benchrAll capabilities Agents, tools and MCP
Teach the model a procedure once and reuse it
A folder with a SKILL.md, optional reference files and scripts. The model reads the name and description always, the instructions only when your request matches.
DocsOfficially documentedTestNot tested by benchr
Claude CodeClaude APIclaude.ai and the Claude appsClaude Opus 5+2
Overview
- Loads in three stages: metadata always (about 100 tokens per skill), the SKILL.md body when triggered (under 5k tokens), bundled files and scripts only when read.
- Runs bundled scripts through bash so the script's code never enters the context window - only its output does.
- Works in Claude Code from ~/.claude/skills or .claude/skills, in the API by skill_id inside the code execution container, and on claude.ai as an uploaded zip.
How it works
- 01Write the description to say both what the skill does and when to use it - that string is what the request is matched against.
- 02Keep SKILL.md under about 5k tokens and push detail into separate reference files the model reads only if it needs them.
- 03Turn anything deterministic into a script rather than instructions. Scripts are cheaper and do not drift.
- 04Install it per surface. Skills do not sync between Claude Code, the API and claude.ai.
Limits
- Names cannot contain the reserved words "anthropic" or "claude", must be lowercase with hyphens, and are capped at 64 characters; descriptions are capped at 1,024.
- On the Claude API a skill runs with no network access and no runtime package installation.
- Custom skills do not sync across surfaces, and claude.ai has no admin-level org-wide distribution.
- A skill is executable trust: the documentation warns that a malicious one can direct the model to misuse tools or exfiltrate data.
Verification
benchr test runs
benchr has not run this capability. No test has been published.
benchr test runs →
Evidence
Sources
- Anthropic - Agent Skills: progressive disclosure, field limits, per-surface constraints Tier 1 · September 1, 2026
Not stated by the source
- How many skills can be installed before metadata cost becomes material
- Whether triggering accuracy varies by model
Source-check history
- September 1, 2026DocsOfficially documentedFirst logged. The three loading levels, the name and description limits, and the no-network API constraint read from the overview.
Changes
Provider changes affecting this
- August 31, 2026Claude Sonnet 5Pricing changed
- July 3, 2026Claude Sonnet 5Pricing changed
- July 1, 2026Claude Sonnet 5Model released
- June 1, 2026Claude Opus 4.8Model released