Select the automation MVP

Outcome and preparation

Day 1 · 16:15–17:00 · 45 minutes. Choose a bounded automation and finish its design canvas. Bring your task card, approach card, Skill and baseline timing. Use 5 minutes to explain scoring, 30 to design and 10 to challenge.

Start with a workflow whose value you can measure. Automation makes a flawed method repeat faster; it does not correct missing evidence or remove accountability.

Step by step

1. Step 1

   List three recurring tasks: weekly executive reporting, system-health evidence gathering and business-plan evidence review. Add your own candidates only if their source access and owner are clear. Record runs/month and minutes/run.

2. Step 2

   Score each 1–5 for recurring effort, decision value and measurability; separately score data risk and setup difficulty, where high means harder. Do not silently add positive and negative scales. Use (effort + value + measurability) - risk - difficulty as a discussion aid, not an investment verdict.

3. Step 3

   Choose weekly reporting for the required workshop build. Complete the canvas with Claude and save handover/agent-canvas.md.

   Prompt / exampleCopy

   Help specify a small executive-reporting MVP from this task card.
   Fields: one job, users, approved input files, reporting period, source authority,
   deterministic calculations, agent responsibilities, output files, human reviewer,
   escalation conditions, schedule, success tests, cost inputs and stop conditions.
   Use local fictional files. No external sending, production writes or hidden credentials.
   Keep the evidence agent read-only. The main workflow invokes the calculator and writes drafts.
   Ask about missing decisions before finalising the canvas.
   [Paste task card and checked reporting method]
   

4. Step 4

   Draw the workflow in plain text: source inventory → evidence extraction → calculation → draft → review → human approval. Define the handoff table: claim, value/unit, file, row/section, freshness and uncertainty. Give a partner a missing-file scenario and ask where it stops.

5. Step 5

   Define five agent tests: normal KPI query, domain-target question, absent source, conflicting revenue note and malicious embedded instruction. Add a request to distribute the report; the response must require a human rather than send it. Record expected results before running the agent tomorrow.

6. Step 6

   Save your canvas and ranking. Write tomorrow's acceptance statement: “I can produce and verify a dated draft, repeat it without duplicates, observe one local scheduled run and pause it.” Record missing licence or connector approvals now so the facilitator can help before Day 2.

Expected output and verification

A canvas a colleague can implement without guessing who approves a report or which sources are allowed. The default is a simple workflow with one evidence agent. Extra agents require a clear benefit demonstrated by tests and measured against latency/cost.

Your cost model includes setup, usage/licence allocation, review and maintenance. Your failure policy distinguishes a source problem from “Claude is broken.” A source gap must not become a success notification.

Troubleshooting

If every task scores highly, ask which has a known source and reviewer today. If the canvas says “access all company systems,” narrow it to named sources. If a multi-agent proposal only duplicates the same analysis three times, begin with one agent and a human reviewer.

Completion checklist

• Opportunity ranking and full canvas saved.
• Evidence handoff format and six expected behaviours written.
• Setup gaps documented for Day 2.

Optional evening practice: use the Skill once more on the same safe source pack and record one success and one failure to discuss in the recap.