Lesson 1.4: Clear instructions: vague against sharp
Learn to turn vague requests into instructions a model can follow and you can check, and why too many rules lower the chance that every one is followed.
By the end of this lesson you will be able to rewrite a vague instruction into a sharp one, test whether an instruction can be checked, spot conflicts and overload in a long prompt, and use the model itself to find the gaps in your request before it answers. You will climb a four-step ladder from a vague wish to a testable rule, see with numbers why ten rules are rarely all followed, and practise on three realistic rewrites from operations, finance and people work. You should have read Lessons 1.1 to 1.3, since this lesson sharpens the parts of the prompt you already know how to assemble. No technical background is needed, and every exercise can be done with any assistant.
An instruction is the smallest working unit of a prompt: one sentence that says what to do or what not to do. Most of the quality of a prompt comes from whether each of those sentences is clear enough to follow and clear enough to check. Many disappointing replies are not model failures. They are the faithful execution of an unclear instruction.
Why instructions fail
When a model does not do what you asked, one of five things has usually gone wrong.
Cause · What it looks like · ExampleAmbiguity · The words allow several meanings · "Make it concise"
Missing definition · A term means something specific to you · "Summarise the active customers"
Conflict · Two instructions pull in opposite directions · "Be brief" and "cover every detail"
Overload · So many rules that some are lost · A prompt with twenty rules
Burial · The key rule sits in the middle of a long text · A "never state prices" line in paragraph six
The fixes are the subject of the rest of the lesson. The first, and the one that does most of the work, is to make each instruction sharp.
The specificity ladder
Instructions range from vague to testable. It helps to think of four rungs, and to climb at least to the third for anything that matters.
1. A wish. "Make it better." There is nothing to act on.
2. A direction. "Make it shorter." The model knows which way to go, but not how far.
3. A measurable target. "Cut it to under 100 words." There is now a number.
4. A testable rule with a reason. "Cut it to under 100 words, because it goes in a text message, and keep the date and the amount." Anyone can check it, and the reason helps the model make sensible trade-offs.
Here is the same climb for another instruction, about tone.
Rung · InstructionWish · "Make it sound professional."
Direction · "Make it more formal."
Measurable · "Use no slang or contractions, and address him as Mr Rao."
Testable with reason · "Use no slang or contractions and address him as Mr Rao, because he is a senior client we have not met."
Predict first: a manager writes "Keep the report short and focused." Is this a wish, a direction or a measurable target? How would two different assistants, or two different colleagues, interpret it? Decide before reading on.
It is a direction at best. "Short" could mean one page or ten, and "focused" could mean any of a dozen topics. Two colleagues would produce two quite different reports. A measurable version is "Keep the report to one page, covering only revenue, costs and the three biggest risks."
The checkability test
A reliable test of sharpness is: could someone who did not write the prompt check whether the instruction was followed? For "be concise" the answer is no. For "under 100 words" the answer is yes, in seconds. Instructions that fail the test are the ones that produce arguments later, because nobody can agree whether they were met.
Instruction · Checkable? · A sharper version"Use a friendly tone" · Hardly · "Use first names, short sentences and one light sentence, such as a thank-you at the end"
"Don't be too technical" · No · "Avoid the words 'API', 'latency' and 'schema', and explain any other technical term in brackets"
"Include the key points" · No · "Include the decision, the owner, the deadline and the cost"
"Under 80 words" · Yes · Already sharp
"Quote the sentence that supports each claim" · Yes · Already sharp
You do not need numbers for every instruction. A style such as "friendly" can legitimately stay soft. But the more an outcome matters, the more you should push it to a rule someone could check.
Define your terms
Every workplace has words with specific meanings that no outsider, and no model, can guess. "Active customer" might mean someone who logged in this month, or someone who transacted in the last 90 days. "Overdue" might mean one day late or past a grace period. "Net revenue" may or may not include refunds. If such a term matters to the task, define it in the prompt, in one line.
text
Definitions:
- Active customer: made at least one transaction in the last 90 days.
- Overdue: more than 7 days past the due date.
Use these definitions exactly. If the data does not let you apply them,
say so instead of guessing.
Suppose you ask for "the share of active customers who are overdue" with no definitions. The model will choose its own. One plausible choice gives 4 per cent and another gives 11 per cent, and neither is wrong, but only one matches your report. The definition costs thirty words and settles the question. A good habit: whenever a number will go into a decision, define the terms that produce it.
To see how much one word can move a number, take a base of 10,000 customers. If "active" means logged in this month, 7,200 qualify. If it means made a transaction in the last 90 days, 6,100 qualify. If it means made a transaction in the last 30 days, 3,800 qualify. The same question, "how many active customers do we have?", has three honest answers that differ by almost a factor of two.
