05. Clear Articulation, Curiosity, and the Art of Prompting

Published 2026-09-14

Core Theme: AI does not reward cleverness; it rewards clarity. If you cannot question what you need with the unquenchable curiosity of a small child — and then translate it into unambiguous words — even the most powerful supercomputer in the world cannot give you the result you want.

Why is a curious three-year-old the best model for prompting?

The most common question asked in any session with officers is some version of this: what coding do I need to learn to become a good prompt engineer?

The answer surprises people, and it is not a rhetorical flourish: becoming a good prompter has nothing to do with learning computer code. It requires becoming a three-year-old child again.

Anyone who has lived with a three- or four-year-old knows the behaviour being described. That child spends the entire day asking questions. There is not a waking moment without one.

• "Amma, why is the sun round? Why is the moon round?"
• "Why isn't the sun there at night? Why doesn't the moon come out in the day?"
• "If I climb to the top of Sigiriya and drop my smartphone, will it float away? Will it fall and break? Will it hit the rock on the way down? Will the wind carry it off? If I jump after it, can I catch it mid-air? Will I float too? What will happen to me?"

Figure 5.1: The Articulation Test — how vague curiosity evolves into precision prompt architecture (illustrated through the Sigiriya hydraulic engineering inquiry).

Look closely at that Sigiriya sequence, because it is not random chatter. It is a systematic exploration of one scenario: the phone falls, the phone breaks, the phone hits the rock, the wind intervenes, the child intervenes, the child falls too. Each question opens a branch the previous answer created. No adult officer habitually explores a problem that thoroughly, and it is precisely the pattern the next chapter will dignify with the name tree-of-thought prompting.

Why does the child do it? One reason: curiosity. No grade depends on it and nobody assigned it.

What happens next to that capacity is the uncomfortable part. As we grow, our education system and our social conventions steadily blunt it. Asking too many questions marks a student as difficult. Asking in a meeting can look like ignorance. We are trained, gradually and thoroughly, to stop asking and to accept quietly whatever we are handed.

That training was always a loss. It is now a measurable professional disadvantage, because the interface to the most powerful tool on an officer's desk is a question.

To win in the presence of artificial intelligence, the buried curiosity has to be dug back up. Prompting is nothing more exotic than this: looking at a problem with curiosity, hunting for its root, and asking a continuous, logical chain of questions. A person who questions that way is already an excellent prompter and usually does not know it.

There is one difference between the child and the professional, and it is the whole subject of this chapter. The child asks and asks without needing precision, because a parent fills in the gaps, guesses the intent, and answers the question that was meant rather than the one that was asked. A model does not do that. It answers what you actually wrote. So the curiosity has to be paired with something the three-year-old does not have — exact expression — and that pairing is what the rest of this chapter builds.

What is articulation, and why does it decide your results?

Curiosity is the fuel. On its own it produces motion without direction. The rudder is articulation — the ability to express an idea clearly enough that it survives the journey from your mind into someone else's, or into a machine's.

The Canadian psychology professor Dr. Jordan B. Peterson has put the case for this skill as well as anyone, and for the public officer of the AI age his formulation reads almost as a specification:

Figure 5.5: The Architecture of Articulation — moving from chaotic intuition through linguistic discipline to machine-executable directives.

"A person becomes more articulate by choosing their words with great care, by practicing saying what they believe to be true, by reading and writing deeply about their own thoughts, and by treating the word with immense respect. The clearer your ideas, the more powerful your thinking; and the more powerful your thinking, the more you become a victor before the world."

Notice the direction of the causation in the middle of that passage, because it is the part people misread. The claim is not that clear thinkers happen to write clearly. It is that the act of writing clearly produces clearer thinking — you find out what you actually believe by being forced to put it in exact words, and vague writing is usually the visible symptom of thinking that has not been finished.

That has a direct and slightly uncomfortable consequence for anyone using these tools. When a model returns something shapeless, the instinct is to blame the model. Far more often, the prompt was shapeless, and the output is an honest reflection of an idea that was never fully formed. The machine has become an unusually fast mirror for the quality of one's own thinking, which is useful and not always flattering.

The most powerful person of the AI age, then, is not the one who memorised computer code. It is the person with the articulate power to tell the machine exactly what they need — in precise words, with zero ambiguity.

Which five human skills actually power a prompt?

Getting a flawless official document out of a model depends on five capacities, and every one of them resides not in the machine but in the officer. Figure 5.6 sets them out.

