Authoring skills and subagents

Our position. We build skills from where the agent struggles and codify actual patterns, not anticipated ones (opens in a new tab). Sub-agents exist to fork context, not to specialize (opens in a new tab): one does the searching and reading in its own window, returns a succinct summary, and keeps the parent lean.

Selection comes first. Only a skill’s name and description preload at startup, SKILL.md loads when the skill becomes relevant, and bundled files load or run only when needed (Anthropic - Skill authoring best practices (opens in a new tab)). The listing budget truncates long descriptions and drops the least-used ones first, so the key use case has to lead (Anthropic - Extend Claude with skills (opens in a new tab)).

Anthropic puts evaluation before instructions: run the task without the skill, write three test scenarios, and add only enough guidance to pass them (Anthropic - Skill authoring best practices (opens in a new tab)). The gain is real but uneven: curated skills lifted the average pass rate from 33.9% to 50.5% in SkillsBench, while self-generated skills fell below the no-skills baseline (arXiv - SkillsBench (opens in a new tab)).

Every model-invoked skill adds context load on each request, and skills drift as models improve. Claude Code’s creator keeps about seven and puts the sustainable number at 5 to 10 actively maintained skills (How Claude Code’s Creator ACTUALLY Automates his work (opens in a new tab)). Deletion tests find the lines that change nothing (Building Great Agent Skills: The Missing Manual (opens in a new tab)).

A subagent attacks the same budget from the other side (Elite Context Engineering with Claude Code (opens in a new tab)). The price is that the work behind its answer is discarded: the main agent sees the summary, not the search that produced it, so check a subagent’s result like any other claim before building on it (Anthropic - Create custom subagents (opens in a new tab)).

Start here

• Authoring skills: How to scope, evaluate, trigger, maintain, and secure skills without wasting the shared context window.

Related

• Authoring skills: triggers, structure, pruning and the debate over skill evals.
• Ecosystem and extension model: how to choose between a skill, a sub-agent and MCP.
• Context Engineering: the reduce-and-delegate frame both extensions serve.
• Context as a versioned artifact: how a team manages many skills and rules.
• Multi-Agent and Orchestration: what changes when sub-agents run in parallel.

Resources

Documentation

• Create custom subagents (opens in a new tab) - Anthropic’s Claude Code documentation on defining subagents, their tools and context, and what the main agent receives back from one.
• Extend Claude with skills (opens in a new tab) - Anthropic’s Claude Code documentation on skills, including the listing budget that truncates long descriptions and drops the least-used first.
• Skill authoring best practices (opens in a new tab) - Anthropic’s guide to writing agent skills: progressive loading, evaluation before instructions, and concise SKILL.md files.

Papers and reports

• SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (opens in a new tab) - arXiv benchmark finding that curated skills raise pass rates from 33.9 to 50.5 percent while self-generated skills fall below baseline.

Posts

• Codify actual patterns, not anticipated ones (opens in a new tab) - LinkedIn post on building skills from where the agent actually struggles rather than from anticipated needs.
• Sub-agents exist to fork context, not to specialize (opens in a new tab) - LinkedIn post arguing a sub-agent exists for its own context window, returning a summary that keeps the parent lean.

Talks and videos

• Building Great Agent Skills: The Missing Manual (opens in a new tab) - AI Engineer talk on writing agent skills, using deletion tests to find the lines that change nothing.
• Elite Context Engineering with Claude Code (opens in a new tab) - Video on reducing and delegating context in Claude Code, including subagents that keep the main window lean.
• How Claude Code’s Creator ACTUALLY Automates his work (opens in a new tab) - Interview video with Claude Code’s creator on skills, loops and skill debt, recommending five to ten actively maintained skills.