The five-part prompt is the structure; context is the fuel. A perfectly structured prompt with no context produces generic output. A perfectly structured prompt with the wrong context produces fluent fabrications. A perfectly structured prompt with the right context produces a draft the journeyman signs off on in 12 seconds. The discipline this lesson teaches is the one between those three outcomes: what to paste into the AI, what to leave out, and how to think about the AI's context window like a shop tool with a real edge. We are going to walk through what belongs in context (job notes, equipment tags, last-three-tickets, brand voice samples) and what never belongs (full credit-card numbers, SSNs from financing soft-pulls, customer home addresses outside the dispatched call, jobsite photos with people or kids in them, anything that crosses into PII for which the shop carries fiduciary duty). The line is bright. The shops that walked it in 2026 captured the lifts; the shops that did not are the ones whose state contractor boards now have files open.

The Context Window as a Shop Tool

Every AI tool a trades shop touches in 2026 has a context window. ChatGPT-4o's window is roughly 128,000 tokens — about 250 pages of text. Claude Sonnet 4.5 carries 200,000 tokens by default. Avoca's voice agent loads a tightly tuned system prompt plus the live transcribed call; the window is shorter but the engineering inside it is tight. ServiceTitan's Titan Intelligence carries the platform's structured data plus a focused prompt. Rilla loads the full audio-derived transcript of a call. The window is the AI's working memory for that single conversation. What you paste in is what it can see; what you leave out, it cannot.

The first practical implication: context is not infinite, and even when it is technically large, the AI's attention degrades on long inputs. The 10-year tech in your bay does not lose track of a homeowner halfway through a service call; the AI does. Research published in 2024-2025 showed that even on a 200,000-token window, model accuracy on retrieved facts drops noticeably past ~30,000 tokens — the "lost in the middle" problem. For a trades shop that translates to a hard rule: paste what is relevant to this conversation, not the entire customer history. The dispatcher does not paste yesterday's whole call log into a re-route prompt; they paste the three calls on the affected tech's territory. The CSR does not paste the customer's 14-year service history; they paste the last three tickets plus the active membership tier.

The second practical implication: context is shared. When you paste a customer's address, repair estimate, financing soft-pull tier, and home-equipment list into ChatGPT, you have shared that data with OpenAI under their data-use policy. ServiceTitan's Titan Intelligence inherits ServiceTitan's data agreements; Avoca inherits Avoca's. Public model APIs (ChatGPT, Claude, Gemini) used outside an enterprise data agreement train on inputs by default unless you opt out at the account level. For a trades shop, this is not philosophical — it is a compliance reality. The Cardinal Rule's verify discipline has a companion discipline: the paste discipline. Paste what the conversation needs. Paste nothing that crosses the PII line. Verify the data-handling agreement of every AI tool the shop uses.

What Belongs in Context

The shop's working principle is the smallest sufficient context to produce a usable draft. Not the largest possible. Not "let me dump everything." The smallest context that contains what the AI actually needs. Here is the field-tested list of what belongs.

Job Notes from the Tech

The tech's voice-dictated notes from the truck are first-class context. "4-ton Goodman heat pump, 14-year-old install, line set leak 18 inches off the condenser, evaporator coil shows visible corrosion, fan motor noisy and original, homeowner on Wisetack inquiry." Six lines, twelve seconds of dictation, and the AI now has the equipment state, the recommendation triggers, and the customer's financing posture. The service manager's structured-note prompt (from Lesson 1) turns this into a 5-section read in 11 seconds. The Comfort Advisor's repair-vs-replace prompt uses the same job-notes context to build the kitchen-table talk-track. The estimator's proposal narrative pulls it for the "here's why your system failed" paragraph. Job notes are the single most reusable piece of context in a service shop.

What belongs in job notes: equipment make/model/age, observed condition (with the tech's specific findings, not "looks bad"), recommendation triggers (leaks, corrosion, age, code issues), customer mood and stated goals. What does not belong: the customer's full address, the homeowner's name and phone, the financing soft-pull result, the customer's credit history. The tech dictates the technical situation; the FSM platform carries the customer identity separately. Keep them separated in the AI's context.

Equipment Tags and the Make/Model Map

"4-ton Goodman heat pump, R-410A, dual-fuel pairing with gas furnace, 14-year-old install" is a richer context than "old heat pump." The AI's training corpus has thousands of pattern-matches for the specific brand-model-age combination; it can surface common failure modes, typical warranty windows, parts availability concerns (especially R-410A in the 2025-2026 transition to R-454B), and replacement system fits. The same prompt with vague equipment context produces vague recommendations. Specificity in the equipment tag is the single highest-leverage context line for tech-facing AI use cases.

