Delegation Review

A cumulative review of Chapters 06-08 (Delegation Not Prompting, The Always-On System, When APIs Don’t Exist). Ten questions, drawn from across the chapters and deliberately mixed together — interleaved practice is harder than chapter-by-chapter review, and that difficulty is exactly what builds retention.

How to use this review:

Pass bar: 9 of 10. Higher than lesson quizzes, because this is where knowledge proves it stuck.
Answer in writing, in your own words, before checking anything. Constructing the answer is the exercise.
With a tutor: it will run the review one question at a time, grade against the key, and generate fresh variants for any retake.
If you miss the bar — or barely clear it — come back tomorrow, not in five minutes. Spaced retakes are dramatically better for retention than immediate ones. The gap is a feature.

Check your understanding

Q1. Four tasks: (a) refactor using your local credentials and local MCP servers, (b) a long research job that must survive your laptop sleeping, (c) processing PDFs sitting in your desktop Downloads folder while you’re away, (d) turning a Slack thread into a work session. Assign each its delegation mechanism and give the one-line reason.

Q2. Trace one piece of knowledge through its whole operational life: dictated rule → skill → used by a routine → corrected after a mistake. Name the mechanism at each arrow.

Q3. You’ve been delegating for a month; your direction-to-execution ratio is 1:1. Diagnose two likely causes and the fix for each.

Q4. Explain what claude mcp add quickbooks --scope project gives a five-person team that the same command without the flag does not.

Q5. The Monday routine runs in the cloud and works; the invoice-PDF job fails there. Explain the difference, and where each kind of job must live.

Q6. Describe a channel-based workflow for something in your world: the event, what pushes in, what Claude does with which capabilities, and what arrives on your phone. Then say what you stopped doing manually.

Q7. A vendor portal has no API. Rank the automation options you’d try in order, and say what determines when you stop descending the hierarchy.

Q8. Why is ‘read the PDF natively, apply the skill, enter data via the browser’ better engineering than doing the entire invoice task through Computer Use — even though Computer Use could do all of it?

Q9. Where does the human approval gate belong in an automated purchasing flow, and what three things should the approval message contain?

Q10. Your teammate says ‘I corrected Claude twice today on the same thing.’ What system failure does that sentence reveal, and what is the one-sentence practice that fixes it?

Answer key — attempt every question first

Answer key

Q1

Model answer: (a) Remote Control — needs your machine’s credentials and local MCP; (b) cloud session — survives the laptop, runs on hosted infra; (c) Dispatch — spawns a session on your paired desktop with its local files; (d) @Claude in Slack — thread becomes session context.

Pass criteria: all four matched; reasons reference the distinguishing property (local deps, persistence, paired desktop, thread context)

Q2

Model answer: Dictation captures it (phone) → formalized into SKILL.md (structure + trigger) → the routine loads the skill and applies it to live MCP data on schedule → the correction is written back into the skill (‘update the skill’), so the next run inherits it. Arrows: capture, formalization, automatic loading + data pairing, persistence of corrections.

Pass criteria: all four stages with the mechanism at each transition; correction explicitly flows back into the skill

Q3

Model answer: Likely causes: tasks delegated too small (watching each step instead of specifying outcomes), and/or missing infrastructure (knowledge not in skills, data not connected — so every task needs hand-holding). Fixes: delegate outcome-sized chunks with checkpoints; invest in the skill/MCP substrate so sessions run without you.

Pass criteria: two distinct causes (granularity, substrate); matching fixes

Q4

Model answer: It writes the server config to .mcp.json in the repo instead of personal settings — committed to git, so all five people get the identical connection on pull, with env-var credentials resolving per person. Without it: five separate manual setups that drift.

Pass criteria: project-file-in-git mechanism; team-parity consequence; per-user credentials

Q5

Model answer: Cloud routines see the repo and hosted/remote MCP servers, and run with your laptop closed — the Monday report’s data (hosted QuickBooks connection) is reachable. The PDF job needs a local filesystem, so it must run on the machine: a local scheduled task, or Dispatch to the desktop.

Pass criteria: resource-locality distinction; each job placed correctly

Q6

Model answer: Any coherent design: real event → webhook/channel pushes in → Claude applies a skill and touches MCP/tools → a summary-plus-approval lands on the phone. The stopped activity must be a checking behavior (inbox, dashboard, portal).

Pass criteria: all four components; capabilities named; a pull behavior explicitly retired

Q7

Model answer: Order: look for any API or MCP server first; then a CLI; then browser automation (Claude in Chrome / Chrome DevTools MCP); Computer Use last. You stop descending at the FIRST level the system actually supports — descend only when the level above does not exist, never for convenience.

Pass criteria: correct order; stop-at-first-available rule stated

Q8

Model answer: Each step runs at its highest available level: native file reading is instant and exact; the skill applies rules deterministically; only data entry, which truly lacks an API, pays the browser cost. All-Computer-Use drags every step down to the slowest, most fragile level and multiplies error surface.

Pass criteria: per-step-highest-level principle; cost of flattening to screen level

Q9

Model answer: The gate goes before money moves — after the order is assembled, before submission. The message carries: what will happen (items, vendor, total), the evidence (stock levels, the rule that triggered it), and any anomaly flags. One decision, full context.

Pass criteria: gate before the irreversible action; message = action + evidence + flags

Q10

Model answer: Corrections are not being persisted — they live and die in conversations, so every session relearns nothing. Practice: end every correction with ‘update the skill (or CLAUDE.md) so you know this next time’ — the correction must land in a file the next session loads.

Pass criteria: diagnoses unpersisted corrections; the persist-it practice stated