Beyond the Buzzwords
Supplemental scenarios for "A Practical Guide to AI in Specialty Coffee" — published in Roast Magazine by RoastLog.
Each scenario below walks through a real challenge, the exact prompt used, what the model produced, and how a follow-up refinement improved the output. These are working examples you can adapt to your own operation.
The same task. Two prompts. Completely different results.
The article walks through a real example: writing a wholesale proposal for a 12-property hotel chain on behalf of Zambrano Coffee. The first attempt used a vague, one-sentence prompt. The second gave Claude everything it needed. The difference in output quality is night and day — and the second prompt took about two minutes to write.
Result: Claude came back with clarifying questions. No proposal was generated. It didn't know what type of business, what product, or what the pricing looked like — so it asked.
Result: Claude produced a complete, multi-page partnership proposal — with tiered pricing tables, financial impact calculations, quarterly review structure, and a decision timeline — without asking a single follow-up question. It even chose to format the output as a Word document on its own.
See the actual output Claude created.
This is a real proposal, generated in a single pass from the prompt above, using Claude Opus 4.5. It's the kind of document that would normally take a few hours to draft from scratch. Claude produced it in under a minute.
From Cupping Notes to Customer Copy
The Challenge
Alma Coffee Roasters just cupped their new seasonal lineup — four coffees, a page of internal notes dense with scoring shorthand and tasting jargon: "enzymatic sweetness, low acidity/malic, phosphoric brightness, dense body." The marketing manager needs retail bag copy, a wholesale sell sheet paragraph, and a short "What's New" email blurb for each coffee — each written for a different audience, all in the same brand voice.
The Input Materials
Internal cupping notes for four coffees (Ethiopia Guji Natural, Colombia Huila Washed, Guatemala Antigua, Kenya AB Nyeri) including origin details, processing method, altitude, and scorer's comments. Also included: three existing retail bag descriptions used as voice and style reference for the model.
The Prompt
"I've attached my internal cupping notes for four new coffees and three examples of our existing retail bag descriptions for voice and style reference. For each coffee, write: (1) a retail bag description of 40–60 words, (2) a wholesale sell sheet paragraph that leads with altitude, varietal, and process, and (3) a short email blurb of no more than three sentences for our "What's New" newsletter. Keep the tone consistent with the reference examples — direct and specific, not flowery. Where the wholesale copy is relevant, include the cupping score."
The Output
Claude produced four complete sets of copy — retail bag descriptions, wholesale paragraphs, and email blurbs — each calibrated to its audience. The retail copy matched the voice of the reference examples closely: short, specific, no adjective inflation. The wholesale copy led with altitude and process as requested and incorporated the cupping scores. The email blurbs were punchy three-sentence blocks. One notable choice: Claude used "phosphoric brightness" verbatim in the Kenya wholesale copy, having correctly inferred that wholesale buyers would understand the term — while the retail bag copy for the same coffee said "sparkling finish" instead.
The Refinement
"The wholesale copy looks good but the email blurbs are still a bit long — cut them to two sentences maximum. Also, the wholesale paragraph for the Colombia Huila doesn't mention the varietal prominently enough; the Castillo and Colombia varietals are a selling point for buyers who care about traceability. Revise that one to lead with varietal and altitude in the first sentence."
Claude tightened all four email blurbs to two sentences without losing the key selling point of each coffee — in every case the first sentence carried the flavor profile and the second was a clean call to action. The Colombia Huila wholesale paragraph was rewritten to open with "1,700–1,900 MASL | Washed | Castillo, Colombia" as a header line, then moved into cup profile and positioning. The revision took seconds and produced copy that was immediately usable without further editing.
Key Takeaway
Providing reference examples teaches the model your voice faster than describing it. Show don't tell: three bag descriptions are worth more than a paragraph explaining your brand tone. The model matched register and length from the examples, then scaled that voice across four different coffees and three different audience formats.
Building Staff Training Materials
The Challenge
Threadline Coffee is rolling out batch brew across all three of their Denver locations. Their baristas know pour-over and espresso inside and out, but nobody has worked with the new Fetco CBS-2152 before. Training is Monday — four days away — and the operations manager doesn't have time to write a full training program from scratch.
The Input Materials
Internal reference document covering equipment specs (Fetco CBS-2152), Threadline's specific brewing parameters (150 TDS water, 200°F, 1:16.5 ratio, 3.5 oz dose, 45-minute hold limit), identified staff knowledge gaps, and a short excerpt from existing training materials showing the preferred voice and tone.
