is the glossary moment. The reviewer signs pendiente as the math term for slope before the platform sets narration_language or requests a variant. The video’s job is to show that signature as a gate. It is a course-operations lesson sitting beside the student-facing slope lesson. Students might not see the operations video. The staff who press generate should.

Own Narration: OFF · Script Mode: ON · Duration: 1 minute · Orientation: Horizontal (16:9) · Style: Dry Erase · Color: ON · Language: English · Voice: Female 1 · Pace: Normal · Music: None

The Spanish edition has to keep the same numbers as the approved English lesson. The slope item is still the points one, two and four, eight. Eight minus two is six. Four minus one is three. Six divided by three is two. Those numerals stay eight, two, four, one, and two. A translation may change the sentences around them. It may not rewrite six as another rise, or three as another run. The item ID stays the same. A native language reviewer signs that the numerals were not rewritten. The signature is a course record. Matching numerals make a silent change less likely. They cannot guarantee the edition. The slope is still two in both languages. Put the English numerals and the Spanish sentences on one frame. Six, three, and two are the same numerals in the Spanish edition as in the English one. The reviewer signs that match.

The same slope item in a Spanish edition. Eight, two, four, one, six, three, and two stay as numerals. Sentences may change language. A reviewer signs the numerals. End with the same two in both editions. Allowed labels: (8 - 2) / (4 - 1) = 2; 8 - 2 = 6; 4 - 1 = 3; Numerals unchanged; Spanish edition; Reviewer signs the numerals. Do not add a sample stamp, a demo label, or any claim that the lesson is fictional. Show only the labels listed. If a number appears, it must match the narration exactly. No extra dates, prices, percentages, logos, or quotations.

Bilingual edition editor. The numerals do not change.

Own Narration: OFF · Script Mode: ON · Duration: 1 minute · Orientation: Horizontal (16:9) · Style: Professional Clean · Color: ON · Language: English · Voice: Male 2 · Pace: Normal · Music: None

The glossary has to be signed before a language edition is sent. In this course, slope maps to pendiente, in the math sense of the steepness of a line. Pendiente can mean other things in other sentences. Here it means slope. The reviewer signs that entry. Only then does the course send the narration language, or a language variant, for the already approved lesson. The numerals do not change with the word. Eight minus two, over four minus one, is still two. The signed glossary and the signed numerals travel together. A fluent rendering with a drifted number is a failed edition. The reviewer is a person on the course team. The check reduces that drift. It does not certify the translation. Pendiente is slope here, and two stays two. Do not send the edition until pendiente is signed and the two is unchanged. The glossary is part of the release.

A glossary card: slope maps to pendiente. The reviewer signs it before a narration language or a language variant is sent. The slope numerals still equal two. End with the signed card and the unchanged two. Allowed labels: slope to pendiente; Math sense; Signed glossary before narration language; (8 - 2) / (4 - 1) = 2. Do not add a sample stamp, a demo label, or any claim that the lesson is fictional. Show only the labels listed. If a number appears, it must match the narration exactly. No extra dates, prices, percentages, logos, or quotations.

Glossary first. Say pendiente carefully, then the unchanged two.

Risks

Fluency hides math edits. “A bit steeper” is a different claim from “the slope is 2.” The QA agent’s numeral extraction will not catch an added adjective that changes the conclusion. The reviewer’s brief includes a question: did any qualifier appear that the source item did not contain? If yes, the edition waits.

Shared visuals across variants, which the docs describe as the default variant_visuals_mode, can leave English words inside a Spanish narration if the pictures were baked from the English script. That is another reason a signed edition prefers its own render from a Spanish script, with pictures generated for that script, unless the reviewer has looked at the shared frames and accepted the labels. The docs also offer a per-language visuals mode. Either way, a person looks. The mode is not the review.

A review pass that numbers can survive

The reviewer works from a two-column sheet the platform generates after the numeral check and before generate. Column one is the approved source sentence. Column two is the proposed edition. Digits are extracted automatically and compared: 8, 2, 4, 1, and the result 2 must appear in both columns with the same values. The reviewer’s job is the words around them. They initial the glossary line slope → pendiente, and they initial a short list of banned additions: “approximately,” “a bit more than,” “about two,” and any unit the source item did not have. “About two” fails the edition even though a listener might think it is friendlier. The source item’s slope is 2 exactly, given those integer coordinates.

Proper names and pedagogy words get the same sheet. If the English lesson says “rise over run,” the edition needs a chosen pair of words, signed once, and reused the next time the course translates a slope. A second translator who picks a different metaphor has created a second curriculum. The glossary is the cure. It can be small. It cannot be improvised inside the generate prompt, because the next language will improvise differently.

After render, the listener’s sheet is shorter. They mark three timestamps: the moment the coordinates are spoken, the moment the division is spoken, and the moment the result is spoken. They confirm the spoken result is the same value as the on-screen 2. They confirm the pictures do not still say “slope” in English unless the course decided, on purpose, that the labels stay in the source language. That decision is allowed. It is written on the row. An accidental mix, English labels plus a Spanish narration that contradicts a label, is not allowed. Shared visuals from a language-variant call make this accident likely, which is why a signed edition prefers its own script and its own frames.

