Skill Manager estimates discovery context from each skill's name and summary. It counts words, Unicode characters, and CJK characters to create a comparable token estimate across the inventory. Treat the result as a planning signal; runtime context also depends on agent loading behavior, model tokenization, full instructions, and referenced resources.
What the estimate means
Agent interfaces need enough metadata to decide which skill fits a request. A short, precise name and description make that discovery layer easier to scan. Skill Manager measures this layer so you can compare the aggregate footprint and find unusually verbose metadata.
The value does not claim an exact billable or runtime token count. Tokenizers vary by model, and agents can load skill content at different moments. A skill can also reference scripts, examples, and assets that remain outside the initial discovery text.
How Skill Manager calculates it
The current estimator joins the parsed skill name and summary, then derives three counts: Unicode scalar length, words separated by whitespace or punctuation, and CJK characters. Its Latin estimate uses the larger of the word count and one token per five characters, then adds one token for every two CJK characters.
$ npx github:Ryan-yang125/skill-manager audit --jsonUse JSON output when you want to sort, track, or compare estimates over time. The report preserves the estimate as an integer for each skill and aggregates it across the inventory.
Read size alongside value
| Observation | Useful next step |
|---|---|
| Large summary with clear usage evidence | Edit for clarity when you own the source; keep the capability available |
| Large summary and overlapping skill | Compare both sources and choose one canonical description |
| Small summary with vague wording | Clarify triggers and outcomes; brevity alone carries limited value |
| Zero evidence and partial log coverage | Expand coverage or observe longer before an archive decision |
Improve the discovery text
- Name the user request that should trigger the skill.
- State the concrete outcome the skill produces.
- Keep setup detail inside the full instructions.
- Remove repeated synonyms that add little routing value.
- Re-run the audit and compare the inventory estimate.
Scope rule: Skill Manager's number models name-and-summary discovery text. Use a model-specific tokenizer and observed agent behavior when you need a full runtime measurement.
Catalog context FAQ
What does the catalog context estimate measure?
It estimates tokens for the skill name and summary used in discovery, giving the inventory a consistent comparison metric.
Is it an exact runtime token count?
It is a planning estimate. Runtime cost depends on agent loading behavior, model tokenizer, full instruction content, and referenced resources.
Should I archive every large skill?
Size is one review signal. Keep a large skill when its purpose, provenance, and workflow value are clear.
Measure your discovery layer
Run the JSON audit and compare context estimates across every local skill root.
Run Skill Manager