How much context does an agent skill cost?
Updated 13 November 2026. Measurements are from a run of the mdprev skill inspector on that date. Token counts are estimates (characters divided by four), not exact counts for any model.
A skill is not free to install. According to the Agent Skills specification, an agent reads the name and description of every installed skill at startup, loads the whole SKILL.md of a skill only when a task matches, and opens other files only when the instructions point to them. The first level is paid in every session, whether or not the skill is ever used. What is inside a SKILL.md explains the levels.
Measured on 20 public skills
We ran the inspector on github.com/anthropics/skills, which held 20 skills on 13 November 2026. Seven of them, from the smallest startup cost to the largest:
| Skill | Always | On trigger |
|---|---|---|
| webapp-testing | ~55 | ~970 |
| frontend-design | ~55 | ~2.3k |
| mcp-builder | ~72 | ~2.8k |
| ~110 | ~2k | |
| pptx | ~184 | ~5.2k |
| xlsx | ~238 | ~2.1k |
| claude-api | ~270 | ~26k |
- All 20 together put about 2.4k tokens in every session, an average of about 120 each. The specification’s figure of about 100 per skill is the right order of magnitude; individual skills here ranged from about 55 to about 270.
- The median description was 324 characters. The longest, 1068 characters, is longer than the 1024 the specification allows.
- Three of the 20 have a SKILL.md above the specification’s suggested 5000 tokens (about 26k, 8k and 5.2k), and one is above its suggested 500 lines (617).
What it adds up to
The startup cost grows with the number of skills, not with how many you use. If the average stayed at about 120 tokens, 100 skills would cost about 12k tokens before the first message. That figure is arithmetic on the measurement above, not a measurement.
Keeping it small
- Write the description for the decision, not the manual. It must say what the skill does and when to use it. Everything beyond that is paid for in every session.
- Keep SKILL.md short. The specification suggests under 500 lines and about 5000 tokens, and moving detail into files under
references/that are read only when needed. - Keep references one level deep. A file that points to another that points to another is easy for an agent to follow only partway.
- Measure before you install a collection. The skill inspector shows the three levels for one skill or a whole repository, and adds up the startup cost of the skills you select.
How far to trust these numbers
Characters divided by four is a rough rule for English prose. Code and text in other languages usually take more tokens per character, and every model counts differently. Use the numbers to compare skills with each other, not to budget to the token. How an agent actually loads skills is up to the agent; the specification describes the model, and implementations may differ.