Lesson 5.7 — Capstone: A Swarm for a Real Job

Introduction to AI · Module 5: AI Agents · About 90 min · Windows PC, PowerShell 5, Claude Code (Max)

Goal: use everything from Module 5 on one real recurring job of yours: instructions, tools, permissions, a plan, subagents and evals. It runs from one command.

Before you start

• Ideas (pick one): weekly receipts → spend summary · weekly todo + calendar review · tidy a copy of Downloads · draft replies to a batch of non-sensitive emails saved as .txt
• My job:

Step 1 · Define "done" (⏱ 0:00–0:10)

mkdir $HOME\capstone -Force | Out-Null
cd $HOME\capstone
notepad DONE.md

Write 3–5 checkable lines, e.g. "out\report.md exists", "every input file is mentioned", "total matches".

• DONE.md written

Step 2 · Inputs and a checker (⏱ 0:10–0:25)

Copy sample inputs into $HOME\capstone\input. Then ask Claude to write the checker for you:

claude

Read DONE.md. Write check.ps1 (Windows PowerShell 5) that prints PASS only if
every line in DONE.md is true for the files in out\, otherwise FAIL plus reasons.
Then run it once; it should FAIL because out\ is empty.

• check.ps1 prints FAIL

Step 3 · Design the team (⏱ 0:25–0:40)

Still in Claude:

Design 2-3 subagents for this job (see DONE.md, input\). Write them to
.claude\agents\ with name, description, tools (least needed) and instructions.
Also write CLAUDE.md for the orchestrator: never modify input\, outputs go
to out\, finish by running .\check.ps1 and fix until PASS (max 3 rounds).
Show me the plan before writing files.

Read the plan and approve it. Then /agents to confirm (press Esc to close it).

Now leave Claude, because the next step is PowerShell: **type /exit and press Enter. Wait until you see PS C:\...> before the next command.**

• Team and CLAUDE.md created

Step 4 · Safety rails (⏱ 0:40–0:45)

mkdir .claude -Force | Out-Null
$json = '{"permissions":{"deny":["Bash(rm:*)","PowerShell(Remove-Item:*)","PowerShell(rm:*)"]}}'
[IO.File]::WriteAllText("$HOME\capstone\.claude\settings.json", $json)

• Delete blocked

Step 5 · One-command run (⏱ 0:45–1:00)

@'
Remove-Item out -Recurse -ErrorAction SilentlyContinue
claude -p "Do the job described in CLAUDE.md using your subagents. Report PASS or FAIL." --allowedTools "Read,Write,Edit,Glob,Grep,Task,PowerShell,Bash,WebFetch"
.\check.ps1
'@ | Out-File run.ps1 -Encoding utf8
.\run.ps1

You should see a summary from Claude, then PASS.

• One command → PASS

Step 6 · Mini eval: 5 cases (⏱ 1:00–1:25)

Make 5 variations of input (e.g. one extra file, one odd format, one empty file) in cases\c1–c5, then:

$pass = 0
foreach ($n in 1..5) {
  Remove-Item input -Recurse -ErrorAction SilentlyContinue
  Copy-Item "cases\c$n" input -Recurse
  $r = (.\run.ps1 | Select-Object -Last 1); Write-Host "case $n : $r"
  if ($r -eq "PASS") { $pass++ }
}
"PASS RATE: $pass / 5"

• Pass rate: / 5

Step 7 · Show and tell (⏱ 1:25–1:30)

Send Skills Coach: the job, the team (agent names), the pass rate, and one thing you'd improve.

✅ Pass check

• 4/5 or more eval cases pass
• The whole job runs from one command (.\run.ps1)
• Told Skills Coach so 5.7 (and Module 5!) gets marked done 🎉

Troubleshooting