Most trades operators who fail at AI in 2026 fail for the same reason: they type a half-sentence into a box and judge the entire technology by what comes back. "Write me a rebuttal for price-shoppers." "Should I send Jose to Marin Park?" "Make a Friday recap." The answers come back vague, generic, and useless — so the operator concludes the AI is dumb. The AI is not dumb. The prompt is dumb. There is a five-part structure used by every CSR who lifted booking percent from 65 to 85, every dispatcher who cut windshield time 18 percent, every Comfort Advisor who closed an extra eight points on jobs above $5,000 with AI in the room. Role, context, task, format, constraint. Five parts. Short. Specific. Repeatable. No "act as a world-class expert in HVAC sales who has won every award." Just the five fields that turn a guess into a usable draft in 14 seconds. This lesson is that structure, with worked examples on the truck, in the dispatch chair, at the kitchen table, on the CSR row, and at the owner's Friday-afternoon desk.

Why Five Parts, and Not Three or Seven

Every trades AI vendor sells a prompt library now. Avoca ships system prompts behind the voice agent. Rilla ships coaching prompts behind the scorecard. ServiceTitan ships Titan Intelligence prompts behind the dispatch summary. Hatch ships nurture prompts behind every text. Every one of those prompt libraries has the same shape, because the shape is forced by how the underlying model actually works. Strip the marketing and there are five fields the model needs to produce output a journeyman will sign off on: who is asking, what situation are we in, what do I want produced, in what shape, and what must I not say. Role, context, task, format, constraint.

Three parts is too few. "Write a rebuttal for a price-shopper" has a task but no role, no context, no format, and no constraint — so the AI guesses the shop is a $79-diagnosis HVAC company in Texas with a flat-rate book, and produces a generic rebuttal that may or may not match the shop's actual voice or financing tiers. Seven parts is too many. The CSR will not type seven fields between calls. Five is the count that survives a Tuesday-morning huddle, fits on an index card, and produces usable output every time. Five is also the count Avoca, Rilla, ResponsiBid, Hatch, and Titan Intelligence converged on by 2026 — not by committee, but by what actually worked in 50,000+ shop deployments.

The discipline is filling them in even when you are in a rush. The CSR who pastes "price-shopper, $79 diagnosis, no free estimates, 12-year shop, Memphis, Wisetack financing" gets a usable rebuttal in 12 seconds. The CSR who types "price-shopper rebuttal" gets a generic paragraph that takes 90 seconds to rewrite. The cost of the five-part habit is 8 seconds of typing. The benefit is the difference between AI as a useful apprentice and AI as a frustrating toy.

The Five Parts, Named and Defined

Here is the structure. Tape it to the wall above the CSR row, the dispatch board, and every Comfort Advisor's tablet. The same five fields work for every role in the shop. The content inside each field changes; the structure does not.

Part One: Role

Tell the AI who is asking and who is being talked to. Not "act as a world-class HVAC genius." Real role. "I am a CSR at a 6-truck residential HVAC shop in Memphis." "I am a Comfort Advisor walking into a Bel Air kitchen for a $24,000 dual-fuel replacement." "I am a dispatcher reshuffling the board after a 2 p.m. call cancel in Marin Park." Real role anchors the AI's vocabulary, tone, and reference frame. The AI has read a million sentences from CSRs and a million sentences from PhD researchers; the role tells it which corpus to lean into.

Role also names the other side of the conversation. "Customer is a 62-year-old homeowner whose furnace died Sunday night, just got the diagnostic, hesitant on price." That sentence tells the AI more than a thousand-word system prompt about HVAC. The role field is the single highest-leverage line in the whole prompt; skipping it is the most common reason trades AI output sounds generic.

Part Two: Context

Tell the AI what just happened, what the situation contains, and any specifics that change the answer. This is where shop voice lives. "Customer just got a $1,400 repair quote, said 'I'm just calling around for prices,' last service was 2019 on the same furnace, financing through Wisetack and GreenSky available, shop policy is no free estimates and $79 diagnostic, brand voice is friendly and direct, never hard-sell." Six lines, twelve seconds, and the AI now writes a rebuttal that uses Wisetack and GreenSky correctly, respects the $79 policy, matches the shop's voice, and references the 2019 service if the rebuttal turns into a relationship play.

