Claude Haiku 4.5 Prompt Cards for a 40-Person Support Team

01 · Context First: The Rule That Fixed Half Our First Drafts

When our support and operations group started using language models daily, the first week produced a predictable mess. Analysts pasted a single sentence into a chat window, got a plausible answer, pasted that answer into a ticket, and moved on. By Friday, three people had published three different explanations of the same billing policy. The model was not the problem. The prompt was.

The fix was a rule we teach on day one: state the context before you state the task. Who is asking. What the output will be used for. What constraints matter. What counts as done. Non-technical staff pick this up quickly, because it mirrors how they already brief a new colleague. A prompt that says we handle enterprise billing disputes, the reader is a customer worried about a double charge, the reply must be under 150 words, and it must not promise a refund date produces far fewer rewrites than one that says write a reply about billing.

Write the rule on the wall. Put it in the template. Correct it in review until it becomes automatic. Most of the variance in output quality across a large team traces back to this single habit.

02 · The Four-Part Prompt Card for Analysts and Coordinators

We hand every new user a prompt card with four labelled parts: Role and audience, Task, Inputs, and Shape of the answer. Staff fill in the four blanks, then send. The card works because it externalises the thinking that experienced prompt writers do silently.

• › Role and audience: who the model is writing as, and who will read the result. A customer-facing note and an internal escalation brief need different voices.
• › Task: one verb. Summarise, classify, extract, draft, check, rank. Two verbs means two prompts.
• › Inputs: paste the source material, name the fields, say what is missing. If the input is a messy ticket thread, say that it is messy and tell the model what to ignore.
• › Shape of the answer: length, structure, tone, and the one thing that must appear or must not appear. Bullets or prose. One paragraph or five. No speculation.

The shape clause is the cheapest quality lever a team has. People who ask for a table get a table. People who ask for nothing get whatever the model finds convenient, and then spend twenty minutes editing it.

03 · Two Worked Examples That Train Faster Than Advice

Training non-technical staff works best when you show a before and after from their own domain. Take a vague prompt: summarise this thread. Now take the same request with the card filled in: summarise the thread below for a shift handover, list open issues as bullets, name the owner of each, flag anything time-sensitive in bold, keep it under 120 words, and omit greeting and sign-off text. Same model, same input, dramatically less editing.

Run the pair through a fast, economical model such as Claude Haiku 4.5 so the lesson costs almost nothing to repeat. Then show the same prompt handled by a heavier reasoning model such as Claude Sonnet 5.5 when the task is a policy comparison or a multi-step deduction, so staff learn that model choice and prompt quality are separate decisions.

Keep a shared library of these pairs. Twenty good examples from real work teach more than any style guide, and new hires read them on their first morning.

04 · Every Bad Prompt Is Also a Metered Token Expense

On a pay-as-you-go relay, prompt discipline has a financial dimension that technical teams feel first and non-technical teams learn later. Every call is metered on real tokens, and long, rambling prompts with pasted context that is not needed inflate both input and output. Over hundreds of daily calls across a whole department, that is a visible line in the budget.

Teach three cost habits alongside the writing habits. First, send only the input that the task requires; a whole ticket history is rarely needed when three quoted lines will do. Second, cap the output in the prompt, because an uncapped answer is a metered answer. Third, ask for exactly one task per call and chain calls when necessary, rather than issuing a single mega-prompt that returns a long document nobody reads.

Teams running heavy, sustained workloads should also think about repeated context. If your staff re-sends the same policy handbook in every prompt, that repeated block is re-billed on every call unless the platform handles it. Quanta is built for exactly this pattern: one balance covers Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and other models, it works with Claude Code, Codex, Gemini CLI and any OpenAI SDK, and cache writes and cache reads are priced separately so repeated context gets cheaper instead of silently compounding. For a rollout where dozens of people share one budget, that visibility is the difference between a controlled pilot and a surprise.

05 · Review Prompts Like You Review Tickets, on a Rolling Cadence

Writing improves under feedback, and prompt writing is no different. Put a light review loop in place from the first week of rollout. Each team lead spends thirty minutes a fortnight sampling prompts and their outputs, then posts two findings: one pattern to adopt, one anti-pattern to drop. Keep the sample small and the cadence steady. This is coaching, not audit.

Give the team simple acceptance checks they can run themselves before sending an answer onward. Did the output answer the actual task. Is every claim supported by the supplied input. Does the format match the shape clause. Would a colleague need to ask a follow-up question to use it. If any check fails, the fix is almost always in the prompt, not the model.

As confidence grows, let experienced users tune the prompt library and test small variations against the same input. Model choice belongs in the same conversation: triage and drafting can stay on an economical model, while complex synthesis can move up when the prompt is already clean. Refinement is cheaper when the prompt is doing its job.

06 · Rolling Out to Forty People Without Forty Different Standards

Standardise the container, not the voice. Give everyone the same four-part card, the same shared library of worked pairs, the same review cadence, and the same naming convention for saved prompts, then let departments adapt the wording to their own work. A rollout that mandates sentences produces resentment; a rollout that mandates structure produces consistency.

Set a budget envelope per team and share the token spend numbers openly each week. When people see that a clear prompt with a capped output is cheaper than a vague one, the behaviour spreads without enforcement. The referral programme helps here too: colleagues who bring new users into the workspace earn 19% of their top-ups, which is a practical way to fund a growing rollout.

Start with one team, run it for two weeks, and expand only when the prompt library holds up under real volume. If you would like to run this rollout on metered, multi-model infrastructure, open a Quanta account and let your first cohort write the next twenty examples themselves.