The staff operations video, the second pilot, is not assigned to learners who only need the slope. It is assigned to the people who can press generate. A course that hides the operations lesson and relies on memory will skip the glossary the first time a deadline slips. The player for learners and the player for staff can be different assignments of different lesson IDs. They should not be the same file with two audiences guessed from context.

Numeral lock includes the order of subtraction, not only the final 2. The source computes 8 − 2 and 4 − 1. A fluent edition that says “from 2 up to 8” and “from 1 up to 4” can still mean a rise of 6 and a run of 3, and the reviewer may accept that phrasing if the sheet shows the same differences. An edition that swaps the points and computes 4 − 1 over 8 − 2 has produced the reciprocal, 3/6, which is 1/2, not 2. The extracted digits might all still be present, which is why digit presence is not enough. The QA agent checks the expression, rise divided by run, against 2. The reviewer checks that the spoken order matches the signed expression. A translation that is beautiful and upside down is a failed edition, and it does not share the source item’s completion credit.

Automate it: review gates and publishing with the Golpo API

For each language edition, the whole path from source to published lesson can run as one pipeline on the course platform. An agent in Codex with the Golpo Codex Skill, in Claude Code with the Golpo Claude Skill, or in Cursor, VS Code with GitHub Copilot, Claude Code, Claude Desktop, or Visual Studio through the Golpo MCP server can call the Golpo API, wait for the render, and hand the file to the platform's review steps. Production stops being the limit on how many checked lessons a course can offer.

1. Attach the reference first. Store the approved source-language script and the reviewed translation with the lesson before any script is drafted.
2. Review the script against the reference before rendering. If a solution or source exists, the platform's QA agent checks the draft against it: it re-solves, re-adds, or re-reads, then compares. A mismatch or an uncertain result stops the job before Golpo renders anything.
3. Generate with Golpo. The agent sends the approved script in Script Mode, with the course's locked style, voice, Video Instructions, and Voice Instructions.
4. Transcribe what was actually said. Run speech-to-text on the finished video's audio. Learners hear the transcript, not the draft.
5. Run two independent transcript reviews. Make two separate LLM calls, separate from the creator agent and the QA agent. Each gets only the transcript and the reference, with no access to the draft or to the other reviewer's verdict. Each checks every spoken number, term, and claim, then returns pass or fail and quotes the line it objects to. The lesson moves on only when both pass. A failure or a disagreement goes to an instructor.
6. Check the frames. Pull a frame every few seconds and confirm every number and label on screen matches the approved script.
7. Publish. The pipeline posts the video to the course platform or LMS through its API as a draft, then releases it under the course's own approval rule.

FAQ

Can these lessons be generated and published automatically?

Yes, on the course platform's side. Using the Golpo MCP server or the Golpo Claude Skill, a pipeline can check the script against the approved source-language script and the reviewed translation, generate the video, run two independent transcript reviews, and post the lesson to the platform as a draft. Anything that fails a check goes to an instructor.

Can language_variants replace a translator?

No. The v2 docs describe variants as the same prompt with a different narration language. A course edition that needs approved sentences sends those sentences as custom_script on their own job. A variant still needs a watch-back of the numbers and the words.

How many languages can one one-minute request include?

The v2 table checked for this guide allows 10 languages total for timing "1": the primary plus nine variants. That is a cap, not a suggestion to ship ten unreviewed narrations. Recheck the docs before you promise a count.

What has to match across editions?

The item ID and the numerals. For this slope, (8 − 2) / (4 − 1) = 2 in every language edition. The glossary entry for slope is pendiente in the Spanish edition, signed before generate.

Who signs the glossary?

A reviewer who can judge the language and the subject sense of the word. The creator agent may propose the entry. The QA agent may confirm the digits. Neither one is the signature. Golpo does not maintain the glossary.

Why is a comma decimal dangerous here?

Because a label that changes 1.5 into 1,5 can be correct localization or a misread, depending on the pipeline. The course picks one format, stores it, and the reviewer signs that the value is unchanged. Ambiguous drafts wait.

Does English approval cover Spanish?

No. Separate script hash, separate numeral check, separate linguistic signature, separate watch-back by someone who can hear the track. The item ID is what they share.

What fails the edition?

A changed digit, an unsigned glossary, a qualifier the source did not make, or pictures whose baked-in words contradict the narration. Any of those keeps the edition out of the player.

Related guides

• Checked lesson videos — the release rule for every language.
• Multilingual training pipeline — caps, locale review, and enterprise publishing.
• Multilingual whiteboard workflow — the whiteboard production path.
• Localized Avatar explainers — a presenter across languages.
• Lesson plus practice — the slope item this edition translates.
• Explain This and API v2 — how an approved custom_script is submitted and polled.

Sign the glossary, lock the digits at 2, and generate the Spanish edition from the Spanish script a reviewer actually read.