The Prompt
"I've uploaded our internal batch brew reference document with our Fetco CBS-2152 specs, our specific water and brewing parameters, and some notes on what our baristas already know and what they don't. Create a complete training guide with four sections: (1) Extraction principles the baristas need to understand — written in our training voice, not a textbook (2) A numbered step-by-step brewing SOP (3) A troubleshooting table covering the most common batch brew problems with causes and fixes (4) A cleaning and maintenance schedule Also generate a lamination-ready recipe card that fits a half-sheet format, and a 10-question multiple-choice quiz with an answer key."
The Output
Claude produced a complete training package: a four-section guide with extraction theory written in a direct, collegial voice matching the tone of Threadline's existing materials; a 10-step SOP with specific parameters embedded at each relevant step; a troubleshooting table with 11 common failure modes, causes, and fixes; and a cleaning schedule broken into between-batches, end-of-day, weekly, and monthly intervals. The recipe card was formatted as a bordered text block with all parameters in a scannable two-column layout, ready to copy into a design tool. The quiz used all multiple-choice format as specified, with an answer key at the bottom. The model correctly embedded Threadline's specific numbers (150 TDS, 200°F, 99g dose) throughout every section rather than using generic values.
The Refinement
"The troubleshooting table is good but it's missing one important scenario: what baristas should do when the TDS reading is dramatically outside range (not just a little off) and they're not sure if the equipment is malfunctioning. Add that row. Also, the quiz question about water TDS is too easy — make it a scenario question instead: "You forgot to check the water filter and the TDS is reading 220 ppm this morning. What do you do?""
Claude added a troubleshooting row for "TDS dramatically out of range (below 1.10% or above 1.65%)" with a two-step response: do not serve, then escalate to the operations manager with a note in the equipment log — exactly the kind of judgment call that distinguishes a good SOP from a basic one. The revised quiz question became a scenario: baristas had to choose between four plausible responses, with the correct answer being to replace the water filter cartridge and run a purge cycle before brewing. The distractor answers ("brew anyway and adjust the grind," "add distilled water to dilute") were realistic enough to test genuine understanding.
Key Takeaway
Specific parameters are better anchors than general instructions. When you give the model your actual numbers — 150 TDS, 200°F, 3.5 oz, 45-minute hold — every section of the output is calibrated to your operation, not a generic coffee training template. The more precise your inputs, the less editing the output needs.
Preparing for a Wholesale Account Review
The Challenge
Campfire Roasting Co.'s quarterly review with their largest wholesale account — the Ridgeline Hotel Group — is tomorrow. The sales manager has 12 months of order history in a spreadsheet and two email threads about a quality issue with their Ethiopia Sidama in September. She needs to walk into the meeting with a clear picture of trends, a plan for addressing the quality concern directly, and a growth proposal for the next quarter.
The Input Materials
A 12-month order history spreadsheet (29 line items across three SKUs: House Blend, Ethiopia Yirgacheffe, Ethiopia Sidama) with monthly unit volumes and revenue. Two email threads from October 2024: the initial quality complaint from the hotel's F&B Director, and Campfire's response and resolution. Also included: a note that the account is opening a fourth hotel property in Boulder in Q1 2025.
The Prompt
"I've uploaded a spreadsheet with 12 months of order history for our largest wholesale account and two email threads about a quality issue with our Ethiopia Sidama in September. Analyze the order data for seasonal patterns, SKU performance trends, and reorder frequency. Then draft talking points that address the quality concern directly and honestly — don't soften it — and propose a growth plan for next quarter that accounts for their new Boulder property opening in Q1 2025. I want to walk into this meeting with a clear picture of the full relationship, not just the recent problem."
The Output
Claude produced a structured review document in four sections. The data analysis identified three clear patterns: a December holiday spike in House Blend (nearly double the typical monthly volume), a gradual Yirgacheffe decline that correlated directly with the Sidama introduction in May, and a September Sidama drop that was clearly event-driven rather than a trend. The talking points section addressed the quality issue directly — naming the specific cause (a high-moisture second harvest lot), the immediate resolution steps taken, and the process changes implemented afterward. The growth plan included a monthly Q1 2025 order calendar with volume projections for the Boulder property, and a recommendation to formally retire the Yirgacheffe by Q4 rather than let it decline to zero. Claude also generated a one-page leave-behind executive summary with a clear action item checklist for the meeting.
