Part 4: Skills and Agents | Garage Education

Chapter 14: Skills as files

Sat, 10 Oct 2026 00:00:00 +0000

A skill is a folder with a SKILL.md file. The file has a name and a one-line description in front matter, then the full instructions. The harness loads only names and descriptions at the start. A useSkill tool loads the full text when the model asks for it.

---
name: triage-error-group
description: Judge the severity of one error group from its rate, the services it affects, and whether a runbook exists.
---

# Steps
1. Read the error rate for the last 15 minutes.
2. List the affected services.
3. Search runbooks for the error fingerprint.
4. Return one severity: low, medium, or critical.

Figure 14.1. A skill costs one line of the prompt until the model asks for it. The figure leaves out the tool calls the model makes while following the steps.

Chapter 15: Skill selection

Sat, 10 Oct 2026 00:00:00 +0000

Skills are indexed by their description embedding in PostgreSQL. For each task, the harness takes the five nearest skills. Only those five appear in the prompt, and useSkill accepts only those names. Measure recall at five on 20 labeled tasks. Include pre-filtering only when the catalog grows past about ten skills.

Figure 15.1. Skill selection for one task. The same five names are the prompt list and the useSkill allowlist. Recall at five, on 20 labeled tasks, says whether the right skill made the cut.

Chapter 16: Sub-agents

Sat, 10 Oct 2026 00:00:00 +0000

A sub-agent has its own system prompt, its own tool allowlist, and its own step limit. The parent calls it through a tool that returns a short summary. A focused agent does one job well.

Figure 16.1. To the parent, a sub-agent is one more tool. Its prompt, tools, and step limit are its own. Chapter 18 fixes the shape of the summary.

Chapter 17: Agent selection

Sat, 10 Oct 2026 00:00:00 +0000

Three patterns exist. A direct agent handles the task itself. A router labels the task and sends it to one agent. A supervisor plans several subtasks and combines the results. Start with direct handling. Add a router only when the cost of a wrong route is clear. Test each pattern with the eval set.

Figure 17.1. The three patterns. A router sends each task to exactly one agent; Agent B is greyed out because it was not chosen this time. A supervisor sends to several and combines the results. Each pattern is tested on the same eval set.

Chapter 18: Context isolation

Sat, 10 Oct 2026 00:00:00 +0000

A sub-agent keeps its full detail in the log. It returns a capped summary with a fixed shape: a finding, a confidence, and a detail id. Set the cap in code, not in the prompt. The parent’s context stays small.

Figure 18.1. Two outputs from one sub-agent. The detail id links the short summary back to the full record. The chapter does not say whether the parent can fetch the detail itself, so that edge is dashed.