A disclosure analyst opens a chat window at 4 p.m. with a question her CFO asked an hour ago: what emission factor should we use for the diesel our fleet burned last year. She types "what is the emission factor for diesel" and hits enter. Two seconds later she has a number, 2.68 kgCO2e per litre, delivered with total confidence and not a single word about where it came from, which year it represents, or which database published it. The number looks right. It may even be right. And if she pastes it into the inventory, she has just introduced a figure into an externally assured disclosure that she cannot defend to the assurer who will ask, three months from now, "show me the source." The prompt was the problem. Not the model.

The Prompt Is the First Control, Not a Convenience

Most people meet a large language model as a search box: type a question, get an answer, move on. For a regulated discloser that mental model is dangerous, because the thing you are producing is not an answer for yourself, it is a figure or a claim that will sit inside an externally assured report and be read by an assurer who can pull any number and demand its basis. In that world the prompt is not a convenience. It is the first control in your evidence chain, the place where you decide whether the output that comes back can be defended or will quietly become a liability.

The discipline this lesson teaches is small and almost mechanical, and it is the single highest-leverage habit in the entire program. It has two halves. First, give the model the context it cannot guess: which framework you are reporting under, which boundary you are inside, which exact datapoint you need, and which source material it is allowed to use. Second, demand provenance or refusal: tell the model that it must cite the source for any figure or claim, and that if the source is not in the material you provided, it must say so rather than supply a plausible number. Context in, provenance or honesty out. That is the whole of prompt discipline for disclosure, and the rest of this lesson is why each half matters and what it looks like when you do it well.

Notice what this is not. It is not prompt engineering as a clever trick to extract better wording. It is the disclosure professional's job, expressed in a chat window. You already know that a number in your report needs a context (the reporting framework and boundary) and a source (the evidence underneath it). All prompt discipline does is move that knowledge to the front of the conversation, so the model produces something shaped like a defensible figure instead of something shaped like a confident guess.

The Context the Model Cannot Guess

A vague prompt fails because the model fills every gap you leave with the statistically most likely continuation, not with your specific reporting reality. "What is the emission factor for diesel" leaves almost everything unspecified, and the model resolves that ambiguity by guessing: it picks a common factor, from some unstated year, expressed in some unit, drawn from a blend of whatever its training data contained. None of those choices are yours, and none are documented. The output is plausible precisely because it is the average of everything, which is also exactly why it is useless as evidence.

The fix is to supply the context that pins the answer to your situation. For a disclosure professional, that context is rarely mysterious. It is the same set of facts you would write into a basis of preparation. Four pieces carry most of the weight.

Which Framework

State the framework you are reporting under, because it changes what counts as a correct answer. A factor that is acceptable for a voluntary footprint may not satisfy the methodology your ESRS (the European Sustainability Reporting Standards, the detailed disclosure rules under CSRD) datapoint requires, and a CBAM (the EU Carbon Border Adjustment Mechanism, which prices embedded emissions in imported goods) embedded-emissions calculation follows its own prescribed rules entirely. Naming the framework tells the model which rulebook governs, and tells you, when you read the output, whether it answered the right question. "Under the GHG Protocol, for a Scope 1 mobile-combustion calculation" is a different and far more answerable request than "for diesel."

Which Boundary

State the boundary: the organizational and operational edge of what you are counting. The same litre of diesel can be in scope or out of scope depending on whether the vehicle is owned, leased, operationally controlled, or a supplier's. If you do not tell the model which boundary you are inside, it cannot help you decide whether an item belongs in Scope 1, Scope 2, or Scope 3, and it will cheerfully assume one. Boundary is also the place where the most expensive silent errors live, because an undocumented exclusion is an assurance finding, so making the boundary explicit in the prompt is also making it explicit in your own thinking.

Which Datapoint

Name the exact datapoint you need, in the units you need it. "The emission factor for road diesel combustion, expressed in kgCO2e per litre, for reporting year 2024" is a request the model can either satisfy precisely or honestly fail. "The emission factor for diesel" invites it to pick. Specificity in the request is what lets you judge specificity in the answer; a precise question makes a vague answer visibly inadequate, where a vague question makes a vague answer look fine.

Which Source It May Use

This is the piece newcomers skip and professionals never do. Tell the model what material it is allowed to draw on. If you have pasted the supplier's response, the utility bill, or the factor database extract, instruct the model to answer only from that text. If you have not provided source material, you must understand that any figure the model returns is drawn from its training data, which is to say from nowhere you can cite. The next lesson is dedicated to this technique of grounding the model on a document; here the point is narrower and prior to it: at minimum, the prompt must make clear whether the model is working from your evidence or from its own memory, because those two produce outputs of completely different evidentiary value.

The Demand That Changes Everything: Cite the Source or Refuse

Context shapes the question. The second half of prompt discipline shapes what the model is permitted to do when it does not know the answer, and this is where disclosure prompting diverges hardest from ordinary use. By default, a language model would rather give you a fluent, plausible answer than admit it lacks one. That default is catastrophic for a discloser, because a plausible-but-unsupported number is the exact failure mode that fails assurance: it looks like data, it reads like data, and it has no evidence underneath it. A hallucinated emission factor does not announce itself. It arrives wearing the same confident voice as a correct one.

