Build and schedule the report

Outcome and preparation

Day 2 · 13:00–14:15 · 75 minutes. Assemble the workflow and observe a local scheduled run. You need the tested agent, Skill and calculator. Use 5 minutes for the sequence, 50 for runs/scheduling and 20 for recovery checks.

A saved schedule, a manual Run now, and a clock-triggered execution are three different pieces of evidence. Record each separately. Local schedules require Claude Desktop to be open and the machine awake. Cloud routines use different source access and are a later extension, not a property of this local MVP.

Step by step

1. Step 1

   Invoke /executive-report with the explicit historical dates below. Ask it to delegate extraction, run the calculator, inspect the results and save its narrative separately as outputs/2026-10-05/narrative-draft.md. Review tool requests.

   Prompt / exampleCopy

   /executive-report
   Use period 2026-10-05, comparison 2026-09-28 and historical as-of 2026-10-12 UTC.
   Delegate evidence collection to evidence-extractor. Run report.mjs for calculations.
   Preserve report.md, calculations.json and run.json. Save the AI narrative as narrative-draft.md.
   Add exact sources, options, unknowns and a human-review checklist. Do not distribute.
   

2. Step 2

   Compare the narrative against the calculator and Module 5 brief. Check all numeric claims against calculator keys, the three target breaches and the absence of a claimed cause for churn. Reject extra deltas calculated from rounded values and unsupported predictions that incidents will continue; ask for a corrected draft. Rerun: the deterministic calculator must say unchanged. Record narrative revisions separately; they must not alter the baseline files.

3. Step 3

   In Code Desktop open Routines → New routine → Local. Name it Northstar weekly report — practice. Select this practice folder, your available model and normal reviewed permissions. For this local folder exercise leave isolated worktree off; a fresh worktree may not contain uncommitted sources. Set Weekly → Monday → 09:00 in your local timezone and record that timezone. Paste the rehearsal instruction below.

   Prompt / exampleCopy

   Historical workshop rehearsal only; these inputs are fictional.
   In the selected practice folder use executive-report with period 2026-10-05,
   previous 2026-09-28, as-of 2026-10-12 UTC.
   Validate sources, delegate extraction, run the calculator and verify the draft.
   For unchanged inputs report unchanged; never overwrite a completed report after source changes.
   Keep drafts local. Do not send, publish or modify inputs. Flag missing evidence or permissions.
   Record whether output was created, unchanged, blocked or failed.
   

4. Step 4

   Save and reopen the routine to verify folder, instructions, timezone and cadence. Click Run now, inspect its session/tool requests and confirm the resulting output path/status. Do not grant broad permanent permissions just to remove prompts; record any permission that would prevent unattended execution.

5. Step 5

   Temporarily choose Edit → Custom and enter an hourly cron expression whose minute is several minutes ahead: for example, 25 * * * * fires around minute 25 of each hour. The Desktop UI inspected for this workshop rejected */5 * * * * with an hourly-minimum message; official documentation describes finer local intervals, so check your installed version. The hourly expression works with the inspected picker. Check the displayed next-run time and allow for its randomized delay. This repeats hourly until you pause it; it is not a one-time schedule. Keep the app open and machine awake. Wait for the actual scheduled session, then inspect history and its timestamp, tools and output. Log clock-triggered observed only if it really fired without clicking Run now. A session may call itself “scheduled” even after Run now; verify the trigger from your action log and history timestamps. If the slot overruns, finish this check after class and leave it pending.

6. Step 6

   Test recovery in a scratch copy: omit service-health.csv; use as-of 2026-10-26 for stale data; change an input after a completed run. Each should stop rather than create a false success. Ask how a missed run, expired connector or permission prompt is surfaced. Inspect history rather than assuming every missed week is replayed.

7. Step 7

   Restore the weekly schedule, then Pause the practice routine. Verify Paused, resume briefly to confirm the control works, and pause again. Save its configuration and execution evidence in handover/rehearsal-log.md. For a real pilot replace historical dates with latest completed-week selection and approved fresh exports before activation.

Expected output and verification

A dated report and narrative, evidence trace, unchanged rerun and routine history showing an actual scheduled execution. The run record's draft status is not human approval. Record manual and scheduled execution separately and leave unknown cost as unknown.

For a real routine, derive the current period from the run's actual date using Monday–Sunday weeks; select its preceding week and reject stale exports. Use UTC for source periods and explicitly record the schedule's local timezone. Do not quietly run the October fixture forever.

Troubleshooting

If the schedule does not fire, inspect app version, awake state, prior runs, history and permission prompts. If a changed input blocks a rerun, investigate and archive the old report manually; do not delete evidence to make a demo pass. If reliable execution while the laptop is off is required, assess cloud routines with cloud-accessible sources, permissions and cost as a separate pilot.

Reference: Claude Code Desktop scheduling. Interfaces and missed-run behaviour can change; the facilitator checks this before each cohort.

Completion checklist

• End-to-end output and unchanged rerun verified.
• Run now and actual scheduled-run evidence recorded separately.
• Recovery scenarios tested; remaining gaps explicit.
• Practice routine is paused at handover.