Lesson 3 — A2 Delegation: Deciding What Actually Gets Handed to Claude

Learning objective

Bloom's level: Analyze. Decide, for each real subtask in your in-progress GFCI/AFCI quiz sheet, whether to hand it to Claude or keep it yourself — and state your reason in terms of capability and the cost of being wrong, not habit or hype.

Plain language: Choosing what you'd actually hand the tool at all — and being able to say why, the same way you can say why an apprentice gets one task solo and another only with you standing right there.

Standard: AIHC.1.A2 — "Decide whether to hand a task to an AI coworker from stated reasons about capability and the cost of being wrong — not novelty, habit, or hype." Plain words: choosing what to hand the tool at all.

Your real artifact today: the GFCI/AFCI placement quiz sheet (NEC 210.8 / 210.12) you're mid-way through building for this week's classroom night.

SSD cycle 1 — capability, not novelty or hype

• SAY: The first half of the standard says: decide from capability — can this thing actually do the task well, per what its own maker documents, not because it's new and shiny. Anthropic's own product page names specific capability areas, and one of them is exactly your situation: "learning (explaining complex topics simply, exam and interview prep)" (doc-set §2, source S9). That's a direct, documented match for quiz-question wording and formatting — not a guess, not hype. Trade parallel: you don't decide what a new apprentice does by how excited you are about the new hire — you decide by what they're actually trained to do.
• SEE: A five-row table listing every real subtask in your GFCI/AFCI sheet, with a "Capability match?" column. Two rows filled in as a worked example: "polish the wording of question 3" → YES, matches S9's named "exam prep" capability area; "give final sign-off that this sheet is code-accurate before it reaches the class" → NO per doc-set §6 (S5): Anthropic's own guidance says "users should not rely on Claude as a singular source of truth" for something this consequential.
• DO: Fill in the "Capability match?" column yourself for two more real subtasks from your sheet: "draft a plausible-wrong answer choice (a distractor) for question 3" and "draft the article citations for the answer key." Say yes or no for each, and why, in one sentence.

SSD cycle 2 — capability isn't the whole decision; cost of being wrong is the other half

• SAY: The standard's second half is easy to skip past: the cost of being wrong, separately from capability. A task can be something Claude is genuinely good at and still be one you watch closer, because of what a mistake costs — same as an apprentice who is plenty capable of wiring a receptacle, but you still look twice when it's the one next to the panel's neutral bonding jumper. Capability answers "can it," cost answers "what happens if it's wrong, and how far does that travel."
• SEE: A two-axis grid: Capability (yes/no) across the top, Cost of being wrong (low/high) down the side. Two cells filled in: "polish wording" lands capable+low cost → free to hand over; "draft the answer key's article citations" lands capable+high cost → still hand it over, but that's exactly why lesson 5 (verification) exists — delegation decides WHETHER to hand something over, checking it afterward is a separate, later skill.
• DO: Place the remaining real subtasks from your GFCI/AFCI sheet onto the grid yourself: the distractor-choice draft, and the final code-accuracy sign-off. For at least three subtasks total (across both cycles), state BOTH your capability reason and your cost-of-being-wrong reason out loud, in your own words.

SSD cycle 3 — the one subtask that never moves off "mine alone," and why

• SAY: Final sign-off that the sheet is code-accurate never moves into the delegate lane — not because Claude can't discuss code sections (it can, and well), but because the cost of being wrong there isn't "one quiz question is off." It's an apprentice walking away believing something incorrect about code that could show up on a real job. Worth knowing, as a parallel, not a rule that technically binds a quiz sheet: Anthropic's own Usage Policy requires an extra human check, plus disclosure that AI helped, specifically for legal, financial, and employment-related material that reaches other people (doc-set §9, source S11). Your quiz sheet isn't the large-scale product that clause is written for — but an apprentice's standing in the program is real, and the same reasoning applies at a smaller table: anything that affects someone's actual standing gets a human's signature, every time, no exceptions carved out for a task Claude happens to be capable of.
• SEE: The completed five-row grid from cycles 1–2, with the sign-off subtask drawn OUTSIDE the grid entirely, in its own box labeled "ALWAYS MINE — not a capability question."
• DO: Write, in one or two sentences, the reason final sign-off never delegates — using BOTH words the standard asks for: capability and cost.

Independent at-bat (subtasks are yours to name this time — no pre-built list given)

You're starting a new sheet this week: overcurrent protection sizing (NEC Article 240). Before you write a single question, list the real subtasks involved (wording, formatting, distractor choices, citations, final sign-off — or however you'd actually break the job down), and sort each one using today's grid — capability + cost of being wrong, stated for every item, no template provided.

Exit ticket (Bloom's climb: Remember → Understand → Apply → Analyze, objective level)

1. (Remember) Name the two things the standard says a delegation decision must be based on.
2. (Understand) Why can a task be something Claude is genuinely capable of, and still land in "watch closer"? One sentence.
3. (Apply) The clipping you brought in claims a study found up to 40% of electrical-inspection work could be automated within a decade. Would you delegate "summarize this claim in plain English" to Claude? Would you delegate "decide whether it's true"? State a capability-and-cost reason for each answer.
4. (Analyze — objective level) Look at your overcurrent-sizing sort from the at-bat. Which subtask was hardest to place, and what made it borderline — capability doubt, cost doubt, or both?

Graded against your own stated reasons, checked for whether they actually name capability and cost — not against whether the sort "feels right."

Ledger write

ledger_append:
  learner_id: L4-PUB-FRANK
  lesson_id: L4-lesson03-a2-delegation-20260719
  standard: AIHC.1.A2
  exit_ticket: {score: "", bloom_reached: analyze}
  auto_mastery: "0.12 ->  (projected ~0.42 — full capability+cost grid completed on a real 5-subtask sheet, correctly isolated the never-delegate task)"
  self_score: ""
  calibration_gap: " — not computed until a real self-score exists"
  journal_entry_prompt: >
    Note it the way you'd log a decision for the training committee's minutes: which real
    subtask from this week's sheet surprised you by NOT being an easy yes-or-no, and what
    finally tipped it one way.
  structure_used: PBL
  referents_used: [electrical-trade code-inspection culture — capable-apprentice / watch-closer framing]
  next_lesson_seed: "A3 Elicitation — asking Claude for its own uncertainty on the panel-clearance question and the clipping's 40% stat, now that he knows what he'd hand over at all."

Rubric self-audit (R1–R13)

Escalation: all load-bearing indicators pass → auto-ship.