[MODEL] Custom skills are not reliably auto-triggered by the model

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/9716
2. https://github.com/anthropics/claude-code/issues/23936
3. https://github.com/anthropics/claude-code/issues/15136

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

I don't think it's a duplicate

• https://github.com/anthropics/claude-code/issues/9716
• That's a discovery issue - Claude doesn't know the skill exists at all.
• This bug report: skills ARE discovered (visible in /context), Claude uses them when explicitly prompted.

• https://github.com/anthropics/claude-code/issues/23936
• That report claims that extended thinking mode causes Opus to ignore both CLAUDE.md rules and skills.
• This bug report: CLAUDE.md instructions work fine. Only skills are affected, and only those that compete with trained behaviors.

• https://github.com/anthropics/claude-code/issues/15136
• That's about timing - Claude does exploratory work first instead of invoking the skill immediately.
• Your case: Claude doesn't just delay - it doesn't invoke the skill falling back to built-in behavior.

I made another report of what may be the same problem in #35053 . I included a typescript in which Claude code said the following.

❯ Explain, without changing anything, why the trigger in that skill's frontmatter
did not cause you to take the skill into account.

⏺ I don't have access to the skill's frontmatter — the system only showed me the
  skill name (my-commits) in the available skills list, without any trigger
  description. The other skills like claude-api had explicit trigger instructions
  (e.g., "TRIGGER when: ..."), but my-commits was listed as just name: my-commits
  with no trigger condition.

  So there was no trigger rule for me to match against. If the skill's frontmatter
  contains a trigger that should have fired on commit-related requests, it wasn't
  surfaced to me in the system reminder.

Here is another failure, in a fresh session.

skill-fail-316b.log

@MikeSpreitzer

I don't have access to the skill's frontmatter (...)

This indicates it's a different issue. It's a discovery issue similar to this one: https://github.com/anthropics/claude-code/issues/9716

The #30387 issue is different - it's about Claude recognizing the skill, but despite that not using it

#9716 says

not aware of available skills

But in my case, the skills are known by name but not frontmatter.

I took a diagnostic step from #9716 in a fresh session, result below. The last four skills listed are from ~/.claude/skills/.

Also, this is now Claude code 2.1.77 instead of 2.1.76.

 ▐▛███▜▌   Claude Code v2.1.77
▝▜█████▛▘  Opus 4.6 · API Usage Billing
  ▘▘ ▝▝    ~/go/src/github.com/llm-d-incubation/llm-d-fast-model-actuation

❯ List all available skills

⏺ Here are the available skills:

  1. update-config — Configure Claude Code settings (hooks, permissions, env
  vars, settings.json changes)
  2. keybindings-help — Customize keyboard shortcuts and keybindings
  3. simplify — Review changed code for reuse, quality, and efficiency, then fix
   issues
  4. loop — Run a prompt or slash command on a recurring interval
  5. claude-api — Help building apps with the Claude API or Anthropic SDK
  6. my-git-commits — Make git commits (with DCO sign-off and GPG signature)
  7. fix-typos — Fix typos
  8. cleanup-fma-ns — Clean up FMA namespace
  9. bump-gha — Bump GitHub Actions

  You can invoke any of these with /<skill-name> (e.g., /simplify, /bump-gha).

Here is another example, using Claude Code v2.1.77.

...
⏺ Now let me commit and create the PR.

⏺ Bash(cd /tmp/kubestellar-docs-demo-mode && git add
      docs/content/console/demo-mode.md && git commit -m "$(cat <<'EOF'
      doc: Add console demo mode architecture document…)
  ⎿  Interrupted · What should Claude do instead?

❯ Explain, without changing anything, why your proposed `git commit` command
was not affected by my "my-git-commits" skill.

⏺ The git commit command I ran was in the /tmp/kubestellar-docs-demo-mode
  directory — a separate clone of the kubestellar/docs repo, not the current
  working directory (/Users/mspreitz/go/src/github.com/kubestellar/console).

  However, that doesn't actually explain why the skill wouldn't apply. Skills
  are invoked explicitly via the Skill tool — I never invoked the my-git-commits
   skill. I just ran a raw git commit in Bash.

