Level 100: How AI works: AI at work and in life · Module 100.5

Two coworkers write together in a notebook beside a closed laptop.
Ask for a draft.

Talking to AI well · Standard gear

What you will be able to do

The course request checklist: a job, the facts, an example, rules, and something checkable.
The course combines advice on stating the job and supplying facts and examples (s1, s5), setting rules and allowing uncertainty (s5, s6), and asking for a checkable answer (s6). It is not a tested formula.
Figure data as text

A request you can check

This is the course’s checklist, not a formula tested as a whole.

PartWhat you writeLesson sources
The jobSay what you want and who it’s for.s1, s5
The factsGive the information it needs for this job.s1, s5
An exampleShow what a good answer looks like.s1, s5
The rulesSet limits and say what to do if unsure.s5, s6
Something checkableAsk for quotes or figures you can verify.s6

What you type into an AI tool is called a prompt. This lesson makes two points. What you put in the prompt changes what you get back. And the first answer does not have to be the last one.

The evidence comes from three research papers, one controlled experiment and one AI company's guide to its own product. They are not equally strong, so we say which is which as we go.

One thing does not change, however well you ask. The AI drafts; you decide. Nothing in this lesson lets an AI send, sign or agree to anything for you.

The wrong idea: short commands work best

Many people type a few words, such as "write an email to a supplier", and take what comes back. The tool has to guess everything else: who the email is for, what it must say, and how long it should be.

Researchers studied 10,538 prompts that people had shared on prompt-sharing websites and chat servers between December 2022 and August 2023.1 They found that these real-world prompts were built from several parts.1 The part that appeared in almost every prompt was what the researchers called the Requirement: the background, the description of the job and the instructions to follow.1 That much describes what people wrote. Separately, the researchers ran a smaller controlled test with GPT-4 on two tasks: writing for search engines, and writing prompts for an image generator. When they removed parts of those prompts one at a time, removing the Requirement had the biggest effect on the quality score.1 Two tasks and one model are a narrow test, not proof for every kind of request.

Anthropic, the company that makes the Claude models, gives similar advice in its guide for its own product (as of November 2026). It says to treat the model like "a brilliant but new employee who lacks context on your norms and workflows", and that "the more precisely you explain what you want, the better the result."2

A request in five parts

This course teaches a request in five parts. The list is the course's own checklist, not a formula from any one study. Each part is backed by the evidence below.

PartWhat you writeExample words
1. The jobWhat you want, and who it is for"Write a short email to a parts supplier we have used for five years."
2. The factsThe information it needs, so it does not guess"Ask to move delivery from 3 March to 17 March. The price stays the same."
3. An exampleWhat a good answer looks like"Here is an email we sent last year that worked well: ..."
4. The rulesLength, tone, what to avoid, and what to do if unsure"Under 120 words. Do not promise anything else. If you need a fact I did not give you, ask me."
5. Something checkableAn answer in a form you can check"Then list every date and number you used, so I can check them against my notes."

1 and 2: the job and the facts

The paper that introduced GPT-3 found that model performance improved when a plain-language description of the task was added to the input.3 Anthropic's guide says that explaining the reason behind an instruction gives the model "more information about your goals" and helps it give more targeted answers.2 It suggests a test: show your prompt to a colleague who knows little about the task. If they would be confused, the model will be too.2

The facts matter because the tool only has what you give it, plus what it learned in training. If you leave out the date, the price or the customer's name, it has to guess, or fill the gap with something that sounds right.

3: an example

The GPT-3 paper tested three ways of asking: with no examples ("zero-shot"), with one example ("one-shot") and with a few examples ("few-shot").3 Performance improved as more examples were added, but the paper also says that more examples are "usually but not always better".3

In the study of shared prompts, examples were tested in only one of the two tasks, the image-prompt task, because the search-engine prompts had none. There, removing the examples lowered GPT-4's score from 0.36 to 0.30, a drop of about 17%.1

What does that score measure? In the image-prompt task, GPT-4 wrote prompts for making pictures. The researchers turned each prompt into a picture with an image generator called Stable Diffusion. Then a second AI model, called CLIP, scored how closely the picture matched the words of the prompt, on a scale from 0 to 1.1 So the 17% is a drop in a word-and-picture match score, judged by a model. It is not a measure of truth, and no person rated whether the answers were useful at work.

Anthropic's guide calls examples "one of the most reliable ways" to steer the format, tone and structure of an answer, and says the examples should closely match your real task.2

4: the rules

Rules tell the tool the length, the tone, and what it must not do. Anthropic's guide says that "being specific about the output you want can help enhance results."2

One rule is worth adding to almost every request: tell the tool it may say it does not know. Anthropic's guide on reducing made-up answers says to give the model explicit permission to admit uncertainty, and claims this "can drastically reduce false information".4 That is the company's claim about its own product, not an independent measurement. "I don't know" is an honest answer, from the tool or from you. Treat it as a next step: find the fact, or ask the person who has it.

A rule in a prompt is a request to the tool, not a lock. "Do not promise anything else" makes a stray promise less likely. This course does not treat it as stopping one. You still read the draft, and you decide whether it goes out.

Not every popular trick earns its place. Many shared prompts begin with a role, such as "act as an expert in search engine optimisation".1 In the researchers' two tests with GPT-4, adding a role made only a marginal difference to the answer.1 The job and the facts did more.

