Which Claude Code skills does your team actually use?
Compare installed Claude Code skills with observed invocations over a representative reporting window. Before removing a quiet skill, check collection coverage, who owns it and whether it is deliberately manual or reserved for rare work. A zero count is a question to investigate, not an automatic deletion rule.
What does Kibble inventory?
Kibble (usekibble.com) is a Claude Code and Codex usage dashboard for teams. For Claude Code capabilities, its public collector looks at personal skills and commands, the .claude directories of checkouts found through local transcript working directories, and the skills and commands of installed plugins.
Plugin names use the plugin namespace. Symlink aliases to the same real directory do not create a second inventory row. Paths are resolved locally and discarded; the origin, version and commit of a skill stay on the machine. This is an inventory of the roots the collector supports, not a guarantee that every built-in or remotely supplied capability is included.
What does an invocation tell you?
An observed Claude Code skill call is evidence that the skill was invoked in the collected window. Kibble also counts supported MCP calls and separates available typed and model-triggered skill signals. An invocation does not prove the instructions were useful, the task succeeded or the attributed cost was caused solely by that skill.
Missing collection, an older client, work outside the window or unsupported records can all produce incomplete evidence. Review collection health before making a cleanup decision.
Is every installed skill a recurring cost?
No universal dollar amount follows from installation alone. Claude Code documents when descriptions and full skill bodies enter context, including manual-only exceptions. Its settings can intentionally prevent automatic invocation.
Kibble's description-token figure is an estimate derived from character length, not a measurement of the exact prompt sent on every request. Caching, truncation, configuration and agent behavior affect the actual tokens and charge. Use the figure to find descriptions worth reviewing; do not multiply it by all sessions and call the result verified savings.
What would a cleanup review look like?
This is a made-up example of three Claude Code skills over four weeks. The names and counts are illustrative, not customer data.
| Skill | Observed calls | Owner's explanation | Next action |
|---|---|---|---|
| release-check | 0 | Only needed during an emergency release | Keep it available; check its manual invocation path |
| old-api-notes | 0 | The API was retired; collection was healthy | Archive it with the owner's agreement |
| review-pr | 18, all typed | Review is deliberately started by a person | Keep manual behavior; inspect usefulness separately |
Four weeks is this example's observation window, not a universal expiry date. A quarterly workflow needs a longer review period. With Kibble's free 30-day retention, keep that limit in mind when interpreting a quiet period.
How do you handle duplicates?
Two names with similar jobs are candidates for a conversation. Ask their owners whether the instructions, permissions and intended workflows are actually equivalent. If they are, choose one maintained definition, update references and observe adoption afterward. Counts cannot establish semantic equivalence without a local human review.
Does the same method work for Codex?
The inventory step does, within the Codex directories supported by the collector. Kibble also counts explicit Codex skill selections and recognized command submissions in local CLI history. Automatic skill reads and skill cost attribution remain unobserved. Never use a zero to label a Codex skill unused. Supported structured MCP calls still provide server-name counts.
The Codex team guide lists the available signals. Compare only columns that both agents actually report.
Who decides what is reported?
Capability reporting is an organization setting. When it is disabled, the CLI follows that policy and ingest drops the capability section. Skill and MCP names can reveal internal terminology, so explain this inventory to the team before enabling collection. The collector sends names and counts, never skill contents, prompts or file paths. Read the full privacy boundary.
After a change, compare another representative window and ask whether the workflow improved. Fewer installed skills is not the objective by itself; a useful, maintained set is.
Kibble shows Claude Code and Codex usage across your team. Counts, never content.
Start free