Context is also where the worked-yesterday detail goes. "Yesterday we lost 4 price-shopper calls on the same objection — they all said 'the other guy quoted $899.'" Now the AI has a real shop pattern to write against, not the generic patterns from its training corpus.

Part Three: Task

One sentence. What do you want the AI to produce. Not "help me with this," not "make this better" — the actual verb. "Write a 4-sentence phone rebuttal." "Draft a 60-word on-my-way text." "Score this Rilla transcript on the 5 close-rate moments." "Build a 3-option good/better/best summary for the proposal." "Generate the Friday recap email to the owner in 7 lines."

Vague task makes vague output. "Help with the rebuttal" produces a 400-word essay the CSR cannot use on a 90-second call. "Write a 4-sentence rebuttal the CSR will read off-screen in under 15 seconds" produces a 4-sentence rebuttal. The verb-plus-shape combination is what makes the task field useful. Trades operators who internalize this stop asking AI to "help" and start asking it to "write 3 sentences," "draft 1 paragraph," "list 5 options," "summarize in 7 lines." Output shape is half the battle, and the task field is where you put it.

Part Four: Format

How should the output be structured. Bulleted list, numbered options, table, single paragraph, on-screen card, text the CSR can read aloud, email-ready draft. Trades work is operational, not academic; the format field is where you tell the AI to produce something a human can act on without restructuring. "Output as 3 bullets the dispatcher can read on the board." "Output as a 4-sentence script the CSR can read in 18 seconds." "Output as a 5-row table with Equipment / Age / Repair Cost / Replace Cost / Recommendation." "Output as an email subject line plus 3-paragraph body the owner can forward without editing."

Format also controls the screen real estate the role works on. A Comfort Advisor on a 10-inch tablet does not want a wall of text. A dispatcher reshuffling at 11:45 a.m. wants the answer in 2 bullets and the rationale below. Tell the AI explicitly. Skip the format field and you spend 30 seconds reformatting before you can act on the draft. Over a day of 80 prompts, that is 40 minutes lost — exactly the time AI was supposed to save.

Part Five: Constraint

What the AI must not do, must not say, or must not assume. This is the field that separates a sloppy AI shop from a verify-disciplined one. Constraint comes in two flavors. First flavor: the forbidden-promise list. "Do not promise free estimates — shop is $79 diagnostic." "Do not promise same-day — territory is booked." "Do not invent rebate amounts — only use the figures I paste from the utility page." "Do not cite NEC code sections — I will pull the exact citation from the AHJ." "Do not draft financing APR or payment math — I will pull from the Wisetack portal."

Second flavor: the brand-voice and tone constraints. "Do not use the words 'partner' or 'family' — sounds corporate." "Do not use exclamation points — we are direct, not chirpy." "Do not address the homeowner by first name — we are a Southern shop and we use Mr./Ms." "Keep it under 60 words." "Do not include emojis." "Match the tone of the example I pasted." The constraint field is where shop-specific judgment lives. It is the part of the prompt that turns a generic LLM into a tool that sounds like the shop, respects the shop's policies, and stops short of the Reg Z and FCRA lines AI is structurally unsafe to cross.

Worked Example: The CSR's Objection Rebuttal

It is 9:14 a.m. on a Tuesday in Memphis. The CSR has just hung up on her third price-shopper of the morning. She has 90 seconds before the next call. She opens the shop's AI tool — could be ChatGPT, could be the Avoca system-prompt window, could be Titan Intelligence inside ServiceTitan, the structure is the same. She types five fields:

Role: CSR at a 6-truck residential HVAC shop in Memphis. Customer is a homeowner who already got a $1,400 repair quote from us last week and is now calling back saying he is "just shopping around."
Context: Shop policy is $79 diagnostic, no free estimates. Financing through Wisetack (0% for 12 months on $1K+) and GreenSky (84-month terms available). Shop voice is friendly, direct, never hard-sell. Last 3 price-shopper losses cited "the other guy quoted less" — we need a rebuttal that re-anchors to value, references our 12-year track record without trash-talking competitors, and offers a financing path before the customer hangs up.
Task: Write a 4-sentence phone rebuttal the CSR will read aloud in 18 seconds.
Format: Output as 4 numbered sentences, no preamble. Include one open-ended question at the end that re-engages the customer.
Constraint: Do not promise free estimates. Do not invent financing rates beyond what I gave you. Do not use "partner," "family," or exclamation points. Stay under 65 words total.

