Mihata
AI Usage2026.05.16

Prompt Engineering Basics: ChatGPT & Claude Templates

Have you ever wondered why the same AI tool produces such different answers depending on who is asking? The difference comes not from the model's intelligence but from prompt design. In this article, drawing on the official prompting guides from OpenAI, Anthropic, and Google, we lay out a practical framework that beginner-to-intermediate business users can put to work tomorrow. We cover the five elements of role, goal, format, constraints, examples, and reasoning steps; Before/After improvements; and ready-to-paste templates organized by business function.

What Prompt Engineering Means (a Short Definition)

Prompt engineering is the discipline of designing and improving the instructions you give to a generative AI in order to draw out the result you want. Anthropic's official documentation positions it as "the practice of refining prompts on top of a foundation where you have defined success criteria and can test against them empirically." Unlike traditional business tools, what determines quality here is not a difficult programming language but the design of your written instructions—plain language, in your case, English.

Why the Basics Are Worth Mastering Now

Generative AI is a tool where vague questions yield vague answers. The opposite is also true: when you give a clear role, goal, constraints, and format, you get reliable output for drafting emails, summarizing meeting notes, building first-pass proposals, translating, or generating SQL and spreadsheet formulas. The official guides from OpenAI, Anthropic, and Google all start with the same first principle: "clear and specific instructions." Getting this one habit right meaningfully changes the consistency of your output.

Five Elements That Change the Result

Reading across the major vendor guides, five elements show up repeatedly as the most reliable levers for business use. You do not need to include all five every time; consciously adding two or three of them lifts quality noticeably.

1. Role (Persona)

Start by telling the AI who it is: "You are a ___." Google's official Workspace guide lists Persona, Task, Context, and Format as the four elements of a strong prompt, with Persona placed first. Setting a role stabilizes vocabulary, tone, and priorities.

Prompt:

"You are a customer success manager at a B2B SaaS company. Answer with a focus on preventing customer churn."

2. Goal and Output Format (Task / Format)

Spell out what the output is for and how it should be shaped. Specifying the format up front—bullet points, a table, a word limit, an email body, JSON—saves enormous editing time later. Google's official Gemini guide explicitly recommends "making the desired output format explicit."

Prompt:

"Summarize the following meeting notes into three sections: (1) decisions made, (2) action items with owner and due date, and (3) open issues for next time. Use bullet points, with no more than five items per section."

3. Constraints

Make explicit what the model should not do or what conditions it must respect. Word counts, tone, how to handle proper nouns, technical-term substitutions, and instructions like "do not invent details—write 'unknown' if the information is missing" are all useful. Negative instructions are also effective at curbing hallucinations. Google's Gemini guide recommends placing the most important constraints at the very end of the prompt.

4. Examples (Few-shot)

Show one to three samples of the input/output pairs you want. Academically this is called few-shot prompting, and it is especially effective when you want consistent tone or phrasing. OpenAI, Anthropic, and Google all converge on the same point: concrete examples convey intent better than abstract instructions.

Prompt:

"Following the examples below, classify the product review. Example 1: 'Packaging was sloppy and disappointing.' → Negative. Example 2: 'Satisfied for the price.' → Positive. Now classify: 'A bit hard to figure out how to use.' →"

5. Reasoning Steps (Chain-of-Thought-style Instructions)

For complex requests, ask the model to think through the problem in steps before producing the final answer. The academic term Chain-of-Thought originally refers to a research technique that exposes intermediate reasoning, but for business use it is easier to think of it as prescribing a procedure: "Write in this order: (1) state assumptions, (2) list options, (3) recommend." Anthropic's official guide also recommends giving Claude time to think (its "thinking" feature) when the task is complex.

Bad vs. Good: Before/After Improvements

Here is how the five elements show up in everyday requests.

Before/After 1: Customer Apology Email

Before (bad):

"Write an apology email."

After (good):

"You are the head of customer support at a B2B SaaS company. The recipient is a corporate customer (manager-level, manufacturing). Last week our outage prevented logins for 12 hours. Write an apology email in polite business English, no longer than 250 words, with a subject line. Conditions: (1) briefly cover the cause and what we are doing to prevent recurrence; (2) say compensation will be discussed separately—do not propose specific terms in the email body; (3) avoid defensive language."

