Turn play into a skill

Project 2: HADR agent

You have already used tools and MCP while playing two episodes. Now turn the procedure that worked into a reusable skill.

Mechanism What it supplies Use it for
Tool A typed action the harness can execute Reading a file, querying a database, or controlling a browser
MCP A standard connection to tools and context Plugging an external system into an agent harness
SKILL Text and resources for a matching task Describes a workflow, checklist, or domain-specific tool

The MCP connection gives Claude access to the simulator. The skill tells a new session how to use that access well. It does not add new tools.

Activity

Use the transcript, scratch incident list, and notes from MCP Play. Work in the same mcp-play/ folder.

Write the skill

Ask Claude to extract the procedure from the episodes:

Read our two MCP Play transcripts and notes. Write the procedure that worked as a skill at .claude/skills/hadr-play/SKILL.md. Include the per-tick checklist, the evidence required before a dispatch, conditions for redirecting a truck, the stopping condition, and the mistakes to avoid. Do not invent rules that the episodes do not support.

A skill is one markdown file in a named folder, with YAML frontmatter:

---
name: hadr-play
description: Play a HADR fire-dispatch episode over the hadr MCP server. Use when asked to play, resume, or demo the game.
---

[The operating procedure.]

Review the result. The frontmatter description is the trigger, so it must say when to use the skill. Put the full procedure in the body. Remove advice such as “dispatch carefully” unless it changes an observable action.

Do not create a skill for a rule or policy that should always be loaded, that belongs in CLAUDE.md. A skill loads only when its description matches.

Test the skill

  1. Start a fresh Claude session in mcp-play/ (skills are discovered at startup).
  2. Ask it to “play a stage-1 episode” without naming the skill. Confirm that the skill loads.
  3. Watch one complete episode. If it stalls or misreads state, change the skill rather than adding corrective chat messages.
  4. If time permits, test one harder scenario or run the same skill from OpenCode. Record what fails when the context, model, or evidence changes.

When you’re done

  1. Open a pull request with the skill and trade reviews with a teammate. Compare the procedures and their evidence.
  2. Run /context, screenshot the output, and upload it to the Padlet for the tokens activity.

Before writing future skills, check whether one already exists in Anthropic’s skills library or with npx skills find <keyword>.