Lesson 4 — A3 Elicitation: Asking Claude What It Would Do, and Where It's Unsure

Learning objective

Bloom's level: Apply. Ask Claude, on a real quiz question and on your clipping's central claim, what part of its own answer it would flag as least certain — and use that flag as your starting map for what to check first.

Plain language: Ask Claude what IT would suggest, or where it's unsure — and treat the answer as something to go inspect, not a courtesy question you're asking to be polite.

Standard: AIHC.1.A3 — "Ask the AI coworker for its read, its proposal, or its uncertainty before finalizing an approach, and treat 'what would you do?' as a working question rather than a courtesy." Plain words: asking Claude what it thinks, and inspecting the answer.

Your real artifact today: the working-clearance quiz question (NEC 110.26) you're drafting, and the clipping's central claim about a study finding "up to 40%" of electrical-inspection work automatable within a decade.

SSD cycle 1 — the working question, not the courtesy question

• SAY: Anthropic's own advice is to talk to Claude "like you would a coworker or friend — naturally and conversationally" (doc-set §3, source S1). A real coworker, you'd ask "which part of this are you least sure about" as a completely normal, working question — not small talk. That's exactly this anchor: solicit Claude's own uncertainty and use it as real input, the same move as a crew chief asking an installer "which run are you least sure passed the pull test?" before deciding where to spend a limited day's worth of inspection time — you don't inspect every run at random, you start where the installer already flagged doubt.
• SEE: A described exchange: Frank asks Claude to draft a working-clearance question for a 240-volt panel, then adds a single follow-up — "which part of your answer are you least sure of, and why?" The illustrative reply flags the specific clearance depth for that voltage/condition combination as the part it holds least confidently, because clearance dimensions shift with voltage class and installation condition (Table 110.26).
• DO: Send your real working-clearance question request to Claude, then send the follow-up exactly as written above. Read the flag it gives you.

SSD cycle 2 — the flag is a map, not a verdict

• SAY: A self-reported uncertainty flag is not proof of an error — it's information about where confidence and correctness might come apart, which is exactly the gap this whole unit keeps returning to. The flag tells you where to spend your checking time first (lesson 5 is that checking step); it doesn't replace checking, and it doesn't mean the flagged part is wrong or the unflagged parts are safe. It's a map, not a verdict — same as an apprentice's "I'm least sure about that torque spec" tells you where to look first, not whether the joint actually fails.
• SEE: A described exchange on your OTHER real artifact: Claude explains the clipping's claim in plain English, then answers "what part of this claim, if any, would you be least confident restating as fact?" — the illustrative flag lands on the specific "40%" figure and the unnamed "a study found," while the broader trend claim (automation affecting skilled trades over time) is treated as more widely discussed.
• DO: Run this exact elicitation on the clipping's claim yourself: ask Claude to explain it, then ask the "least confident" follow-up. Write down, in one sentence, exactly which part got flagged.

SSD cycle 3 — asking even when you already have a plan

• SAY: The standard's point cuts both ways: elicitation isn't just for when you're stuck. Asking "is there a different way you'd approach this, and why" on something you already consider finished still counts as a working question, not wasted motion — Anthropic's own capability framing includes "research (gathering information across sources)" (doc-set §2, source S9), which is exactly the kind of alternate-angle Claude can surface even on a question you already like. And per the standard, a divergent answer is data you weigh — not something you're obligated to adopt.
• SEE: A described exchange: Frank already considers his box-fill question (from lesson 2) done. He asks anyway: "is there a different way you'd approach this question, and why?" The illustrative reply proposes adding a cross-reference to the box-fill table for a different box size, as an optional second part — Frank reads it, and the SEE shows him noting "not using it this time — my classroom sheet only tests one box size per question," which is itself the correct move: weighing the divergence, not automatically taking it.
• DO: Pick a question you already consider finished from an earlier lesson. Ask Claude the "different way you'd approach this" question. Write one line: what it proposed, and whether you're adopting it or not — either answer is fine, as long as you can say why.

Independent at-bat (no template this time — just the two moves, named)

Pick a question still in progress (your overcurrent-sizing sheet from lesson 3 works well). Run both moves from today, unscaffolded: (1) ask what Claude is least sure of in its own answer, and (2) ask what it would do differently. For each, write down what came back and what you did with it.

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

1. (Remember) Name the two elicitation moves practiced today.
2. (Understand) Why is an uncertainty flag a map and not a verdict? One sentence.
3. (Apply) Here's a fellow committee member's situation: he already has a finished answer key and doesn't think he needs Claude's input anymore. What would you tell him about asking anyway?
4. (Apply — objective level) From your at-bat: which of the two flags — "least sure of" or "what would you do differently" — actually changed what you plan to check first in lesson 5, and what specifically will you check because of it?

Graded against your own recorded exchanges, not against whether Claude's flagged part later turns out to be an actual error — that verdict is lesson 5's job.

Ledger write

ledger_append:
  learner_id: L4-PUB-FRANK
  lesson_id: L4-lesson04-a3-elicitation-20260719
  standard: AIHC.1.A3
  exit_ticket: {score: "", bloom_reached: apply}
  auto_mastery: "0.05 ->  (projected ~0.35 — largest single-lesson jump so far, matching the diagnostic's flag that this anchor had no prior trade parallel)"
  self_score: ""
  calibration_gap: " — not computed until a real self-score exists"
  journal_entry_prompt: >
    Tell it like you'd tell a bowling buddy about a spare you almost missed: what did Claude
    flag as unsure that you wouldn't have thought to double-check on your own?
  structure_used: PBL
  referents_used: [electrical-trade code-inspection culture — crew-chief / pull-test framing; bowling league — journal-register bridge only]
  next_lesson_seed: "B1 Verification — check the flagged clearance dimension and the flagged '40%' figure against actual independent sources; this is where the map from lesson 4 gets walked."

Rubric self-audit (R1–R13)

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