While advancing the go-tour-i18n multilingual translation project, I had previously validated a complete ChatGPT-based translation workflow.
This was previously accomplished through:
I completed the full translation validation for the Simplified Chinese zh-CN of A Tour of Go.
However, while advancing the translation for the second language, ja-JP, I discovered a new issue:
A ChatGPT translation workflow that had already been successfully validated could not reliably deliver translation results in a new task.
This article documents this validation process.
The ChatGPT Translation Workflow Validated During the Chinese Phase
go-tour-i18n is not a simple text translation project.
The project uses TranslationUnit as the minimal translation unit.
There are two types of TranslationUnits:
- Page: A complete course page (
.article). - Example: Translatable comments in a complete Go example file.
During the Chinese phase, I validated that:
- ChatGPT can read the retranslation batch exported from GitHub;
- ChatGPT can complete the TranslationUnit translation based on the manifest and input files;
- The translation results can enter the project’s unified process, validation, and release workflow.
Related process:

From the directory structure, it can be seen that:
inputsraw-responsescandidatesvalidationresult.json
These workflow stages already exist.
This indicates that:
The ChatGPT translation results do not remain in the chat text, but have already entered the unified processing workflow defined by the go-tour-i18n project.
Adjusting from GitHub Writes to Artifact Imports
In the early design, I wanted ChatGPT to directly complete:
GitHub batch
↓
ChatGPT
↓
Generate translation results
↓
GitHub write
↓
validationHowever, during actual use, the stability of GitHub operations such as writing a large number of files, committing, and pushing was insufficient.
Therefore, the workflow was adjusted to:
Reference:
“go-tour-i18n Translation Workflow Adjustment: From ChatGPT GitHub Writes to Local Artifact Imports”
GitHub batch
↓
ChatGPT
↓
raw-responses artifact
↓
Local import
↓
process
↓
validationThe purpose of the adjustment:
It was not to reduce ChatGPT’s translation responsibilities, but to minimize the uncertainty caused by the AI directly modifying the repository.

By generating:
chatgpt-ja-JP-008-raw-responses.zipThen executing locally:
go run -mod=readonly ./cmd/tour-i18n retranslation import \
--locale ja-JP \
--batch-id chatgpt-ja-JP-008 \
--archive ~/下载/chatgpt-ja-JP-008-raw-responses.zipThen continuing:
go run -mod=readonly ./cmd/tour-i18n retranslation process \
--locale ja-JP \
--batch-id chatgpt-ja-JP-008To enter the subsequent processing workflow.
Preparing the New ja-JP Batch
This test batch:
chatgpt-ja-JP-009Contains:
unit_count: 2020 Page TranslationUnits:
flowcontrol/1
flowcontrol/2
...
flowcontrol/14
moretypes/1
...
moretypes/6The batch export process was normal.

It can be confirmed that:
- The batch_id is correct;
- The locale is
ja-JP; - The unit_kind is
page; - The unit_count is 20.
Therefore, the current issue is not:
- A batch export failure;
- Missing input files;
- An error in the TranslationUnit definition.
Execution Process in a New Session
The new ChatGPT session was able to complete:
- Identifying the project background;
- Reading the batch;
- Reading manifest.json;
- Confirming the input files.
However, after entering the translation generation phase, no actual result files were produced.

Current status:
manifest.json read
inputs 20/20 read
ja-JP translation files not yet generated
raw-responses not yet generated
zip artifact not yet generatedThat is to say:
The preceding input reading phase was normal.
However:
TranslationUnit
↓
raw-responses file generation
↓
artifact packagingThis phase was not completed.
Attempting to Adjust Project Instructions
To prevent the new session from re-evaluating already validated capabilities, I adjusted the ChatGPT project settings.
I added:
- Verified that ChatGPT can read the retranslation batch;
- Verified that ChatGPT can read the manifest and inputs;
- Verified that ChatGPT can translate Page;
- Verified that ChatGPT can translate Example code comments;
- Verified that ChatGPT can generate raw-responses artifacts.

However, after retesting following the adjustment:
The execution behavior still did not stably recover.
Therefore, it can currently be confirmed that:
The issue is not insufficient project documentation.
Analysis of the Current Issue
This validation cannot prove:
That ChatGPT lacks translation capabilities.
Because:
The 103-page translation during the Chinese phase has already proven that ChatGPT can complete high-quality TranslationUnit translations.
What currently cannot be resolved is:
ChatGPT translation results
↓
Stably generating the files required by the project
↓
Entering the git repository
↓
Continuing process / validationThis complete chain.
Missing Links in the Current Workflow
Current actual status:
TranslationUnit input
↓
ChatGPT
↓
?
↓
raw-responses
↓
go-tour-i18n
process
↓
validation
↓
review
↓
promotionWhere:
?
Indicates that a stable translation result delivery method is currently missing.
This is also the issue that needs to be further resolved at present.
Current Conclusions
This validation yielded several conclusions:
First:
ChatGPT’s translation quality has been validated during the Chinese phase.
Second:
AI translation capability and AI engineering execution capability are not the same issue.
Third:
For long-term maintenance and multilingual expansion projects, it is not enough to only verify that “the model can translate”; it is also necessary to verify:
- How translation results are stably saved;
- How they enter the project repository;
- How they are executed repeatedly;
- How they are automatically validated.
The current go-tour-i18n already has:
- TranslationUnit design;
- Batch management;
- Process workflow;
- Validation workflow.
The next step requires re-evaluating:
How to choose a stable translation result delivery method.
This record serves as a validation of the stability of an AI-assisted engineering workflow.