Before/After 2: Meeting-Notes Summary

Before:

"Summarize this meeting note."

After:

"Summarize the meeting notes below for an internal Slack channel. Format: (1) decisions made (max 3), (2) action items with owner and due date, (3) open issues for next time (max 3). Each line should be under 80 characters. Do not infer information that is not in the notes; write 'TBD' if it is missing."

Before/After 3: First Draft of a Proposal

Before:

"Write a proposal for a new service."

After:

"You are a product manager for SaaS aimed at small and mid-sized businesses. Based on the following assumptions, draft only the 'Background,' 'Problem,' and 'Proposal Overview' sections, 200–300 words each. Assumptions: target is wholesalers with 30–100 employees; competitors are Company A and Company B; our differentiator is operational support grounded in field interviews. Do not introduce any numbers that are not given in the assumptions; mark anything that needs verification as 'requires research.'"

What these examples share is being explicit about role, format, and constraints (what not to do). The clarity of the request, not the model's raw ability, is what determines the output.

Ready-to-Use Templates by Business Task

Writing prompts from scratch every time is tedious, so here are fill-in-the-blank templates for common business tasks. Replace anything in [brackets] with your situation.

1. Email Reply Draft

Prompt:

"You work in [industry / role]. Draft a reply to the email below in polite business English, under [word count] words. Conditions: (1) put the conclusion first, (2) for any uncertain dates, write 'we will follow up with options separately,' (3) do not commit to anything that is not stated. Incoming email: '[paste the body here]'"

2. Meeting-Notes Summary

Prompt:

"Summarize the meeting notes below for internal sharing. Format: (1) decisions (max 5), (2) action items with owner and due date (max 5), (3) open issues for next time (max 3). Do not infer information not in the notes. Each line should be under 100 characters. Notes: '[paste here]'"

3. Proposal Draft

Prompt:

"You are a planner working with the [industry] sector. Based on the assumptions below, draft the five sections—Background, Problem, Proposal, Expected Effect (KPI hypothesis), and Risks—in 200–300 words each. Assumptions: [target] / [competitors] / [our strengths]. Use only numbers stated in the assumptions; mark everything else as 'requires research.'"

4. Japanese-to-English Business Translation

Prompt:

"Translate the Japanese email below into business English. Conditions: (1) the recipient is [counterpart's role / industry]; (2) use a formal register; (3) include standard opening and closing phrases; (4) prefer industry-standard English terms over forced literal translations. Source: '[paste here]'"

5. SQL or Excel Formula Generation

Prompt:

"Based on the table definition below, write a single [SQL query / Excel formula] that satisfies the requirement. Also include (1) a one- or two-line description of what it does, and (2) how it behaves with edge cases (NULL, empty strings, future-dated values). Table definition: '[columns and types]'. Requirement: '[what you want]'. Environment: [BigQuery / PostgreSQL / Excel 365 / etc.]."

6. Tone Adjustment / Rewrite

Prompt:

"Rewrite the text below for [audience] in a [tone: e.g., softer / more technical / more concise] register. Conditions: (1) keep all factual content unchanged; (2) do not modify numbers or proper nouns; (3) keep the length within ±20% of the original. Source: '[paste here]'"

If you prefer a quick reference table, the matrix below summarizes the must-have elements per task and the most common pitfalls.

Task

Must-have elements

Common pitfall

Email reply

Role, word count, register, "do not invent"

Committing to dates that were never agreed

Meeting-notes summary

Format (decisions / actions / issues), item caps, no inference

Missed items or over-summarization

Proposal draft

Section structure, word count per section, "requires research" rule

Fabricated numbers slipping in

Translation

Counterpart attributes, formality, opening/closing phrases

Over-localization that loses precision

SQL / formula

Table definition, environment, edge-case handling

Syntax errors due to dialect differences

Rewrite

Target reader, tone, "do not change facts"

Unauthorized additions or embellishment

Tendencies on ChatGPT, Claude, and Gemini

The same prompt can land differently across models, and tendencies differ. That said, models update frequently, so it is risky to make absolute statements like "this model always behaves that way." The table below distills the design philosophies visible in each vendor's official documentation as a rough guide.