Models do this too. Asked for "active customers" with no definition, an assistant will pick whichever meaning is most common in its training, which may not be yours. A definition line turns that hidden choice into a visible one.
Conflicting instructions
When two instructions cannot both be met, the model has to choose, and you will not control which one it prefers. The classic conflict is "be brief, but be thorough". Another is "use only the attached document" combined with "add relevant examples from your own knowledge". A subtler one hides in a long prompt: a tone instruction near the top asks for a casual style, and a later pasted example is formal. The model blends them.
Fixes are straightforward once you see the conflict. Prioritise: "If you cannot cover everything within 150 words, keep the cost and the deadline and drop the background." Separate: ask for a one-line answer, then a longer one for those who want details. Remove: decide which rule really matters and delete the other. When you reread a long prompt, scan for pairs of instructions that pull against each other, as an editor scans a draft for contradictions.
Overload: why ten rules are not followed ten times
Every added rule is one more thing the model has to satisfy, and each has a chance of being missed. If each rule is followed 95 times in 100, then with one rule you are almost safe. With ten rules, the chance that all ten are followed is 0.95 raised to the power 10, which is about 60 per cent. With twenty rules, it is about 36 per cent. These numbers are illustrative, since real rules differ in difficulty and models often do better than this, but the shape is right: the more rules, the more likely at least one is broken.
What to do about it is practical and not mysterious.
Keep the list short. Ask which rules would actually change your decision to use the output, and delete the rest.
Rank them. Mark the two or three that must never be broken. "Must": never state a price not in the sheet. "Should": keep it under 120 words. "Nice": use a friendly sign-off.
Group related rules. Five formatting rules under one heading read more easily than five scattered sentences.
Check the output against the list. Ask the model to review its own answer: "Before you reply, check your draft against each rule above and fix any that you broke." Lesson 3.4 develops this.
Move the critical rule. Put it where it is hard to miss, and repeat it near the end of a long prompt.
Say what to do, not only what to avoid
A prohibition tells the model what to leave out, but not what to put in its place. "Don't write long paragraphs" leaves the model free to write ten medium ones. "Write in short paragraphs of at most three sentences" gives it something to follow. Negative instructions also sometimes call attention to the very thing they ban, in the way that "don't think of an elephant" does. Use them for hard limits, and pair them with a positive version: "Do not mention competitors. Focus only on our own features."
Negative only · Positive and negative"Don't use jargon." · "Use plain words a school student would know. Avoid jargon."
"Don't be too long." · "Keep it under 120 words."
"Don't make up numbers." · "Use only numbers from the table. If one is missing, write 'not provided'."
"Don't be rude." · "Stay polite and calm, even if the customer is angry."
## Let the model help you find the gaps
You do not have to find every ambiguity alone. The model can interview you before it answers, which is one of the most effective ways to close gaps in a prompt. Add a line such as:
text
Before you answer, ask me up to three questions if anything in this request is
unclear. If nothing is unclear, go ahead.
If you do not want to be interrupted, ask for the assumptions instead: "At the end, list any assumptions you made." Either approach surfaces guesses that you can then replace with facts. Over time you will notice the questions the model asks most often, such as "who is the audience?" and "how long?", and you will start answering them in the first prompt, which makes the first prompt better.
Three rewrites
Rewrite 1: an operations request. A logistics supervisor writes: "Summarise today's delivery problems and tell me what to fix." This is a wish with two jobs. The sharp version is: "Below is today's delivery log. List the three most frequent causes of delay, with a count for each. For each cause, suggest one action that could be done this week by the depot team. If the log does not show a cause, write 'not shown'. Use a table with the columns Cause, Count and Action." It has a number of items, a definition of the output, a boundary and a format.
Rewrite 2: a finance request. A manager writes: "Explain the variance." The model has no idea which variance, of what, to whom. The sharp version is: "Explain in plain words, for a non-finance director, why the marketing budget was 18 per cent over plan in Q2. Use only the figures below. Give the biggest two reasons with the rupee amount of each, then one sentence on whether it will repeat in Q3. Do not speculate beyond the figures. Under 120 words."
Rewrite 3: a people request. An HR partner writes: "Write feedback for Priya, she's doing well." This is a direction with no content, and the result will be generic praise. The sharp version is: "Draft feedback for Priya, a junior analyst, for her annual review. Use these facts only: she built the weekly dashboard that saved the team about four hours a week, she presented to the director twice, and she needs to ask for help earlier on blocked tasks. Structure it as two strengths and one area to develop, each with a specific example. Keep the tone warm and honest, and avoid generic phrases such as 'hard-working' or 'team player'. Under 150 words."
In each case the sharp version is longer, but each added phrase closes a specific gap: audience, number, source of facts, format, boundary or tone. None is padding.
