Token Usage for Discovery and Full Workflow Loading with agentskills.io

The agentskills.io standard implements a progressive-disclosure model that costs approximately 30 tokens per skill during discovery (metadata only) and 500–2,000 tokens per skill when loading full workflows (complete instructions and scripts).

The mukul975/Anthropic-Cybersecurity-Skills repository demonstrates this token-efficient architecture through its collection of 754 cybersecurity skills. By separating lightweight discovery metadata from heavy execution content, the platform enables AI agents to scan entire skill libraries without exhausting LLM context windows.

The Progressive-Disclosure Token Model

The agentskills.io specification structures skill access in two distinct phases to optimize token consumption. This architecture is documented in README.md at lines 142–149, which defines the cost boundaries for each access pattern.

Discovery loads only YAML front-matter containing tags, domain classifications, and framework mappings. Full workflow loading retrieves the complete skill body including Markdown instructions, verification steps, and helper scripts.

Discovery Phase: Scanning Skill Metadata (~30 Tokens)

During discovery, agents read only the structured metadata for each skill without processing the actual implementation content. According to the source documentation in README.md (line 142), each skill's front-matter consumes approximately 30 tokens.

With 754 skills in the repository, an agent can scan the complete skill index in a single pass:

30 tokens × 754 skills ≈ 22,000 tokens

This lightweight payload fits comfortably within standard LLM context windows. The aggregated metadata lives in index.json at the repository root, which serves as the discovery payload source.

Full Workflow Loading: Deep Execution Content (500–2,000 Tokens)

Once relevant skills are identified through discovery, agents load the complete workflow definitions. As implemented in the skill directories (e.g., skills/performing-memory-forensics-with-volatility3/), full workflows include:

• Detailed Markdown instructions
• Step-by-step verification procedures
• Referenced helper scripts and code samples

This complete content ranges from 500 to 2,000 tokens per skill depending on complexity. By loading full workflows only for selected skills, agents preserve context capacity for actual execution and analysis.

Practical CLI Implementation

The built-in CLI tools respect this token budgeting strategy through separate commands for metadata versus full content.

List all skills (discovery only, ~30 tokens each):

# Install the skill collection to local index

npx skills add mukul975/Anthropic-Cybersecurity-Skills

# Query only front-matter for all 754 skills

skills list

Load specific skill workflows (500–2,000 tokens each):

# Retrieve complete definition including scripts

skills show performing-memory-forensics-with-volatility3

The tools/README.md file documents these CLI utilities and their token-conscious design.

Programmatic Metadata Access

For custom agent implementations, access the discovery index directly without parsing individual skill files:

import json

# Load index.json containing only front-matter for all skills

with open("index.json") as f:
    index = json.load(f)

# Filter by tags without paying full workflow token costs

ir_skills = [s for s in index if "incident-response" in s["tags"]]
print([s["name"] for s in ir_skills])

This approach enables efficient skill matching against the 22,000-token discovery dataset before selectively loading high-cost workflow content.

Key Files Supporting Token Efficiency

File	Purpose
README.md	Documents token model at lines 142–149
index.json	Aggregated front-matter for all 754 skills (~22k tokens total)
skills/<name>/	Individual directories containing 500–2,000 token full workflows
tools/README.md	CLI documentation for token-aware commands

Summary

• Discovery costs ~30 tokens per skill via YAML front-matter in index.json, enabling full library scans (~22k tokens for 754 skills).
• Full workflows cost 500–2,000 tokens per skill and include complete Markdown instructions and scripts stored in individual skill directories.
• Progressive disclosure allows agents to identify relevant skills cheaply before paying the higher token cost for execution details.
• CLI tools (skills list vs skills show) and the index.json API provide direct access to the token-budgeted architecture.

Frequently Asked Questions

How many total tokens are needed to scan all 754 skills?

Scanning the complete skill library requires approximately 22,000 tokens (30 tokens × 754 skills). This discovery-phase payload uses only the metadata from index.json, allowing agents to identify relevant capabilities without loading full workflow content.

What content differentiates discovery tokens from full workflow tokens?

Discovery loads only YAML front-matter including skill names, tags, domain classifications, and framework mappings. Full workflows load the complete skill body—Markdown instructions, verification steps, and helper scripts—which typically requires 15–65 times more tokens than discovery.

Where does the discovery metadata reside?

The aggregated discovery metadata lives in index.json at the repository root. This file contains the compiled front-matter for all 754 skills and serves as the primary data source for the skills list command and custom agent indexing.

Why is there a range (500–2,000) for full workflow tokens?

Token consumption varies by skill complexity. Simple skills with brief Markdown instructions cost closer to 500 tokens, while complex workflows containing multiple verification steps and substantial helper scripts (as seen in forensic analysis skills) approach the 2,000-token upper bound.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}