Aspect

ChatGPT (OpenAI)

Claude (Anthropic)

Gemini (Google)

Core principles emphasized

Clear instructions, reference text, task decomposition, time to think, external tools, systematic testing

Clear and direct instructions, multishot examples, reasoning steps, structured input (e.g., XML tags), use of system prompts

Clear and specific instructions, explicit constraints, output format, few-shot examples, added context, prompt decomposition

Long or structured prompts

Stable when broken up with section headings or bullets

Tends to follow long instructions well when structure is signaled with headings or delimiters

Risk of important instructions being buried; place key constraints at the end

Role setting

"You are a ___" stabilizes role and tone effectively

Pairing role with explicit "success criteria" aligns evaluation

Use the four elements: Persona / Task / Context / Format

Where it tends to struggle

Will infer when up-to-date or company-specific info is not provided

Becomes cautious with vague requests, sometimes losing brevity

Tends toward concise output unless conversational tone is requested

The common takeaway: covering the five elements—role, goal, format, constraints, examples—delivers consistent quality on any of these models. When you build internal prompt standards, write for portable structure rather than tying yourself to a single model.

What Not to Do

1. Do Not Paste Confidential or Personal Data Verbatim

Customer information, unpublished financial figures, performance reviews, and raw contract text may be handled differently by each AI service depending on the plan and settings—covering training use and log retention. If your company has not yet established internal rules, simply enforcing "mask names, company names, and unique numbers before pasting" already reduces risk significantly.

2. Do Not Treat Output as Final Without Verification

Generative AI can produce plausible-sounding errors (hallucinations). For legal, financial, medical, or hiring matters, always have output reviewed against primary sources or by a qualified professional before sharing externally. On the prompt side, adding guardrails like "write 'unknown' if the information is missing" or "do not output a number whose source is unclear" makes errors easier to catch.

3. Do Not Expect a Finished Product in One Shot

For long proposals or complex analyses, do not try to produce the final version in a single prompt. Splitting the work into stages—structure first → draft each section → fact-check → tone adjustment—is faster overall and produces more consistent quality. The official guides from OpenAI, Anthropic, and Google all recommend decomposing complex tasks.

4. Do Not Lock In a Single "Magic Prompt"

Optimal prompts shift as your business and the models change. Templating prompts and improving them as a team beats a single "secret recipe" held by one person.

Mihata's Support

Mihata helps small and mid-sized businesses with the practical work of designing AI-enabled workflows, building prompt template libraries, and writing internal usage guidelines. We work alongside teams to figure out which model fits which task, how to deploy ChatGPT, Claude, or Gemini safely, and how to expand a successful pilot horizontally across the company. If your AI rollout has stalled at a single use case, we are happy to talk. You may also find these companion articles useful: the ChatGPT business-use guide and our comparison of ChatGPT, Claude, and Gemini.

Summary

  • Prompt engineering rests on five elements: role, goal/format, constraints, examples, and reasoning steps.
  • The official guides from OpenAI, Anthropic, and Google all recommend the same fundamentals: clear and specific instructions, concrete examples, and decomposing complex tasks.
  • Adding one or two lines of role, format, and constraints in the Before/After style is the simplest way to stabilize output quality.
  • Building a library of templates by task removes the bottleneck where only a few power users can get good results.
  • Treat per-model tendencies as rough guidance and design prompts for portability across models.
  • Always observe three rules: handle confidential data carefully, verify final output, and break complex requests into stages.

Copy these templates into your internal wiki or prompt library and start where the lift is highest—email replies or meeting-notes summaries. Iterate on your own house style from there. That is the highest-ROI starting point we have seen across SME engagements.

Feel free to contact us

Whether you have questions about AI, IT, or design, need a consultation,
or want to request a quote — don't hesitate to reach out.

Contact Us