The Refinement
"The quality issue talking points are good but they read a little formal for what is actually a strong relationship — the emails show Sandra responded really positively when we handled it fast. Soften the tone slightly so it sounds like a conversation between partners rather than a corporate accountability statement. Also, add a specific recommended order date for the December holiday surge based on the pattern in the data."
The revised talking points opened with an acknowledgment of the positive resolution rather than the problem itself — framing it as "we know you handled this well and we want to close the loop" rather than "here is what went wrong." The December surge recommendation became a concrete date: "Submit your December holiday order by November 15 — your December 2023 data shows a 40–80% volume increase over baseline, and we want to make sure you're fully stocked before the holiday window opens." The action item checklist on the leave-behind was updated to include that order date as a named deliverable.
Key Takeaway
Real data produces specific analysis. When you upload actual order history instead of describing it in prose, the model can surface patterns you might not have noticed — like the Yirgacheffe decline tracking precisely against the Sidama introduction, or the December surge being almost exactly predictable from the prior year. The output is only as useful as what you feed it.
Turning an Origin Trip into a Content Strategy
The Challenge
Silo Coffee just returned from a two-week origin trip to Oromia, Ethiopia — 10 days, 340 photos, raw trip notes, three new coffees to launch. The founder wants to build a content campaign around the trip but doesn't want to post everything at once or produce the kind of generic "we visited a farm" content that's indistinguishable from every other specialty coffee brand. The coffees launch in four weeks.
The Input Materials
Raw trip notes from 10 days in Oromia covering key people met (farmer Berhane and his daughter Selamawit, exporter Frehiwot), specific places visited (Benti Nenka washing station, Aricha station), memorable observations, and a photo inventory of 60 edited selects. Also included: a brand voice guide excerpt explaining Silo's tone (honest, specific, story-driven, not precious) and their existing social media channels.
The Prompt
"I've uploaded my raw trip notes from our Ethiopia origin trip, a summary of our best photos, and our brand voice guide. We have four weeks before the coffees launch. Create a 4-week content calendar with specific posts for Instagram and LinkedIn, plus two email newsletter drafts — one for Week 3 (people and story) and one for Week 4 (coffee launch). For each Instagram post, write the full caption in our voice. The strategy should build audience interest gradually — don't lead with the product. Use the specific people and details from the trip notes, not generic coffee-origin language. Also provide a hashtag strategy."
The Output
Claude produced a 4-week calendar with a clear narrative arc: Week 1 builds curiosity through landscape and process imagery, Week 2 teaches the audience something real about how the coffee is made, Week 3 focuses on the people (Berhane, Selamawit, the coffee ceremony), and Week 4 launches the product to an audience that's already been on the trip. All 11 captions were written in Silo's voice — specific and grounded, no superlatives, using real names and altitudes from the trip notes. The two email newsletters were distinct: the Week 3 email was a first-person letter from the founder about Berhane and Selamawit, and the Week 4 email was a clean product launch with three coffees presented individually. The hashtag strategy recommended 4 primary tags, 6 trip-specific tags to rotate, and 2 brand tags — and explicitly advised against hashtag dumps on LinkedIn.
The Refinement
"The Instagram captions are good but a few of them are too long for the medium — Instagram works better when captions can be read in under 20 seconds. Trim the three longest captions to 100 words or fewer. Also, the Week 3 email is good but the P.S. line feels a little forced — either make it more natural or cut it. And can you add a LinkedIn post for Week 2 that specifically addresses the women-owned cooperative angle from the trip notes? That's an audience that resonates on LinkedIn."
The three longest Instagram captions (Posts 4, 7, and 10) were each trimmed to under 100 words while keeping the most specific detail from each — the fermentation smell in Post 4, Selamawit's agronomy ambitions in Post 7, and the cupping table description in Post 10. The P.S. in the Week 3 email was rewritten to feel like something the founder would actually add as an afterthought rather than a planted promotional line. The new LinkedIn post for Week 2 focused specifically on the three women-owned cooperatives and Frehiwot's role in the supply chain — framed as a perspective on what "direct trade" actually means in practice, which performed well with the professional audience the post was targeting.
Key Takeaway
Context and constraints are the model's raw material. The more specific your inputs — real names, real places, real observations — the less the output sounds like generic content marketing. Defining tone, platform, and narrative structure upfront means the model isn't guessing what kind of content you want. Describing your brand voice in a reference document and then showing specific examples of it produces better results than describing the voice abstractly.