Lesson 2 — A1 Specification: Writing a Request Claude Can Actually Use

Learning objective

Bloom's level: Apply. Write a request to Claude that states your goal, what you're handing it, what's fixed, and what "done" looks like — before you ask it to draft anything — on a real quiz question you need finished this week.

Plain language: Ask clearly enough, up front, that Claude doesn't have to guess what you actually need — the same way an apprentice doesn't have to guess what a job requires if you hand over a real work order instead of "go wire that room."

Standard: AIHC.1.A1 — "Write task requests that state goal, inputs, constraints, and what 'done' looks like before handing work to an AI coworker; when output misses, revise the request before re-rolling the output." Plain words: asking clearly.

Your real artifact today: the box-fill quiz question (NEC Article 314.16) you still need to finish for this week's classroom sheet, plus service-entrance conductor sizing for the independent at-bat.

SSD cycle 1 — the request is a work order, not a wish

• SAY: A vague ask gets a vague answer, from Claude same as from an apprentice. Anthropic's own getting-started guidance says the tool works best when you "speak to Claude like you would a coworker or friend — naturally and conversationally," and that "more specific requests get better-targeted help" (doc-set §3, source S1). A specification — the technical term for what the standard is asking you to write — is nothing more than a work order: what the job is (goal), what materials you're supplying (inputs), what's fixed by code and can't move (constraints), and what a passing inspection looks like (done-criteria). Four things. Same four things you'd never send an apprentice out without.
• SEE: A two-column card. Left, "VAGUE ASK": "give me a box fill question." Right, "WORK-ORDER ASK": "Write one multiple-choice box-fill question for my apprentice classroom, using NEC Article 314.16. Inputs: a 4x4x1½ square box, two 12 AWG cables, one switch device. Constraints: match my classroom's existing format (question, four choices, cited article number). Done: the correct answer's box-fill count is shown with its math, not just stated." Below each: a plausible illustrative Claude response — the vague ask produces a generic question with no article citation and an unverified count; the work-order ask produces a question matching his format with a shown, checkable calculation.
• DO: Before reading further, actually type the vague version — "give me a box fill question" — into your own free claude.ai account and read what comes back. Notice, in your own words, what it left for you to guess.

SSD cycle 2 — when the output misses, fix the request, not the output

• SAY: The standard's second half matters as much as the first: "when output misses, revise the request before re-rolling the output." Anthropic's own advice backs this exactly — "don't hesitate to refine a request or ask a follow-up rather than starting over" (doc-set §3, S1). This is the same move as a failed rough-in: you don't just tell the apprentice "try again" and hope — you correct the instruction (which article, which dimensions, show your count) and have the same job redone, in the same conversation, so the correction actually lands.
• SEE: An annotated exchange: Claude's first answer to a half-specified ask states a box-fill count without showing the math — the number happens to be right, but there's no way to check it from the answer alone. Instead of re-asking from scratch, the request gets ONE line added: "show your box-fill count as a calculation, not just a number." The redo, in the same thread, now shows the math — checkable.
• DO: Take your real box-fill question in progress. If Claude's first answer to you (from cycle 1) skipped showing its math, add exactly that one line to your existing message and ask again in the same conversation — free-tier claude.ai keeps "memory across conversations" (doc-set §4, S15), so you don't have to re-type your classroom's format from scratch.

SSD cycle 3 — plan the whole ask before you start typing

• SAY: Anthropic's own usage-best-practices page gives advice that happens to double as good specification discipline: "Start by planning your conversations," "Be specific and concise," and "Batch similar requests in one message" rather than sending them one at a time (doc-set §5, source S3). For a full quiz sheet — several questions, not just one — that means handing over the whole batch in one clear message: the topic, how many questions, your format, and what "done" looks like for each — not five separate back-and-forths. It also happens to respect something real about the free plan: usage resets on a session basis every five hours (doc-set §5, S1) — Anthropic doesn't publish an exact message count for that window (a real gap in their own documentation, not this lesson's guess), so the practical habit is the same either way: plan the batch, don't burn the session trickling it in one line at a time.
• SEE: A "one clean message" card holding all of: topic (box fill, Art. 314.16), quantity (3 more questions), format (his classroom's), and done-criteria (shown math, cited article) — next to a "five separate messages" version asking for the same three questions one at a time, with a session-reset icon marking that the second path spends more of the same window for an identical result.
• DO: Batch your remaining real box-fill questions — however many you still need for this week's sheet — into one message with all four parts filled in, and send it once.

Independent at-bat (lighter scaffold than lesson 1 — no filled-in example given this time)

Pick a topic you haven't touched yet in this unit: service-entrance conductor sizing. Write one full work-order request, unscaffolded — goal, inputs, constraints, done-criteria, batched into a single message — for one quiz question on that topic, and send it. Use only the four-part shape from today; no template is provided this time.

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

1. (Remember) Name the four parts of a work-order-style request.
2. (Understand) Why does adding one missing line to an existing request usually beat starting a brand-new chat from scratch? One sentence, grounded in Anthropic's own advice.
3. (Apply) Here's a co-committee-member's vague ask, brought to you for a second opinion: "make me some quiz stuff on grounding." Rewrite it as a work order.
4. (Apply — objective level) Show your actual sent message for the service-entrance conductor question from the at-bat, and label which words in it are doing the job of goal / inputs / constraints / done.

Graded against your own sent requests and Claude's actual replies — not against how good the request merely sounded to you before sending it.

Ledger write

ledger_append:
  learner_id: L4-PUB-FRANK
  lesson_id: L4-lesson02-a1-specification-20260719
  standard: AIHC.1.A1
  exit_ticket: {score: "", bloom_reached: apply}
  auto_mastery: "0.15 ->  (projected ~0.45 — first full DO cycle completed, work-order habit demonstrated twice in-lesson)"
  self_score: ""
  calibration_gap: " — not computed until a real self-score exists"
  journal_entry_prompt: >
    In your own words: what's the one line you added to a Claude request today that you would
    NOT have thought to add a week ago? Write it like the note you'd leave an apprentice on
    a corrected work order.
  structure_used: PBL
  referents_used: [electrical-trade code-inspection culture — work-order / rough-in / inspection framing]
  next_lesson_seed: "A2 Delegation — what to hand over at all, now that he can ask clearly; batch of subtasks from a real GFCI/AFCI sheet."

Rubric self-audit (R1–R13)

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