self-improving-agent generates skill names with reserved 'claude' keyword

[prefrontalsys]

Summary

The self-improving-agent skill (under engineering-team/self-improving-agent) generates new skills from session patterns. When those patterns involve Claude Code itself, the generated skill names include the reserved fragment claude — e.g., claude-code-settings, claude-code-maintenance, claude-code-plugins, claude-mcp-tools.

Per the Claude Code skill spec, claude and anthropic are reserved name fragments that should not appear in user-created skill names.

Observed behavior

Over normal usage, the si agent auto-created these skills:

• claude-code-maintenance
• claude-code-plugins
• claude-code-settings
• claude-plugin-development
• claude-workspace-audits
• claude-mcp-tools

All of these contain the reserved claude fragment in their directory/skill name.

Expected behavior

The skill creation logic should either:

1. Prefix with cc- instead of claude-code- (e.g., cc-settings, cc-maintenance)
2. Check generated names against a reserved fragment list (claude, anthropic) before writing the skill directory
3. Strip the reserved fragment and use the remaining descriptive portion

Workaround

We manually renamed all 6 skills to use the cc- prefix and updated cross-references. This works but will recur if the si agent creates new Claude-Code-related skills.

Environment

• Claude Code CLI
• self-improving-agent installed via plugin marketplace

[manja316]

We manage 40+ skills and hit this exact problem. The cc- prefix workaround is solid but here's what we ended up doing to prevent recurrence:

Name validation at creation time:

RESERVED_FRAGMENTS = {"claude", "anthropic", "openai", "mcp-official"}

def validate_skill_name(name: str) -> str:
    parts = name.lower().split("-")
    if any(frag in RESERVED_FRAGMENTS for frag in parts):
        # Strip reserved fragments, keep descriptive parts
        clean = [p for p in parts if p not in RESERVED_FRAGMENTS]
        return "-".join(clean) if clean else f"custom-{name.replace('claude-', '')}"
    return name

For the self-improving-agent specifically, the fix belongs in the skill template generation step — right before it writes the new directory name. A post-generation rename creates orphaned references in any skill that cross-references by name.

The cc- prefix convention works well in practice. We also use domain prefixes like security-, deploy-, test- which naturally avoid the reserved fragment collision.

[sophymarine]

The reserved-name bug is exactly the kind of thing that becomes painful once an agent starts generating files from its own observations. A practical guardrail is to validate names before write time, and to keep the generated skill tied to a source-backed lookup step rather than free-form labels. OpenRegistry can help in adjacent agent workflows when the skill needs live company or registry facts, with cross-border UBO walks across 27 government registries in one query. https://openregistry.sophymarine.com