Better Prompting at Work

Across the teams we work with, the gap between someone who gets useful output from a model and someone who gives up after three attempts rarely comes down to knowing special phrases. It comes down to how much information reaches the prompt and how it is organized. The magic openers circulating online, along the lines of “act as a world-class expert,” change little, because they add no knowledge about the task. Research bears this out: across 2 410 factual questions and 162 roles written into the prompt, adding the role alone did not improve answer accuracy over the version with no role at all.¹

What follows is the practice we use at RS Management daily on documents, analyses and correspondence, with concrete before and after pairs.

Five parts of a good prompt

• Role: who the model is writing as, and for whom. “A financial analyst preparing material for the board” narrows vocabulary and level of detail far more effectively than any adjective such as “professional.”
• Context: what the model needs to know so it does not have to guess. Industry, project stage, relationship history, constraints, decisions already made. Include only what the task needs: a 2023 study showed that models are easily distracted by irrelevant context and make more mistakes on problems padded with unnecessary detail (Shi et al., ICML 2023).
• Task: one specific action expressed as a verb. “List the risks” and “judge which risk is largest” are two different tasks and come out better split into two steps.
• Output format: length, structure, tone, language, table layout. Without it the model picks a format on its own, usually one that misses the need.
• Example: one or two samples showing what a good answer looks like.

The order matters less than whether all of them are there. When an answer disappoints, one is usually missing, most often the output format or the context.

Three pairs from everyday work

The meeting note

Before:

Summarize these meeting notes.

After:

You are assisting a project manager. Below are raw notes from a 90-minute status meeting with the client and two vendors. Write a summary for a director who was not present and has two minutes to read it. Format: three sentences of substance, then a list of decisions taken, then a table of actions laid out as action / owner / due date. If an owner or due date was not stated explicitly, write “not agreed” rather than filling it in with a guess.

That last sentence removes the most common defect in such summaries: politely inventing deadlines nobody agreed to.

The reply to a price increase

Before:

Write an email to the vendor saying we do not accept the increase.

After:

Context: three years of work with a service vendor, a framework agreement with an annual addendum, the vendor has announced a 12% rate increase from January, justified by labour costs. We want to keep the relationship and to spread the increase over two stages. Write a reply of up to 150 words, matter-of-fact and calm, no exclamation marks, closing with one question about their willingness to discuss the schedule. Below are two of our earlier emails to this vendor, keep to that tone.

The two pasted emails do more work here than any description of style.

The sales spreadsheet

Before:

Analyse this sales data.

After:

Below is monthly revenue for 14 product categories over the last 18 months. Task: identify the categories where the year-on-year trend changed direction in the most recent quarter. For each, give the size of the change and the month of the reversal. Format: a table with columns category / year-on-year change / month of reversal / short comment. Do not interpret causes. If the data for a category is incomplete, list it separately below the table.

“Do not interpret causes” earns its place: asked for open-ended analysis, a model will happily supply explanations that cannot be derived from the numbers in front of it.

Ask for a plan before the work

For any task longer than a page, it is worth asking for a plan first: 5 to 7 bullets stating what each section will cover and where the data comes from. Correcting a plan takes 2 minutes. Correcting a finished eight-page document takes 30 minutes. A plan also surfaces misunderstandings early, when the model has read the task differently than it was meant. This single habit changes more than the rest of the list combined.

Iterate on the model’s own mistakes

The usual reflex after a weak answer is to fix it by hand. It pays more to treat each defect as a missing instruction and add it to the prompt. The model produced generalities instead of numbers, so the prompt gains a line requiring numeric values. It wrote three pages instead of one, so a hard word limit appears. After a few rounds the prompt produces a predictable result on the first attempt.

Proven blocks are worth keeping. Ours live in one text file sorted by category: summaries, correspondence, data analysis, document review. Copying a block and swapping in fresh context takes under 1 minute, and the quality stays consistent.

Verify before anything goes out

Model output needs a check before it travels further: figures, dates, names and quotations from source documents. It helps to ask the model to point to the passage each claim came from, then verify two or three at random. If those spot checks stay clean for 4 weeks, they can be thinned out. If they fail, the problem usually sits in the prompt that let the model step outside the source material.

A separate question is what goes into the chat window at all. Personal data within the meaning of GDPR, commercial terms and confidential material call for a tool the organization has approved, and a clear understanding of where that data is processed and how long it is retained. For sensitive data the decision is worth confirming with the people responsible for security and compliance.

What a better prompt will not fix

Prompt structure helps only when the task can be done with the material provided. A model will not invent data it never received. It will not answer a badly framed question well, including one the author cannot answer. And it does not substitute for system access: when the information sits in a database, an HR system or a document repository, the answer is integration, for example through MCP (Model Context Protocol), a standard for connecting models to company data sources, rather than an ever longer prompt.

The practical starting point is simple: pick three tasks repeated every week, write a five-part prompt for each, test them over a few days and revise after each disappointing result. After a month, the difference in time and quality is visible without any measurement.

Footnotes

1. Zheng et al., measurement of how a role written into the prompt affects answer accuracy: https://arxiv.org/abs/2311.10054. ↩