Progressive Disclosure

Agent Skills是Anthropic发布的能力封装标准，定义为一个包含Skill.md文件的目录，可包含说明、脚本和其他资源

Progressive Disclosure is originally a classic principle in user interface design, referring to showing complex information only when the user needs it.

Agent Skill adopts a lazy-loading mechanism whereby the Agent first reads only the metadata (name and description) of each Skill and loads the full instructions only after matching the user's need, so having multiple Skills coexisting does not increase Token consumption

Skill's progressive loading mechanism has two levels: the first level exposes only name and description to form a minimal skill catalog; the second level reads the full SKILL.md and supporting files only after a hit.

The core design philosophy of Agent Skills is Progressive Disclosure, meaning that only the information and tools truly needed at the current moment are exposed to the agent at the appropriate time.

The Skill mechanism uses on-demand loading: it first sends the metadata of all Skills to the large model, and after the large model selects the matching Skill, the full instructions are loaded.

Skills采用渐进式披露设计，平时只向模型提供技能的名称和描述，具体执行规则在需要时才加载

Skill由三层结构组成：概要层（约100字）、主体层（约3000字）和细节层，采用渐进式披露策略

A typical form of progressive disclosure in agent scenarios is "Agent Skills": writing the operating instructions for repetitive tasks as separate documents, keeping only a brief index in the main prompt, and reading the full text on demand.

Context management counters information accumulation in the loop through means such as compaction and progressive disclosure, determining what content is fed to the agent each round.

Progressive disclosure means that the Agent gradually loads and discloses relevant skill information at runtime according to task needs, rather than exposing all capabilities to the model at once.

The core of the progressive disclosure mechanism is that at session start only the skill's YAML front matter description is loaded, and the full workflow instructions are loaded only when a task triggers it, thereby greatly reducing context cost.

Progressive disclosure puts metadata first, compressing first-layer overhead to just dozens of tokens per skill.

Too many tools cause context bloat, consuming tokens and reducing model selection accuracy, so the lightweight version adopts a progressive disclosure strategy to lower invocation error rates and latency.

Progressive Disclosure is borrowed from human-computer interaction and information architecture design; its core idea is to present only the necessary information when the user needs it, avoiding cognitive overload.

Skill uses a layered loading design: at startup only the names and descriptions of all Skills are loaded; after a task is issued, descriptions are matched and the full content is then loaded, with on-demand reading during execution.

Progressive disclosure originates from the field of interaction design, referring to gradually revealing information only when the user needs it to avoid the cognitive burden of presenting all information at once; in the AI agent context it is used to manage the context window.

Progressive disclosure, through the three-level hierarchy of "metadata → detailed guidance →附属 resources," lets the model first make routing decisions using low-cost information and then expand high-cost content on demand, thereby avoiding the context window being occupied by irrelevant content.

A progressive disclosure strategy can reduce token consumption: first show the model only the skill categories, let the decision model choose a category, then choose the specific skill, and load only the tool descriptions actually needed.

Progressive Disclosure is a classic principle in interaction design, first systematically articulated by the Nielsen Norman Group.