The AI gives back a 4-sentence rebuttal that uses Wisetack 0% accurately, re-anchors to the 12-year track record, ends with "What's pushing the timing on getting this fixed?" and does not promise a free estimate. The CSR pastes it into her cheat sheet. Total time: 38 seconds. By Friday she has 6 rebuttals built this way, and her booking percent has moved from 67 to 79.

Worked Example: The Dispatcher's Re-Route Decision

It is 11:45 a.m. The Marin Park 1 p.m. install just cancelled — homeowner's sister is in the hospital, they need to reschedule for next week. Jose was on the install with two helpers. The dispatcher has 75 minutes to redeploy or lose the day. ServiceTitan Dispatch Pro suggests pulling Jose to a Bel Air diagnostic at 2 p.m. The dispatcher knows Jose just finished a callback on that street yesterday and the homeowner specifically asked for him. He also knows the rookie tech is sitting idle in the south territory and has not hit revenue all week.

Role: Dispatcher at a 6-truck residential HVAC shop. Decision is 11:45 a.m. with Marin Park 1 p.m. install canceled (Jose plus 2 helpers).
Context: Bel Air diagnostic at 2 p.m. is on the board with homeowner who specifically asked for Jose. Rookie tech (Marco) is idle in the south territory, has not hit revenue target this week. Install crew helpers can be reassigned to a stacked maintenance route in the east territory. Comp plan rewards revenue per truck, not call count. Dispatch Pro suggests Jose to Bel Air, but the model does not see customer-name request or the rookie's revenue gap.
Task: Produce 3 ranked options for the next 4 hours with a one-sentence rationale for each, plus the override risk if Dispatch Pro's auto-route is followed.
Format: Output as a 3-row table with columns Option / Action / Rationale / Risk. Add a 1-sentence overall recommendation below the table.
Constraint: Do not assume comp data I did not give you. Do not invent customer history beyond what I stated. Do not recommend overtime — shop policy is no overtime without owner approval before 3 p.m.

The AI gives back a 3-row table. Option 1: Jose to Bel Air, Marco kept on south — preserves customer request, costs install yield. Option 2: Marco to Bel Air, Jose home early — costs the customer-named close, gains rookie revenue. Option 3: Jose to Bel Air, Marco shadow as ride-along coaching — preserves both, costs 30 minutes of close friction. The dispatcher picks Option 3. The call closes for $19,400. Marco watches a real kitchen-table from a senior tech for the first time. The override is logged with the reason.

Worked Example: The Tech's Job-Notes Voice Prompt

The tech just finished a 2-hour diagnostic on a Goodman heat pump in a 14-year-old install in Bel Air. The homeowner is on a payment plan, the system has a leaking line set, and the tech recommended repair-or-replace. He has 4 minutes before his next stop and needs to leave service-manager-readable notes that will actually be acted on.

Role: Senior tech at a residential HVAC shop. Just finished a Goodman heat-pump diagnostic, 14-year-old install, homeowner is hesitant.
Context: Equipment is a 4-ton Goodman heat pump with R-410A, line set is leaking, evaporator coil shows visible corrosion, condenser fan motor is original and noisy. Homeowner is on a payment plan for an unrelated debt and asked about financing. Repair estimate is $2,400 for line set + leak repair; replace estimate is $9,800 for a 17 SEER2 system, partial replace (outdoor unit only) is $6,200. Customer mood is hesitant but engaged. I dictated the following voice notes: "Goodman heat pump 14 years old, line set leak found about 18 inches off the condenser, coil looks bad lots of green stain, fan motor noisy original, homeowner wants payment options, mentioned Wisetack to him said we'd send numbers."
Task: Structure my voice notes into the shop's standard service-manager-readable format.
Format: Output as 5 labeled sections: Equipment, Condition, Recommendation, Customer Mood, Next Step. Each section is 1-3 sentences. No preamble.
Constraint: Do not invent readings or measurements I did not give you. Do not cite specific code sections. Do not invent the Wisetack approval status — I only said I'd send numbers.