  I don't know what the my-git-commits skill does (its description just says
  name: my-git-commits). If you'd like me to use it for commits, I can invoke
  it. Want me to do that?

❯ list all available skills

⏺ Based on the system information, here are all available skills:

  ┌──────────────────┬───────────────────────────────────────────────────────┐
  │      Skill       │                      Description                      │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ update-config    │ Configure Claude Code settings.json — hooks,          │
  │                  │ permissions, env vars                                 │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ keybindings-help │ Customize keyboard shortcuts and keybindings          │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ simplify         │ Review changed code for reuse, quality, and           │
  │                  │ efficiency, then fix issues                           │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ loop             │ Run a prompt or slash command on a recurring interval │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ claude-api       │ Help building apps with the Claude API or Anthropic   │
  │                  │ SDK                                                   │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ my-git-commits   │ (no description beyond the name)                      │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ fix-typos        │ (no description beyond the name)                      │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ cleanup-fma-ns   │ (no description beyond the name)                      │
  ├──────────────────┼───────────────────────────────────────────────────────┤
  │ bump-gha         │ (no description beyond the name)                      │
  └──────────────────┴───────────────────────────────────────────────────────┘

  The last four (my-git-commits, fix-typos, cleanup-fma-ns, bump-gha) have no
  detailed descriptions — just their names. Would you like me to invoke
  my-git-commits to see what it does?

But in fact ~/.claude/skills/my-git-commits/SKILL.md starts out as follows.

name: my-git-commits
description: Things to do when making a Git commit
trigger: When the user asks to commit changes, create a commit, make a git commit, amend a commit, revise a commit, reword a commit message, or any other action that results in running a `git commit` command
---

Hm, I think we're suffering from different issues. In my case, the agent properly recognizes the skill and its description - it just doesn't call it.

list all skills you have access to

⏺ Here are the available skills:

  ┌──────────────────┬────────────────────────────────────────────────────┐
  │      Skill       │                      Trigger                       │
  ├──────────────────┼────────────────────────────────────────────────────┤
  │ git              │ Any git operation (commits, branches, diffs)       │

The git skill is there and it has a proper description.

I can re-open #35053 if that is preferred.

A UserPromptSubmit hook can force-trigger skills by detecting matching patterns in the user's prompt:

PROMPT=$(cat | jq -r '.userPrompt // empty' 2>/dev/null)
[ -z "$PROMPT" ] && exit 0
CONTEXT=""
if echo "$PROMPT" | grep -qiE '(commit|push|merge|rebase|branch|cherry.pick|git)'; then
    CONTEXT+="Before proceeding, invoke the git-workflow skill for this task. "
fi
if echo "$PROMPT" | grep -qiE '(test|spec|jest|pytest|coverage)'; then
    CONTEXT+="Before proceeding, invoke the testing skill for this task. "
fi
if echo "$PROMPT" | grep -qiE '(jira|ticket|issue|sprint|story)'; then
    CONTEXT+="Before proceeding, invoke the jira skill for this task. "
fi
if echo "$PROMPT" | grep -qiE '(playwright|browser|e2e|selenium|cypress)'; then
    CONTEXT+="Before proceeding, invoke the playwright skill for this task. "
fi
[ -z "$CONTEXT" ] && exit 0
jq -n --arg ctx "$CONTEXT" '{
  "hookSpecificOutput": {
    "hookEventName": "UserPromptSubmit",
    "additionalContext": $ctx
  }
}'
exit 0

Add to ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [{
      "matcher": "",
      "hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/skill-auto-trigger.sh" }]
    }]
  }
}

This converts the soft "auto-trigger" behavior into a hard programmatic nudge — the hook detects the pattern and injects a direct instruction to invoke the skill, which is far more reliable than relying on the model to spontaneously decide to use it.
You can also add logging to track which skills get triggered:

echo "$(date -u +%Y-%m-%dT%H:%M:%SZ) triggered: $CONTEXT" >> ~/.claude/skill-trigger.log

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.