5: something checkable

Ask for the answer in a form you can check: the dates and numbers it used, its steps, or the exact words it relied on in a document you gave it. Anthropic's guide suggests making an answer "auditable" by asking the model to cite quotes and sources for each of its claims.4 The same guide warns that its techniques reduce made-up answers but "don't eliminate them entirely", and says to always check critical information.4 A list of sources is something to check, not proof. (Lesson 100.8 covers checking.)

The checkable list is a work note, not a receipt. If your workplace allows it, keep your prompt and the AI's answer in your own work record, kept apart from any receipt. Do not put them in a receipt. In the AiGovOps Foundation's practice, a receipt carries "hashes, decision, tier, policy hash, key id"; it never carries "prompts, outputs, payloads, personal data".5

That does not mean nobody may ever see your notes. This course's rule: if a reviewer or checker needs the prompt or the answer itself, they see it only through a route your workplace has approved. Holding a receipt, or a code taken from it, does not give anyone permission to read the notes behind it. Lesson 100.P covers receipts.

Real case: a clear, checkable job, done faster

In a 2022 experiment, posted as a paper in February 2023, researchers recruited 95 professional programmers and randomly split them into two groups.6 Both groups had the same clearly defined job: build a web server (an "HTTP server") in the JavaScript language, as quickly as possible.6 The job came with twelve automatic tests, and a programmer finished when their work passed all twelve.6 Among the programmers who finished, those with access to GitHub Copilot, an AI coding assistant, took 55.8% less time on average than those without it.6 That is a cut in the time taken on one task, not a 55.8% rise in overall output. Whether a programmer finished at all was a separate result: the Copilot group's success rate was higher, but the difference was not statistically significant.6

This study did not test prompts. It is in this lesson for a different reason: it shows what a clear job with a built-in check looks like. Everyone knew exactly what "done" meant, and the tests told them whether the work was right.

Keep the result in its place. It is one task in one language. The authors say the benefits may vary across tasks and languages, and that they did not study the quality of the code.6 The paper was posted on arXiv as a preprint in February 2023, and its arXiv record, as read in November 2026, lists no journal publication.6 The authors worked at GitHub, the maker of Copilot, and at Microsoft Research and MIT.6

The second wrong idea: the first answer is the best answer

The Self-Refine paper starts from a simple observation: "Like humans, large language models (LLMs) do not always generate the best output on their first try."7 Its authors had a model write an answer, then write feedback on its own answer, then rewrite the answer using that feedback, and repeat.7 Across seven tasks, from replying in a conversation to maths reasoning, people and automatic measures preferred the refined answers. On average, task performance improved by about 20 points (an absolute gain) over the first answers.7

In that study the model gave feedback to itself. At work, you can give the feedback. The study did not test people writing feedback, so using its findings for your own follow-ups is this course's suggestion, not a measured result.

Specific feedback works better than vague feedback. The authors compared specific feedback (for example, "avoid repeated calculations in the for loop") with generic feedback ("improve the efficiency of the code"). Specific feedback gave the best results.7 On one task, called sentiment reversal, the score was 43.2 with specific feedback, 31.2 with generic feedback and 0 with none.7

A few rounds help, then the gains shrink. The improvement grew smaller with each extra round, and it did not always go up every time.7

A model is not a fact-checker for itself. On maths problems, the gains were small. The authors traced this to the model not being able to tell whether its answer had a mistake. ChatGPT's feedback on its own maths answers was "everything looks good" 94% of the time.7 So asking "are you sure?" is not a check. Checking against a real source is.

Good follow-up questions

These follow-ups are the course's suggestions, built on the findings above.

At Big Widget Company (a made-up example)

Big Widget Company and the people in it are invented for this course. The studies above are real.

Priya, a buyer in procurement, needs to ask a supplier for a later delivery date.

The AI drafted. Priya decided. She is the gate, and the "Do not promise anything else" line in her prompt is not. Suppose the supplier replies asking for a higher price. Changing the price is not Priya's call, so she does not ask the AI to agree to it. She holds the reply and takes it to her manager, who owns supplier terms.

Three-yes check (the prompting questions only; lesson 100.9 covers Get to Yes, Stay at Yes and Recover to Yes in full).

Summary

Words to know: prompt, context, example (few-shot), follow-up, iterate.

Sources

  1. s2 Yihan Ma, Xinyue Shen, Yixin Wu, Boyang Zhang, Michael Backes, and Yang Zhang (2024). The Death and Life of Great Prompts: Analyzing the Evolution of LLM Prompts from the Structural Perspective. Paper. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  2. s5 Anthropic (2026). Prompting best practices (Claude documentation). Official doc. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩ ↩
  3. s1 Tom B. Brown et al. (2020). Language Models are Few-Shot Learners. Paper. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩
  4. s6 Anthropic (2026). Reduce hallucinations (Claude documentation). Official doc. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩ ↩ ↩ ↩
  5. s7 AiGovOps Foundation (2026). The AiGovOps Foundation café — start with your problem (Practice home page; object table, Receipt row). Official doc. Checked 2026-11-17 by source-checker round-3 re-check (separate Claude Code session 1b682d3d). ↩ ↩
  6. s4 Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. Paper. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  7. s3 Aman Madaan et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. Paper. Checked 2026-11-17 by source-checker round-2 re-check (separate Claude Code session 9d48e5af). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩

Next: the quiz