Model ignores auto-invoke skill instructions, requires redundant memory/feedback to comply

Description

When a skill is configured with auto-invoke behavior (e.g., \"Auto-invokes when a coding task is done\"), the model still ignores it and manually performs the steps itself instead of invoking the skill. The user then has to explicitly remind the model to use the skill.

Reproduction

1. Create a skill with auto-invoke instructions (e.g., a /ship skill that says "Auto-invokes when a coding task is done" and handles verify + commit + push)
2. Ask Claude to make a code change
3. Observe that Claude manually runs the verification steps (svelte-check, vite build) itself instead of invoking the skill
4. Claude also skips the commit and push steps entirely
5. User has to manually tell Claude to use the skill

Expected behavior

The model should recognize that a skill matches the current context (coding task completed) and invoke it automatically, as the skill description instructs.

Actual behavior

The model ignores the auto-invoke instruction and performs a subset of the steps manually. When called out, the model's instinct is to create a "feedback memory" file to remind itself to follow the original instruction — which defeats the purpose of the skill system entirely.

Why this matters

The skill system exists so users can define workflows once and have them run automatically. If the model needs a secondary persistence mechanism (memory files) to follow the primary one (skill instructions), the skill system isn't functioning as designed. Users shouldn't have to repeat themselves through multiple layers of configuration.

🤖 Generated with Claude Code