How to use the Localization Pipeline agent skill (localize)

What the localize skill makes the agent do

A skill is a set of instructions, packaged as a SKILL.md file, that an AI agent loads when you ask for that kind of work. This one covers the full localization pipeline for a game, from finding text that is not ready for translation through to a QA pass before release.

The skill has ten modes: scan, extract, validate, status, brief, cultural-review, vo-pipeline, rtl-check, freeze and qa. Each mode has a narrow job, so you can run a single check or the whole cycle.

• scan finds hardcoded strings and localization anti-patterns in source code. It is read-only.
• extract compares source code against the string table in assets/data/strings/ and drafts new keyed entries with translator context.
• validate checks translations for missing entries, placeholder mismatches, length problems, plural forms, orphaned keys and stale translations. It is read-only.
• status produces a coverage matrix per locale. It is read-only.
• brief writes a translator context document for an external team.
• cultural-review, vo-pipeline, rtl-check, freeze and qa cover cultural sensitivity, voice-over, right-to-left layout, string freeze and the full release QA cycle.

Install the localize skill

Download the localize ZIP from this page. The skill is MIT licensed and its folder contains a SKILL.md file. You can install it in one of two ways.

For Claude Code, unzip the download into ~/.claude/skills/ to make it available in all your projects, or into .claude/skills/ inside one project to limit it to that project. For the Claude apps, open Settings, then Capabilities, then Skills, and upload the ZIP there.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Check what the skill needs before you run it

The skill needs Claude Code or a compatible SKILL.md agent, a game project with source code, and a string table in assets/data/strings/. It also uses a bash-compatible shell, because it calls the yaml-helper.sh configuration script to read settings.

Without a string table or game source code, the skill does not produce useful results. It reports NOT ASSESSED instead. Treat that as a missing prerequisite, not as a clean result.

• Claude Code or a compatible SKILL.md agent
• A game project with source code
• A string table in assets/data/strings/
• A bash-compatible shell for yaml-helper.sh

Trigger the skill with a mode

Invoke the skill with /localize followed by one mode name. If you give no mode, the skill prints its usage list and stops, so pick a mode first. The first run is usually scan, because it shows where the code is not yet ready for translation.

Scan only reports findings. Extract is the mode that proposes changes, and it asks before writing anything. Brief writes a file into production/localization/, named with the locale and date.

• Start with /localize scan to find hardcoded text
• Use /localize extract to draft keys for untracked strings, then approve the diff
• Use /localize validate after translators return files
• Use /localize status for a coverage table per locale
• Use /localize brief when you need a document for a translation vendor

Example workflow for a project preparing a second language

Suppose you are adding French and German to a game that has a string table but few keys yet. Run scan first and review the file paths and line numbers it reports. Strings it flags should be wrapped in a localization function before extraction.

Next run extract. The skill generates new entries following the convention category.subcategory.description, each with a context field that gives the translator the screen, the maximum length and any placeholder meaning. You then approve the diff, and only the new entries are written to assets/data/strings/strings-en.json.

Once translators have supplied files, run validate and then status. Use the status matrix to see what is missing or stale. Run brief for the vendor, and run rtl-check or qa before sign-off.

Safety notes: scripts and network

The skill runs a script. Its configuration is read through yaml-helper.sh using a bash command, and the skill's permissions limit that command to the resolve_config call. Read the script before you install the skill if your environment requires it.

The skill does not use the network. Most modes only read files. The files it can write are the new entries in the string table after your approval and the brief document in production/localization/.

• Runs scripts: yes, through yaml-helper.sh
• Uses network: no
• Read-only modes: scan, validate, status
• Writes only after approval: extract
• Writes a file: brief

Tips for getting reliable results

The skill can only report on code it can find. If it says the code root is unresolved, it reports NOT ASSESSED rather than zero hits, which means a clean report from an unresolved root does not mean nothing needs localizing. Check that the project layout is what the skill expects before you trust the output.

The agent flags issues and drafts context, but it does not replace a professional translator or a native reviewer. Use cultural-review and the translator brief as inputs to human review.

• Keep string keys in the category.subcategory.description format so extraction stays consistent
• Give every string a clear context entry, because validate checks length against it
• Run freeze before translation starts so later source changes show up as freeze violations in status
• Treat NOT ASSESSED as a missing prerequisite, not a pass