Brand-specific tags also load the AI's voice differently. A Trane technician's vocabulary differs from a Goodman technician's; a Carrier dealer's proposal language differs from a Lennox dealer's. The shop's brand voice matters; the equipment brand's vocabulary in context matters too. Comfort Advisors who include "preferred replacement is Trane XR16 dual-fuel" in their pivot prompt context get talk-tracks that reference Trane's actual warranty structure and efficiency story rather than a generic HVAC pitch. The specificity compounds.

Customer History — The "Last Three Tickets" Pattern

Full customer service history is too much context; one ticket is not enough. The pattern that works is the last three tickets — most recent service, second most recent, third most recent — with date, equipment touched, work performed, and outcome (resolved/recurring/upsell-offered). Three tickets gives the AI enough to spot patterns (the third coil clean in 18 months — water issue?), enough to surface the relationship arc (member since 2019, financed last replacement through GreenSky), and enough to anchor talk-tracks to the customer's actual recent history. More than three becomes noise; the model's attention dilutes and the surfaced patterns weaken.

What belongs in the three-ticket context: date, equipment, work performed, outcome, advisor or tech on the job, and any open items. What does not belong: the customer's full PII (address, phone, full name), payment method, credit-card last-four, financing application history, soft-pull result history. The three tickets are technical and operational; the customer identity stays in the FSM platform. The shop's privacy posture is to never paste identifying details into an AI tool when an internal identifier (ticket number, customer ID) accomplishes the same thing.

The "last three tickets" pattern is also how the CSR avoids restarting the relationship every call. The AI surfaces the membership status ("Premium member since 2021"), the recent ticket summary ("June 2025 — replaced condenser fan motor, no further issues"), and the open thread if any ("3 months ago — declined the heat-pump retrofit estimate, said spouse wanted to wait until winter"). The CSR's opening line stops being "what's the issue today" and becomes "Mr. Patel — happy to help, I see we replaced your condenser fan motor in June and everything has been working well; what's going on now?" The customer feels known. The booking happens.

Brand-Voice Samples

The AI cannot guess what the shop sounds like. It can match what you show it. Three to five paste-in examples of the shop's actual voice — a real Google review reply that worked, two CSR rebuttals that closed, a proposal narrative from the senior advisor's best close, a Friday recap the owner approved without editing — load the brand voice into context. This is not the same as the "Constraint" field from Lesson 1 ("no exclamation points, no partner language"); constraint is what to avoid, voice samples are what to emulate. Both work together.

The shop's voice library lives next to the prompt library. Senior CSR's best 5 rebuttals. Comfort Advisor's best 5 kitchen-table closes (transcribed from Rilla). Owner's best 3 Friday recaps. Marketing manager's best 5 Google review replies. The AI gets the constraint plus 2-3 voice samples; the output reads like the shop wrote it because the shop's actual examples are in the context window. Generic-sounding AI output is almost always a missing-voice-sample failure.

External References — Manuals, Spec Sheets, Utility Pages

Manufacturer spec sheets, utility rebate program pages, and IRS guidance documents belong in context when the AI's training data is structurally stale. R-454B charge weights, current Section 25C cap interpretations, current state heat-pump rebate amounts, current Wisetack/GreenSky/Synchrony promo windows — all of these change faster than model training cycles. Pasting the relevant excerpt from the actual current source into context grounds the AI on verified current data, not its 2024 memory. This is the trades-shop version of retrieval-augmented generation, done manually by copy-paste, and it is the difference between a usable proposal and a Reg Z lawsuit.

What Never Belongs in Context

The PII line in 2026 is bright and not negotiable. The shop carries fiduciary duty on customer data. State data-protection regulations (CCPA in California, CTDPA in Connecticut, VCDPA in Virginia, and a growing list) impose direct obligations on shops that hold and transmit customer data. AI tool data-handling agreements vary; even when they are clean, the shop's discipline is to never test whether they are. The negative list is short, specific, and absolute.

Credit Card Numbers

Full or partial credit-card numbers do not belong in any AI prompt, in any tool, ever. The payment data lives in the payment processor (Stripe, Square, ServiceTitan Payments, Authorize.Net) under PCI DSS scope. Any AI tool that ingests credit-card data drags the shop into PCI compliance scope. The fix is simple: AI tools never see payment data. The Comfort Advisor pastes "approved up to $22K at 8.99% / 84 months" — not the card number, not the last four, not the bank routing. The CSR pastes the customer ID — not the credit card on file.

SSNs from Financing Soft-Pulls

Wisetack, GreenSky, and Synchrony soft-pull SSN inputs live in the lender's portal. The lender's adverse-action notice template (if decline) and approval-tier output language (if approve) is what the Comfort Advisor pulls into the proposal — never the SSN. AI tools do not see SSNs. Pasting an SSN into ChatGPT — even by mistake — is an FCRA exposure event and a state attorney general's office's open file. Shops that train CSRs and advisors on the AI workflow include this as a hard line in week-one training, with a written acknowledgment in the personnel file.