Figure 5.6: The 5 Essential Human Skills of Prompt Engineering — critical reading, explicit writing, structured logic, critical thinking, and precise articulation.

Critical reading: ingesting the context

Read the problem, the complaint, or the draft bill properly, without shortcuts. This is the least glamorous item on the list and the one most often skipped, usually under the honest pressure of a deadline.

The consequence of skipping it is worth stating bluntly: asking an AI to solve something you have not read yourself is one blind man asking another for directions. You will receive a confident answer, and you will have no means of judging whether it addresses the real problem — because you do not know what the real problem is.

Explicit writing: making the instruction precise

Strip out the unnecessary words and convert the goal into direct, clear, imperative instructions.

Most first drafts of a prompt are twice as long as they need to be and half as precise, because they describe the situation at length and the requirement only vaguely. The discipline is the same one that makes a good office memo: say what you want done, plainly, near the beginning.

Structured logic: sequencing the work

Do not demand a complex administrative task in one lump. Decompose it into a logical sequence of sub-routines: review, then eligibility verification, then calculation, then recommendation.

The reason this works connects back to the mechanics in Chapter 03. A model generates each part of its answer conditioned on what it has already produced, so a task broken into ordered stages gives it correct intermediate ground to build on. It also gives you checkpoints — you can catch an error at the eligibility stage rather than discovering it inside a finished recommendation.

Critical thinking: auditing the answer

Examine what comes back for logical gaps, contradictions and false inferences, and do it deliberately rather than by impression.

This requires active resistance, because the output is designed to read as though it has already been checked. Fluent formatting, confident phrasing and a professional tone are properties of the generator, not evidence about the content. The question to hold in mind is not "does this look right?" but "what would be true if this were wrong?"

Precise articulation: arranging the prompt

Finally, arrange the role, the context, the constraints and the format so cleanly that the machine has no room to misunderstand. This is the skill the remaining sections of this chapter are devoted to, because unlike the other four it has an actual structure that can be taught.

What makes an instruction precise?

Ambiguity is not a stylistic weakness in public administration. In the public service a single ambiguous word in a letter can cost a citizen their land title or their pension, or open the road to a High Court case. The four principles in Figure 5.7 are how ambiguity gets removed from a prompt before it does that damage.

Figure 5.7: The Four Pillars of Precise Articulation in Prompt Engineering — Zero-Ambiguity, Negative Constraints, the Intent-Action-Result (IAR) audit model, and Semantic Precision.

Pillar one: zero ambiguity

Remove from the prompt every subjective word that could carry more than one meaning.

The instruction "write this very professionally" feels precise and is not. "Professional" means brisk and commercial to one reader, formal and deferential to another, warm and plain-spoken to a third. Handing the model that word means handing it the choice, and it will resolve the ambiguity by picking whatever is statistically most common — which is how so much AI output arrives sounding like a generic corporate circular.

Replace it with the properties you actually want: "write this in a direct, authoritative tone, leading with action-oriented verbs, in simple sentences a person with an ordinary school education can understand." Longer, unquestionably. But every adjective you specify is a decision you have taken back from the model, and the specification is reusable once written.

Pillar two: constraints as scaffolding

Telling the model what it must not do matters as much as telling it what to do, and this is the pillar most people never adopt at all.

Negative constraints rule out the failures you can predict: "Do not use passive voice. Do not use empty corporate buzzwords like 'synergy' or 'optimization.' No more than two sentences per paragraph." Each of those lines removes a known irritation permanently rather than being fixed by hand in every draft.

Structural constraints dictate shape rather than substance: "Format the output as a 3-column table (Cost, Risk, Velocity)." Specifying the container is a surprisingly effective way of improving the contents, because a three-column table cannot be filled with waffle — it forces the model to commit to a value in every cell.

The word "scaffolding" in the figure is exact. Constraints are not restrictions that reduce what the model can do for you. They are the structure inside which it can build something usable.

Pillar three: the Intent–Action–Result audit

The third pillar is a habit rather than a technique: before pressing Enter, read your own prompt against three questions.

Intent asks what the institutional purpose actually is — for example, "persuade the Ministry of Finance to approve the budget allocation." This is the part most prompts omit entirely, and its absence is why so much output is technically responsive and strategically useless. A document written to inform reads differently from one written to persuade, and the model cannot infer which you needed.

Action asks what specific mechanical task is being delegated: "compare the expenditure reports of the 3 attached Divisional Secretariat offices." Narrow and checkable.

Result asks what form the finished document must take: "a one-page executive summary with a clear approve / do not approve recommendation." Naming the deliverable prevents the most common disappointment of all, which is receiving a perfectly good piece of work in the wrong shape for the meeting it was needed for.