The AI gives back a 5-section service-manager note in 11 seconds. The service manager reads it Thursday morning, surfaces the repair-vs-replace as a Comfort Advisor lead, and the advisor is in the kitchen by Friday. The tech saved 9 minutes of typing. The lead did not fall through the cracks because the notes were structured rather than a 3-line text dump.

Worked Example: The Comfort Advisor's Financing Pivot

It is 7:20 p.m. at a kitchen table in Bel Air. The Comfort Advisor has presented the $24,400 replacement. The homeowner has just said "that's a lot more than I was thinking." The advisor has 90 seconds before the conversation collapses. He pulls his tablet and opens the prompt template.

Role: Comfort Advisor at a kitchen table in Bel Air. Homeowner just heard $24,400 for a heat-pump replacement and said "that's more than I was thinking."
Context: Wisetack soft-pull at the door showed approved up to $22,000 at 84 months 8.99%. GreenSky available for the gap at 9.99% over 120 months. Homeowner is 58, household income roughly $145K, system is a failed 18-year-old furnace plus an aging condenser, current power bill running $340/month in winter. Replacement is a 17 SEER2 dual-fuel with a heat-pump primary. Shop voice is direct and respectful; no high-pressure close tactics. Section 25C credit applies up to $2,000 on the heat-pump portion.
Task: Write a 60-second pivot talk-track that re-anchors from sticker to payment, references the 25C credit and the Wisetack approval, and ends with one open question.
Format: Output as a single paragraph, 100-120 words, conversational tone the advisor will read off the tablet.
Constraint: Do not invent Wisetack APR — use exactly 8.99% / 84 months. Do not promise the 25C credit amount above $2,000. Do not pressure or use scarcity language ("only today," "limited time"). Do not address the homeowner by first name.

The AI gives back a 110-word paragraph that anchors at $310/month after the 25C credit, references the failing furnace's likely repair cost over the next 3 winters, and ends with "What would feel manageable for the monthly piece?" The homeowner asks about the 120-month GreenSky option to drop the payment lower. The advisor pulls up the GreenSky portal, runs the actual numbers, closes at $22,800. The pivot lived in the AI-drafted talk-track the advisor verified in 30 seconds and read in 60.

Worked Example: The Owner's Friday Recap

It is 4:15 p.m. on Friday. The owner has 18 minutes before he leaves to coach his kid's soccer practice. He needs to read the week, decide one priority for Monday, and respond to two emails. He opens his AI tool and pastes the week's numbers.

Role: Owner of a 6-truck residential HVAC shop. Friday 4:15 p.m. recap before the weekend.
Context: Booking percent 78 (up from 73 last week). MPR 31 (flat). Financing close 24 (down from 28). Recall percent 3.4 (up from 2.8 — one tech responsible for 4 of the 6 recalls). Average ticket $612 service / $14,200 replacement. GLSA ROAS 3.9x (down from 4.2). Three Avoca catches this week — one mis-heard name, two forbidden-promise hits the CSR floor stopped. Two open complaints — Bel Air install dispute over linesets, Marin Park warranty question on a 9-month-old condenser. Marco the rookie hit revenue target Tuesday for the first time.
Task: Draft a 7-line owner recap I can read in 90 seconds, plus name one Monday priority and one celebration.
Format: Output as 7 numbered lines plus 2 labeled lines (Monday Priority / Celebration). No preamble.
Constraint: Do not invent numbers I did not give you. Do not propose actions on the open complaints — those need owner judgment, not AI suggestion. Do not soften the recall trend — call it out.