Full Customer Addresses Outside the Active Call

Addresses for the dispatched call live in the FSM platform's dispatch view; AI tools that book or route receive the address through the platform integration, not through pasted context. When a CSR pastes a customer's full street address into a ChatGPT prompt to "draft a follow-up email," they have created a record of that address in OpenAI's logs that is outside the shop's data control. The fix is the customer ID — the FSM platform's internal identifier — used in AI prompts instead of the address. The AI drafts the talk-track; the platform fills in the address at send time.

Jobsite Photos with People or Kids in Them

AI photo-annotation tools (the proposal-narrative ones, the photo-tagging ones) are useful for equipment photos. They are not useful for photos that contain people, especially children. Even when the AI tool's terms allow it, the shop's discipline is to never paste a photo with a recognizable person into an AI tool. The risk is twofold: data-retention by the AI vendor (the photo lives in their logs), and downstream surface in the shop's social or marketing channels in a way the homeowner did not consent to. The discipline is: equipment-only photos for AI annotation; people-photos never. The tech crops the photo before uploading.

Customer Identifiers When the Conversation Does Not Need Them

Even when the data is not strictly PII, the shop's discipline is to minimize identifiers in AI prompts. The CSR drafting a price-shopper rebuttal does not need to paste the customer's full name; they need to paste the call situation. The Comfort Advisor drafting a financing pivot does not need to paste the customer's last name; they need to paste the financing situation and the equipment context. Pseudonymize on the way in. "Customer is a 62-year-old homeowner whose furnace died Sunday night" is richer context than "Mrs. Hendricks at 4214 Marin Park Drive" and carries zero PII exposure. The shop's voice samples and prompt templates standardize on the pseudonymized pattern.

The "Paste the Last 3 Tickets" Pattern in Practice

The single most reusable context pattern in 2026 shops is the "last three tickets" paste. It works for the CSR's relationship re-anchor on inbound calls. It works for the Comfort Advisor's pre-kitchen preparation. It works for the service manager's recall-pattern check. It works for the owner's customer-recovery email drafting. Same pattern, five roles, five outputs. Here is what it looks like operationally.

The CSR receives an inbound call from a customer. The FSM platform surfaces the customer ID. The CSR opens her prompt template and pastes the three most recent tickets in a standardized format — date, equipment touched, work performed, outcome, tech/advisor on the job, any open items. She fills in the role (CSR), task ("draft an opening line that acknowledges the relationship and pivots to today's issue"), format ("2 sentences, no preamble"), and constraint ("no first-name address, no exclamation points, do not invent ticket details I did not paste"). The AI returns: "Mr. Patel — I see we replaced your condenser fan motor in June and everything has been working well since. What's going on now?" The CSR opens the conversation knowing the customer. The customer feels known. The booking happens at 92% vs. the floor's 78%.

The Comfort Advisor uses the same paste before walking into a Bel Air kitchen. The pattern surfaces the homeowner's recent decisions ("declined the heat-pump retrofit 3 months ago"), the equipment posture, and the financing posture ("Wisetack approved last year, GreenSky inquiry never closed"). The advisor walks in with the conversational arc pre-loaded. The pivot lands faster.

The service manager uses the same paste for recall investigation. Tech had 4 of 6 weekly recalls; paste the last three tickets per recall, run a clustering prompt, surface the pattern (3 of 4 were condenser fan motors from a specific parts batch). The Monday coaching conversation is data-driven; the tech is not defensive because the parts batch is the surfaced root cause, not Carlos's technique.

Context Hygiene and the Shared Shop Discipline

Context discipline is not an individual habit. It is a shop habit, enforced by training, captured in the prompt library, and audited on the bulletin board. Shops that try to enforce it role-by-role with no shared discipline fail by week 6 — someone pastes a full SSN by mistake, the shop's PII boundary is breached, the owner discovers it on a quarterly review, and the AI program loses credibility. The fix is shop-wide standards, codified into the prompt library's templates and the verify checklist.

Three operational practices anchor the shop discipline. First, every prompt template in the library pre-specifies the context fields that should be filled — and explicitly names the fields that should not be filled even if available. The CSR's price-shopper template has a field for "customer call context (3 sentences max, no name)" — the field's hint makes the pseudonymization automatic. The Comfort Advisor's pivot template has "financing approval tier and term — no SSN, no credit-card details." The constraint is built into the structure.

