Podcast insights
Agent Skills & Progressive Disclosure: How Builders Structure Agent Knowledge
Every agent that ships eventually outgrows its prompt. Agent skills — folders of instructions the model pulls into context only when it needs them — are how Anthropic's own engineers are solving this. We mined real podcast and conference transcripts for the specifics: what goes in a SKILL.md, how front-matter loading keeps context clean, and what happened to eval scores when one team swapped a 400-line prompt for skills. Every quote is verbatim.
From 4 episodes on PodLearn
A skill is packaged, composable knowledge the agent loads on demand — not another tool call.
skills are packaged and composable information that Claude has the ability to pull into context whenever Claude realized that it needs that information to complete a particular task
Tool, skill, or subagent? Decomposing an agent that outgrew its prompt · Claude
The dividing line: system prompt for always-needed context, skills for sometimes-needed knowledge.
Leave the system prompt only for the information that Claude needs in its mind, regardless of the task that you give it. Skills are fantastic for packaging information that Claude is going to need some of the time, not all of the time.
Tool, skill, or subagent? Decomposing an agent that outgrew its prompt · Claude
A real refactor: a 400-line system prompt cut to 15 lines by moving business logic into skills — and evals climbed from 83% to 92%.
We actually simplified our system prompt to 15 lines long, and we replaced all of our business logic with skills. … at the end of this, my eval score is about 92%.
Tool, skill, or subagent? Decomposing an agent that outgrew its prompt · Claude
SKILL.md anatomy: only the front matter loads upfront; the body and reference code load when the skill is actually invoked.
just the front matter of the skill is loaded in instead of sort of all of your tool descriptions, you know, which can consume quite a lot of the context window up front. … the entire rest of the body of the skill is loaded in if it's instantiated, followed by say some references
Anthropic Workshop: Build Agents That Run for Hours · AI Engineer
The runtime behavior is agentic, not static: the agent decides it needs a capability, then goes and reads the SKILL.md itself.
I think generally skills are a form of progressive context disclosure. … it's a way of like the agent being like, okay, I need to do this, let me find out how to do this and then let me read in the skill.md
Claude Agent SDK — Full Workshop (Thariq Shihipar, Anthropic) · AI Engineer
Skills are how agents pick up capabilities other people — or other agents — codified, making them shareable procedural memory.
we launched skills, which let agents pick up brand new capabilities that either other agents have designed and shared with them or humans and users that they interact with have designed for them
Episodes this draws from
- Tool, skill, or subagent? Decomposing an agent that outgrew its prompt
Claude
- Claude Agent SDK — Full Workshop (Thariq Shihipar, Anthropic)
AI Engineer
- Anthropic Workshop: Build Agents That Run for Hours — Ash Prabaker & Andrew Wilson
AI Engineer
- Memory and dreaming for self-learning agents
Claude
Full transcripts are free to read with a PodLearn account.
FAQ
What is a SKILL.md file?
A SKILL.md is a markdown file that packages procedural knowledge for an AI agent: a short front-matter block (name and description) plus a detailed body of instructions, often with reference files or runnable scripts alongside it. Anthropic's own example is a front-end design skill — essentially a very detailed, expert-written prompt the agent reads only when doing front-end work.
How does progressive disclosure work in agent skills?
Only the skill's front matter — its name and a one-line description — sits in the agent's context by default. When the agent decides a task needs that capability, it reads the full SKILL.md body, and only then any linked references or code. Knowledge loads in stages, exactly when it's needed.
When should I write a skill instead of adding to the system prompt or building an MCP tool?
The system prompt is only for information the agent needs regardless of task; skills are for knowledge needed some of the time (policies, procedures, formats); tools are for atomic actions the model can't do itself. In the Anthropic workshop's refactor this cut a 400-line prompt to 15 lines while lifting evals from 83% to 92%.
Read every word — free
Every quote above links to its full, searchable transcript on PodLearn. Create a free account and read, search, and turn any episode into a lesson.