Run those three questions over a prompt and the weak one is usually obvious in seconds — almost always because Intent is missing.

Pillar four: semantic precision

The last pillar is the one that feels like a trick and is in fact a consequence of the architecture described in Chapter 03. Using your field's exact technical or legal terminology, rather than everyday language, anchors the model's attention in the relevant expert region of its training data.

Recall that words occupy positions in a high-dimensional meaning space, and that a model's answer is assembled from the region your words point at. Vague words point at a vague region — the average of everything ever written on the topic. Precise professional vocabulary points at the neighbourhood where the expert material lives.

The effect is easy to observe. "Write good software code" returns something serviceable and generic; "write a stateless functional component in TypeScript, using memoization for efficiency" returns work of a different order. "Make a good promotional plan for farmers" returns platitudes; "using the Jobs-to-be-Done framework, identify the core struggle of the rural farmer applying for the fertilizer subsidy, and design a communication strategy that speaks to it" returns something you could take into a meeting.

The practical implication is worth drawing out, because it reverses a common assumption. Expertise has not been made redundant by these tools — it has become the thing that determines how much you can get out of them. Knowing the right words is knowing where to point.

How is a complete prompt structured?

A prompt built to handle the public service's complex material — a case file, a gazette, a financial report — is not a sentence or two. It is a carefully designed work of architecture, with five distinct layers, each doing a job the others cannot.

Figure 5.8: Prompt Architecture (Full Architecture) — Role, Context, Task, Constraints, and Output Format.

Layer one: role

The expert identity the model is to adopt. "Act as an Additional Secretary with 20 years in the Sri Lanka Administrative Service, expert in the Establishments Code and government financial regulations."

One sentence, and it sets the vocabulary, the priorities and the standard of evidence for everything that follows. The specificity is doing the work — twenty years and a named code, rather than "an expert."

Layer two: context

The background the model cannot know: the situation, the relevant circular numbers, the parties involved. "A dispute has arisen between the local council and the traders' association over the removal of unauthorized shops in the Divisional Secretariat area. Under Circular No. 12 of 2024…"

This layer is where most prompts are thinnest, because the author already holds the context and forgets that the model does not. Everything you leave out, the model will supply from its training data — generically, and possibly from another country's practice entirely.

Layer three: task

The specific operation to perform, stated as an instruction. "Draft a 5-step mediation action plan to bring the two parties to a settlement."

Note that it is one task, not four, and that the number of steps is specified. Ambiguity about scope is ambiguity about the answer.

Layer four: constraints

What must not be done, the legal limits, the length. "Propose nothing that exceeds the powers granted by the Local Government Act. Do not include assumed laws. Limit the response to 300 words."

The middle instruction is the one to notice. "Do not include assumed laws" is the negative constraint from Pillar two doing exactly the work it exists for — closing off the model's most dangerous habit before it starts, rather than catching it in review.

Layer five: output format

The shape the answer must arrive in. "Present with headings, numbered steps, and a closing Markdown table of [Risk | Responsible Officer | Remedy]."

Specifying format is not fussiness about presentation. That closing table forces a named responsible officer and a remedy against every risk — which means the structure you asked for has quietly required the model to produce content it would otherwise have left vague.

Read the five layers together and the underlying logic is clear enough to remember without the diagram: who you are, what you need to know, what to do, what not to do, and what it should look like. Every good prompt answers those five questions. A prompt that disappoints is almost always one that skipped two of them.

Why articulation is the skill worth practising

This expressive power is the skill Idasara Academy returns to relentlessly, and the reason it is treated as foundational rather than optional (Meta-Learning Superpower).

Consider where the failures actually occur. A student fails an examination, or a graduate fails a job interview, rarely because the knowledge was absent. They fail because the knowledge could not be assembled into clear, ordered expression under pressure (Certificates vs Capability). Knowledge that cannot be articulated cannot be assessed, cannot be applied, and — in a working world where the interface to your most powerful tool is a sentence — cannot be used.

The encouraging part is that articulation is a practised skill rather than a talent, and it improves with use in a way most skills do not: every clear prompt you write is a small exercise in stating exactly what you want, and it compounds. For a public officer, practising precise expression in file work and in commanding these tools does double duty. It gets better work out of the machine today, and it makes you an administrator without equal in your service.

In the next chapter, building on these principles, we study, hands-on, the advanced prompting methods used in the public service and the Divisional Secretariat "Promptathon" model.