Reading your own prompt as a stranger would
A quick edit pass catches most of these problems. Print or copy the prompt, and go through it with four questions, marking each sentence.
1. Could I check this? Put a tick beside every instruction a stranger could verify, and a circle beside every one they could not.
2. Do I know what this word means here? Underline terms that carry a specific meaning in your workplace.
3. Does anything contradict this? Draw a line between any two rules that pull against each other.
4. Would the answer change if I deleted this? Strike through sentences that make no difference. They are padding, and they dilute the rules that matter.
Suppose a first draft has fourteen sentences. After the pass, you might circle four soft instructions to sharpen, underline two terms to define, draw one line between "be brief" and "include all details" and strike through three sentences of background that change nothing. The prompt gets shorter and sharper at the same time, which is the usual result. Fewer rules, each one clear, beat a long list of vague ones, and the arithmetic of overload explains why.
A last check: the definition of done
Before you send a prompt that matters, finish this sentence: "A good answer will..." If you can complete it in a line, such as "a good answer will be a table of three causes with counts, each with one action", you have a definition of done and the prompt probably says it. If you cannot, the prompt is still a wish. This is the same discipline from Lesson 3.1 of the Data Science course: no analysis without a decision attached, and no prompt without a picture of the answer.
Common mistakes
Using soft words as if they were rules. "Concise", "professional" and "good" are wishes. Give a number, an audience or an example.
Leaving company terms undefined. Define "active", "overdue" and "net" when they matter.
Contradicting yourself. Scan for pairs of rules that cannot both be met, and say which wins.
Piling on rules. Rank them, group them and cut the ones that do not change your decision.
Writing only prohibitions. Say what to do as well as what to avoid.
Burying the key rule. Put it first or last, and repeat it in a long prompt.
Never asking the model what is unclear. Invite questions or ask for assumptions.
Accepting a reply without checking each rule. Tick the rules off against the output.
Key takeaways
Most instruction failures come from ambiguity, undefined terms, conflict, overload or burial.
Climb the specificity ladder from a wish to a direction, a measurable target and a testable rule with a reason.
A good instruction is checkable by someone who did not write it. Soft style words can stay soft, but anything that matters should be testable.
Define workplace terms in the prompt, and resolve conflicts by ranking or removing rules.
Many rules lower the chance that all are followed. At 95 per cent each, ten rules are all met only about 60 per cent of the time, so keep the list short and rank it.
Say what to do as well as what to avoid, and ask the model for its questions or assumptions to find gaps.
Check yourself
1. Put these in order from vaguest to sharpest: "Under 80 words", "Make it shorter", "Make it better", "Under 80 words because it is a text message".
"Make it better", "Make it shorter", "Under 80 words", "Under 80 words because it is a text message".
2. Why is "Be concise and cover everything in detail" a problem, and how would you fix it?
The two rules conflict. Rank them, for example "Under 150 words; if you cannot cover everything, keep the cost and deadline and drop the background."
3. If each rule is followed 90 per cent of the time, what is the chance that five rules are all followed?
0.9 to the power 5 = 0.59, about 59 per cent.
4. Rewrite "Don't make it too long" as a positive, checkable instruction.
"Keep it under 120 words."
5. What line could you add to make the model reveal its guesses?
"At the end, list any assumptions you made," or "Ask me up to three questions if anything is unclear."
Try this
You need any assistant and about forty-five minutes.
1. Collect three prompts you have used, or would use, at work. Write each exactly as you would type it.
2. For each, mark every instruction as a wish, direction, measurable target or testable rule. Underline every undefined term and every pair of conflicting rules.
3. Rewrite each prompt so that every important instruction is at least a measurable target, with definitions for any term that needs one, and a fallback for missing information.
4. Add the line asking the model to raise up to three questions, and run the original and the rewritten prompt in fresh chats. Note which questions the model asks of the original.
5. Score both outputs with this rubric, one point per yes.
Check · Original · RewrittenDid it follow every stated rule? · ·
Is the length and format what you wanted? · ·
Did it avoid invented facts? · ·
Would you send it with light edits? · ·
6. Stress-test overload. Write a prompt with eight formatting rules, run it three times, and count how many rules were followed each time. Use the rule-count simulator with your observed rate.
7. Write down your three most useful sharpening moves.
A good result is three before-and-after prompts, each with a visible gain on the rubric, and a short list of the vague words you will now replace. With this lesson Module 1 is complete: you know what a prompt is, how a model reads it, how to structure it and how to make each instruction sharp. Module 2 teaches the techniques that build on these foundations, beginning with the single most powerful addition to a prompt: examples.
Tags: Prompt Engineering Instructions Clarity Constraints Beginner