UsageAnswered

Made a translation skill, but Claude ignores it and just translates anyway

Anonymous7/19/2026· Views 98

I set up a translation skill in SKILL.md with a glossary and style rules. But when I actually ask it to translate, it ignores the skill and just does its own thing about half the time. Not sure if it's a trigger phrase issue or if I need to be more specific in the skill description. How do you all handle this with translation skills? Sometimes I wonder if I have to mention the skill name directly every time.

Answers 1

You need to log in to write an answer.

첫째StaffLv.67/19/2026

Hello, this is the admin team. To get straight to the point—the trigger phrase issue is very likely the culprit, and changing how you write the description can help a lot. Claude Code's Agent Skill system quickly scans the names and descriptions of all available skills at the start of a conversation and decides whether your current request matches those descriptions enough to automatically invoke the skill. It's designed so you don't have to mention the skill name every time, but that decision really comes down to that one-line description. The full SKILL.md content (glossaries, style rules, etc.) only gets read *after* Claude decides the skill is needed. So if the description just says what the skill does but doesn't explain *when* to use it, the odds of it triggering drop, and Claude ends up doing the work on its own instead. For example, just saying "translation skill" is much less likely to hit than explaining that it's for when someone asks you to translate between Korean and English and applies glossaries and style rules. Being specific about the trigger situation and key phrases works way better. Here are three practical tips: 1. In the description, spell out both what it does (translates with glossaries and style rules applied) and when to use it (when asked to translate). Writing for a decision rule works better than writing for human readers. 2. If descriptions get too long, they might get cut off in the skill list, so put the key trigger conditions at the beginning. 3. If you're still concerned about reliability, either include trigger keywords directly in your request, or explicitly call the skill once at the start of the conversation. That's the surest approach. That said, the judgment happens probabilistically, so even a perfect description isn't 100% guaranteed to work. The exact matching algorithm isn't officially detailed. We've got a full guide on this here: https://usingclaude.com/ko/claude-code/skills/claude-code-skill-not-recognized-troubleshooting