OpsfleetAgent Bootcamp

Task 8 · 8 tasks

Agent skills

Package company know-how as skills that the agent loads only when a question needs them.

25 minIntermediate
Alice’s ask

The “House Rules” problem

HR has a fixed format for the headcount report, and Finance has strict rules for anything about pay. Pasting every team's rules into the system prompt makes it long, expensive and confusing. Ship each rule book as a skill instead: the agent sees a one-line description of each and opens the full instructions only when a question calls for them.

You

Build it

  1. Write a skill

    Challenge

    Turn HR's headcount rules into phase1/starter/skills/headcount-report/SKILL.md:

    • Count active employees only.
    • One row per department, including departments with nobody active (count 0).
    • Columns: Department, Active employees, Share of company (%, one decimal, computed in SQL).
    • Largest department first, then a Total row.
    • Finish with the line Source: datastream_corp.db, active employees only.

    skills/compensation-policy/SKILL.md next to it is a finished example; the agent gets both.

    Docs
    Hint 1

    A skill is a folder with a SKILL.md: YAML frontmatter with a name (the folder's name, lowercase with hyphens) and a description, then Markdown instructions. See the Agent Skills specification.

    Hint 2

    The description is all the model sees until it loads the skill, so it decides whether the skill is ever used. Say what the skill is and when to use it, e.g. “Use when asked for headcount, employee counts per department or a staffing overview.”

    Solution
    phase1/starter/skills/headcount-report/SKILL.md
    ---
name: headcount-report
description: The standard DataStream Corp headcount report. Use when asked for headcount, employee counts per department or a staffing overview.
---
# Headcount report

1. Count only employees with `status = 'active'`.
2. One row per department (`employees.department_id = departments.id`); keep departments with no active employees
   (LEFT JOIN, count 0).
3. Columns, in this order: Department | Active employees | Share of company (%, one decimal). Compute the share
   in the SQL query (`ROUND(100.0 * count / total, 1)`), never by hand.
4. Sort by active employees, largest first, then add a final **Total** row.
5. Answer with the markdown table, then exactly one line: `Source: datastream_corp.db, active employees only.`

  2. Give the agent its skills

    Challenge

    Build the agent with the database tools and every skill in SKILLS_DIR.

    Work in phase1/starter/t8_skills.py (look for TODO; an unfinished one prints [starter] TODO …) and run it with uv run bootcamp.py phase1 t8 --starter. The reference solution is phase1/t8_skills.py; uv run bootcamp.py phase1 t8 runs it.

    Docs
    Hint 1

    Skills come from the AgentSkills plugin in strands.vended_plugins.skills. A plugin bundles hooks and tools; this one adds the skill list to the system prompt before each call and gives the agent a skills tool that returns a skill’s full instructions. See Strands: agent skills.

    Hint 2

    Pass the folder that contains the skill folders, and hand the plugin to the agent with plugins=[...] (not tools=).

    Solution

    Paste each block below into phase1/starter/t8_skills.py in place of the definition with the same name (the TODO), including any import line that comes with it. Nothing else in the file changes.

    phase1/starter/t8_skills.py · replace build_agent() (keep its import line)
    from strands.vended_plugins.skills import AgentSkills


def build_agent(mcp_client: MCPClient) -> Agent:
    """The data engineer with the database tools plus every skill under `SKILLS_DIR`."""
    return Agent(
        name="Data Engineer",
        model=make_model(),
        system_prompt=SYSTEM_PROMPT,
        tools=mcp_client.list_tools_sync(),
        plugins=[AgentSkills(skills=SKILLS_DIR)],
    )

  3. Run it

    terminal · your starter file
    uv run bootcamp.py phase1 t8 --starter "Give me the headcount report"
uv run bootcamp.py phase1 t8 --starter "What is Alice Chen's salary?"
uv run bootcamp.py phase1 t8 --starter "Which projects are in planning?"
    terminal · reference solution
    uv run bootcamp.py phase1 t8 "Give me the headcount report"
uv run bootcamp.py phase1 t8 "What is Alice Chen's salary?"
uv run bootcamp.py phase1 t8 "Which projects are in planning?"

  4. Experiments

    • Watch for [tool] skills: headcount-report: the agent loads a skill before it queries. The projects question loads none, so it pays nothing for the rules it does not need.
    • Delete the “keep departments with no active employees” rule from your skill and ask for the report again. Is Research still in the table?
    • Make the headcount description vague (“Reporting helper.”) and ask again. Does the agent still find the skill?
    • Write a third skill, for example a project-status report, and ask for it. The agent code does not change.

Check your work

Phase 1 has no automated test: you check it by running the task and looking for the result below.

Headcount: [tool] skills: headcount-report, then a [tool] query_db that counts status = 'active' (long SQL is cut at 160 characters). The table has seven rows, largest first: Engineering 481 (40.0), Sales 360 (30.0), Marketing 180 (15.0), HR 84 (7.0), Finance 60 (5.0), Operations 36 (3.0), Research 0 (0.0), then Total 1,201, and ends with Source: datastream_corp.db, active employees only. No Research row, or 483 / 362 / 1,207, means a rule was skipped.

Salary: [tool] skills: compensation-policy; no individual salary, an offer of a rounded department average instead, and the line Compensation figures are confidential; do not forward.

Projects: no skills line, a [tool] query_db with status = 'planning', and exactly two projects: Brand Refresh Initiative and Office Expansion Project.

Failed? Run the same prompt on the reference (drop --starter). If the reference fails too, it's the model: rerun or try another MODEL_ID (uv run bootcamp.py llm models).

Under the hood

This is progressive disclosure: only each skill's name and description sit in the system prompt, and the body arrives as a tool result when the model asks for it. Prompts stay short however many skills you add, and each team can own its skill as a file in the repo. Claude Code reads the same SKILL.md format; this repository's .claude/skills/ folder holds skills for working on the bootcamp itself.