The AI gives back 7 lines that name the recall-trend concern, flag the financing-close drop as the likely GLSA ROAS cause, and recommend recall coaching as Monday priority. The celebration is Marco's first revenue-target Tuesday. The owner reads it in 80 seconds, forwards it to the service manager, makes the soccer practice. Last quarter that same recap took the owner 35 minutes. Without context and constraint, the AI would have softened the recall trend.

Building the Shop's Prompt Library

The five-part structure is most powerful when it stops being one-off and becomes a library. Every role in the shop should have 8-15 saved prompt templates by end of month two — the role's daily situations pre-built with role, format, and constraint locked, leaving only context and task to fill in real-time. The CSR has rebuttal templates for the top 6 objections. The dispatcher has reshuffle templates for the 4 common board-disruption patterns. The tech has voice-notes templates for the 5 service-call shapes. The advisor has financing-pivot, options-explanation, and objection-rebuttal templates. The owner has Friday-recap, weekly-1-on-1, and quarterly-review templates.

The library lives wherever the shop already lives — a shared Google Doc, a Notion page, the prompt-library feature inside Titan Intelligence or Avoca, a pinned Slack post. The shape does not matter; the existence matters. Shops with a real library hit their AI metric targets at month two; shops still typing prompts from scratch at month four are still calibrating. Every catch from the bulletin board feeds back into a constraint update. Every successful close feeds back into a format refinement. Every CSR's best rebuttal becomes everyone's.

The library is also the artifact that survives turnover. When the senior CSR leaves for a higher-paying gig, the 12 rebuttal templates she built are the shop's, not hers. The new CSR onboards in 3 days against a library that took 4 months to build. That is the leverage of the five-part discipline applied across a shop. Not faster prompts — compounding institutional memory.

What the Five Parts Do Not Do

The five-part structure is necessary, not sufficient. It produces a usable draft. It does not produce a verified artifact. The Cardinal Rule from L1 still applies: every output that touches a customer, an estimate, a permit, or a regulator gets a 30-second verify pass with the five checkpoints — numbers, names, parts, warranty, financing/regulatory. The five-part prompt makes the draft better; the verify makes the draft safe. Skip either and the workflow breaks in a different place.

The five parts also do not replace judgment. The dispatcher chose Option 3 over Dispatch Pro's Option 1; the AI surfaced the trade-offs, the human picked. The advisor read the pivot; the human delivered it with eye contact the AI cannot do. The AI is the apprentice; the journeyman picks. The five-part structure makes the apprentice's draft worth the journeyman's time. That is the entire point.

Key Takeaways

• Five fields, in this order, every time: Role, Context, Task, Format, Constraint. Skip any one and the output drifts. Three fields is too few; seven is too many. Five is the count that survives a Tuesday huddle and fits on an index card.
• Role anchors vocabulary; context anchors specifics. "CSR at a Memphis HVAC shop, customer is a 62-year-old homeowner whose furnace died" is the single highest-leverage line in the whole prompt. Skip role and you get generic SaaS prose.
• Task is the verb plus the shape. "Help with the rebuttal" produces a 400-word essay. "Write a 4-sentence rebuttal the CSR reads in 18 seconds" produces a 4-sentence rebuttal. Output shape is half the battle.
• Format controls the screen real estate the role actually works on. Dispatcher gets 2-bullet output; advisor gets a 120-word paragraph; owner gets 7 numbered lines plus 2 labeled lines. Tell the AI exactly. Reformatting is wasted time.
• Constraint is two flavors: forbidden promises and brand voice. "Do not invent Wisetack APR" and "Do not use exclamation points" are the same field — both protect the shop. This is the field that separates a sloppy AI shop from a verify-disciplined one.
• Build the prompt library by end of month two. 8-15 templates per role with role, format, and constraint locked. Context and task get filled real-time. The library compounds; it survives turnover; it is how tribal knowledge becomes shop infrastructure.
• The five-part prompt is necessary, not sufficient. Draft quality goes up; verify discipline still runs. Every customer-, estimate-, permit-, or regulator-touching output gets the 30-second pass with five checkpoints. The prompt makes the draft better; the verify makes the draft safe.