You now have all four parts. This chapter combines them into a workflow you can adapt: "Write this week's release notes and prepare them for Slack." Each part does one job, and you can remove any part and still understand what is left.
The design
| Part | File | Job |
|---|---|---|
| Skill | .claude/skills/release-notes/SKILL.md | House format and rules |
| Subagent | .claude/agents/changelog-scout.md | Read merged PRs on a small model, return a compact list |
| MCP server | Markdown Convert MCP | Convert the Markdown draft to Slack mrkdwn |
| Hook | .claude/settings.json (PreToolUse) | Block any Slack send tool unless .release-approved exists |
Step 1: the skill
---
name: release-notes
description: Write weekly release notes from merged PRs in our house format. Use when the user asks for release notes, a changelog or "what shipped". Skip for single-PR descriptions.
---
# Release notes
1. Delegate PR collection to the `changelog-scout` agent with the date range.
2. Group into Features, Fixes, Internal. Drop dependency bumps.
3. One line per item: user-facing change, then (#PR).
4. Open with one sentence naming the most important change.
5. Produce Markdown. If the user wants Slack, convert with `markdown_to_slack` before handing it to the Slack tool. Do not send; show the draft and wait.Step 2: the subagent
---
name: changelog-scout
description: Collect merged PRs for a date range and return title, number and one-line user-facing summary for each. Read-only.
tools: Bash, Read
model: haiku
---
Run `gh pr list --state merged --search "merged:>=<START>" --json number,title,body --limit 100`.
Return one line per PR: `#<number> | <title> | <one-line summary>`. No commentary. Max 60 lines.The scout reads up to 100 PR bodies; the main chat receives at most 60 short lines.
Step 3: the MCP server
Connect Markdown Convert MCP (chapter 7). The skill tells the agent when to use markdown_to_slack: only for an API send, never before a paste into Slack's composer.
Step 4: the hook
#!/usr/bin/env bash
# .claude/hooks/require-approval.sh (PreToolUse, matcher: slack tools)
if [ ! -f .release-approved ]; then
echo "Sending is blocked until a human creates .release-approved. Show the draft instead." >&2
exit 2
fiRegister it with a matcher for your Slack tool's name (MCP tools appear as mcp__<server>__<tool> in Claude Code).
Step 5: run it
- New session. Ask: "Release notes for everything merged since Monday, for Slack."
- Watch for: the skill's structure, a
changelog-scouttask, amarkdown_to_slackcall, and the hook blocking the send. - Review the draft. When happy,
touch .release-approvedand ask the agent to send. Delete the file afterwards.
Review checklist
Use this on any workflow you build:
- One job per part. No skill that tries to enforce, no hook that tries to teach.
- Descriptions match real requests. You tested the trigger phrases.
- Output contracts. Every subagent returns a bounded, structured result.
- Least privilege. Subagent
toolsand MCP servers limited to what the job needs. - A human gate before anything external. Sends, deploys and deletes need an explicit approval, enforced by a hook.
- Receipts. You can tell which model ran and which tools were called.
Where to go next
- Browse ai-agent-skills for 17 ready-made thinking skills (pre-mortem, five whys, tradeoffs, six hats and more).
- Validate your own skills with the AI Agent Skill Validator.
- Continue with the Prompt Engineering for Coding Agents course: better prompts make every one of these parts work better.
Quick check
-
Which part keeps 100 PR bodies out of the main conversation?The
changelog-scoutsubagent. -
What guarantees the notes are not posted before review?The PreToolUse hook that blocks the send tool until
.release-approvedexists. -
You remove the skill. What breaks?Format consistency and the "when to convert" judgment. The scout, converter and hook still work, but the agent improvises the notes.