Second, the 30-second verify pass from L1 adds a sixth checkpoint when context is involved: "did I paste anything I should not have?" The verify is on the input now, not just the output. A 5-second skim of the prompt before submission. If a name, address, SSN, or card number slipped in, edit before submit. The discipline is cheap; the cost of skipping it is unbounded.

Third, the shop's AI vendor list is audited quarterly for data-handling agreements. Which tools train on inputs by default. Which tools have enterprise data agreements that opt out of training. Which tools retain logs and for how long. Which tools have SOC 2 Type II reports. This is owner-level work, not CSR-level; the owner reviews the list quarterly and signs off on which tools the shop uses. Shops that skip this step end up with three CSRs using three different tools under three different data agreements, and no shop-wide answer to "where does our customer data go?" The audit is dull. The audit is mandatory.

Reading the Context Window Strategically

The skilled AI user in 2026 thinks about the context window like a dispatcher thinks about the board: limited space, high-stakes allocation, prioritized by leverage. The "lost in the middle" research from 2024-2025 showed that information at the start and end of a long context gets attended to better than information in the middle. The practical implication for shop prompts: put the role at the start, put the constraint at the end, and put the most important context (job notes, the customer's actual situation) close to the beginning of the context block. Less important context (brand voice samples, external references) can sit in the middle where attention is weaker — they shape the output but do not need to be retrieved as facts.

Short prompts also beat long prompts on accuracy. Pasting 12,000 tokens of customer history dilutes attention; pasting the last three tickets focuses it. "More context is better" is wrong past a point. Stop adding context when the answer stops improving. The prompt library codifies this — every template has a defined context shape.

Strategic reading also means picking smaller, focused windows over the largest available. Avoca's voice-agent system prompt is tight; that is a feature. Titan Intelligence's structured-data context is curated; that is a feature. ChatGPT's 128K window invites the operator to dump too much; that is a hazard. The right tool for a CSR rebuttal is the most focused window, not the biggest.

The Context Discipline as Competitive Edge

Two shops with the same AI tools, the same prompt structure, and the same five-part discipline will produce wildly different output quality if their context discipline differs. Shop A pastes pseudonymized job notes, equipment tags, last three tickets, and 2-3 brand-voice samples. Shop B pastes the entire customer file because "more is better." Shop A's output reads like a senior employee wrote it. Shop B's output reads like a confused generalist wrote it. Shop A also has zero PII exposure events; Shop B has accumulated several without realizing it. By month 6, Shop A is compounding metric movement (booking % up, close rate up, recall % down) and Shop B is doing the same volume of AI work with weaker output and growing regulatory risk.

The context discipline also shapes M&A valuation. PE acquirers (Wrench Group, Authority Brands, Apex Service Partners, Sila Services) increasingly run AI-tool data-handling diligence during platform acquisitions in 2026. Shops with documented context discipline clear diligence in 30 days; shops without it carry an unmodeled liability that affects the platform's risk underwriting.

And context discipline is the foundation for L3 and L4 workflows. Multi-step AI workflows chain context across tools — what the CSR's AI sees gets handed to the dispatcher's AI gets handed to the service manager's AI. If each handoff is hygienic, the chain works. If any handoff pastes raw PII, the chain has a leak. Shops that build the discipline at L2 graduate to L3 and L4 workflows in months rather than quarters.

Key Takeaways

• The smallest sufficient context beats the largest possible context. Job notes, equipment tags, last three tickets, 2-3 brand-voice samples — that is usually the right size. More dilutes the AI's attention; less starves it.
• Context is shared. Whatever you paste into ChatGPT, Claude, or Gemini lives in their logs under their data-handling agreement. Enterprise tools (Titan Intelligence, Avoca) inherit their vendor's agreement. Audit the data agreements quarterly; owner signs off on the vendor list.
• The "last three tickets" pattern works across five roles. CSR opens with relationship knowledge; Comfort Advisor walks in pre-loaded; service manager surfaces recall patterns; owner drafts recovery emails; marketing manager builds Hatch nurture sequences. Same pattern, different outputs.
• What never belongs in context: credit-card numbers (PCI scope), SSNs from financing soft-pulls (FCRA exposure), customer addresses outside the active call (use the customer ID), jobsite photos with people or kids in them (data-retention risk), and customer identifiers when pseudonymization works.
• Pseudonymize on the way in. "Customer is a 62-year-old homeowner whose furnace died Sunday night" is richer context than the full name and address, and carries zero PII exposure.
• The 30-second verify gains a sixth checkpoint when context is involved: "did I paste anything I should not have?" 5-second skim of the prompt before submission. Built into every template.
• Context discipline is competitive edge. Two shops with the same tools and prompt structure produce wildly different output if their context discipline differs. PE diligence, state regulatory audits, and L3/L4 workflow chaining all rely on the boundary the shop set at L2.