Agent Skills laboratory

How AI agent skills load procedures without becoming tools

A skill is a portable package of procedural context. Its metadata helps an agent decide when the procedure is relevant; SKILL.md supplies the workflow; optional references and scripts load only when the task needs them.

Choose a task, inspect the catalog match, and advance from metadata to SKILL.md to one on-demand reference. Then inject ambiguity, a trust denial, or a missing resource and watch activation stop at the exact boundary.

A compact Agent Skill folder unfolding into metadata, procedural instructions, examples, and selectively loaded reference material
Agent Skills · Progressive disclosure of procedural context
Inspect
Metadata · SKILL.md · references · activation
Changes
Instructions loaded into working context
Does not imply
API access · executable authority · durable state
Skill activation microscoperesource-loaded
  1. 01catalogcompleted

    Only discovery metadata is visible; no skill procedure has entered context.

  2. 02activatedcompleted

    The chosen SKILL.md is now in context; referenced files remain unloaded.

  3. 03resourcecompleted

    The procedure reached the relevant step, so exactly one referenced resource entered context.

Task

Verify upcoming agent conferences and draft a briefing with organizer citations.

A host can see the catalog without paying the context cost of every procedure. Activation and reference loading are separate, observable decisions.

Catalog metadata · always visible
catalog eligibleConference research

Verify dates and locations against official organizer pages, then create a cited event brief.

Trigger: conference dates, event research, or a sourced conference briefing
catalog eligibleSupport response

Draft a customer reply grounded in an approved support policy.

Trigger: customer support cases and policy-grounded replies
catalog eligibleSite publishing

Validate and publish an existing website through its configured host.

Trigger: website deployment or publication
Exact context disclosure3 artifacts · 1347 serialized bytes
01
metadata-catalogskills/catalog.json
[
  {
    "id": "conference-research",
    "name": "Conference research",
    "description": "Verify dates and locations against official organizer pages, then create a cited event brief.",
    "trigger": "conference dates, event research, or a sourced conference briefing",
    "trusted": true
  },
  {
    "id": "support-response",
    "name": "Support response",
    "description": "Draft a customer reply grounded in an approved support policy.",
    "trigger": "customer support cases and policy-grounded replies",
    "trusted": true
  },
  {
    "id": "site-publishing",
    "name": "Site publishing",
    "description": "Validate and publish an existing website through its configured host.",
    "trigger": "website deployment or publication",
    "trusted": true
  }
]
02
skill-instructionsskills/conference-research/SKILL.md
---
name: conference-research
description: Verify conference facts and produce cited briefings.
---

# Conference research

1. Prefer the event organizer's page over aggregators.
2. Record the event name, start date, end date, location, and source URL.
3. Treat an unverified date as unknown; never infer it from a search snippet.
4. Before drafting, load references/source-policy.md and apply its source rules.
03
skill-referenceskills/conference-research/references/source-policy.md
# Source policy

- Accept an organizer, venue, or program page that names the event and date.
- If two official pages conflict, surface the conflict and the retrieval dates.
- Keep the source URL beside every date and location in the evidence table.

Primary research

Inspect the mechanism at its source.

Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.

Required metadata and folder formatAgent Skills specificationOriginal source ↗Loading and precedence in a real runtimeOpenClaw skillsOriginal source ↗On-demand procedural learningHermes Agent skillsOriginal source ↗

Keep experimenting

Change the layer, keep the system visible.

AI Lab indexSee every interactive laboratoryBrowse the suite →Transformer microscopeGo beneath the agent runtime into the modelInspect the matrices →