So you must explicitly override the default. The instruction is simple and you should treat it as non-negotiable boilerplate in any prompt that could produce a figure or a factual claim: cite the source for every figure and claim, and if the source is not in the material provided, say "not in the provided material" rather than supply a number. You are giving the model permission, and more than permission, an instruction, to refuse. You are telling it that an honest "I do not have this" is a success and a confident invented number is a failure. That inversion is the heart of the discipline.

In disclosure, "I do not have a source for that" is a correct answer and a confident unsourced number is a misstatement waiting for an assurer. Prompt for the honest refusal, not the fluent guess.

Why does this work, when a model cannot truly know whether it is hallucinating? Because the instruction changes the target the model is optimizing toward. Left to its defaults, it aims at "produce a fluent, helpful-looking answer," and an invented factor hits that target perfectly. Told to cite or refuse, it aims at "produce an answer with a source attached, or decline," and now an invented factor misses, while "not in the provided material" hits. You have not made the model incapable of error. You have made the kind of output you want, the sourced answer or the honest gap, the kind it is trying to produce. Combined with grounding on real source material, this is most of what stands between you and a fabricated number in a public disclosure.

There is a second benefit that matters in the assurance file. When the model returns "not in the provided material," it has just told you, in seconds, that you have a data gap. That is not a failure of the workflow; it is the workflow working. A disclosed, visible gap is something you can fill with a documented estimate or a supplier follow-up. A gap papered over with a confident guess is a landmine. Prompting for refusal turns the model into an early detector of exactly the holes you most need to find.

Worked Example: The Vague Prompt Versus the Disciplined Prompt

Watch the same question asked two ways, and watch the two answers diverge into a defensible figure and a liability. The task is real and ordinary: the analyst needs an emission factor for the diesel her owned fleet burned in 2024, for the Scope 1 section of an inventory that will be externally assured.

The Vague Prompt and Its Plausible Poison

She types: what is the emission factor for diesel

The model answers, instantly and confidently: "The emission factor for diesel is approximately 2.68 kgCO2e per litre." Clean, specific, and ready to paste. And almost entirely undefendable. There is no year, so she cannot know if it matches her 2024 reporting period. There is no named database, so she cannot cite it. There is no statement of scope (is this tank-to-wheel, well-to-wheel, combustion only?), so she cannot confirm it matches her boundary and methodology. The 2.68 is a real-looking number that the model assembled from the average of its training data, and if it goes into the inventory, the first question the assurer asks ("which database, which year, on what basis") has no answer in the file. The number is not wrong so much as it is orphaned: a figure with no parents, which in disclosure is the same as a misstatement, because she cannot prove it is right.

The Disciplined Prompt and Its Defensible Answer

She tries again, this time doing the disclosure professional's job in the prompt: I am calculating Scope 1 mobile combustion under the GHG Protocol for reporting year 2024. The boundary is our owned fleet. I have pasted below an extract from the UK DEFRA 2024 conversion factors. Using only the pasted extract, give me the emission factor for road diesel in kgCO2e per litre, and cite the exact row. If the figure is not in the pasted extract, reply "not in the provided material" and do not supply a number from elsewhere. Below it, she pastes the relevant rows of the factor table.

Now one of two things happens, and both are wins. If the factor is in the pasted extract, the model returns it with the citation she demanded: the value, the units, and the exact row from the named, dated DEFRA 2024 source. That is a figure she can drop into the inventory and defend on sight, because its provenance travels with it. If the factor is not in the extract, the model replies "not in the provided material," and she has learned in two seconds that she needs the right table, rather than learning it from the assurer in three months. Same question, same model, opposite evidentiary outcome. The first prompt produced a number she would have to defend from memory. The second produced either a sourced number or an honest gap, which are the only two things a discloser should ever accept from a machine.

Why That Difference Is the Whole Job

The gap between those two interactions is not a gap in the model's capability. The same model produced both. It is a gap in the discipline of the person at the keyboard. The vague prompt asked the model to be a confident oracle, and it obliged. The disciplined prompt asked the model to be a sourced research assistant constrained to provided evidence, and it obliged that too. The model is a mirror for the rigor of your request. In ordinary life a sloppy prompt costs you a slightly worse answer. In disclosure it costs you a number you cannot defend, and the cost lands months later, in front of an assurer, when it is most expensive to fix.

Building the Habit So It Survives a Deadline

A discipline that only works when you are calm and unhurried is not a discipline; it is a luxury. The pressure of a reporting deadline is exactly when people revert to the search-box prompt and exactly when an unsourced number is most likely to slip through. So the goal is to make context-and-citation so routine that it is faster to do than to skip.

The practical move is to keep a short standing preamble you paste at the top of any disclosure prompt, something like: "I am a sustainability discloser. For any figure or factual claim, cite the source. If a figure is not in the material I provide, say so and do not supply one from elsewhere. State any assumption you make explicitly." That preamble is the cite-or-refuse rule and the no-silent-assumptions rule, frozen so you never have to retype them. Later in this level you will see this idea formalized into a reusable system prompt; for now, a copied paragraph is enough to change your default behavior, which is the entire point.

Then add the situational context each time: framework, boundary, datapoint, and the source material if you have it. Two or three sentences of context, plus the standing preamble, plus your actual question. It feels like more work than typing "what is the factor for diesel." It is perhaps thirty seconds more. And it is the thirty seconds that decides whether the output is evidence or exposure. Across a reporting cycle, that trade is not close.

One last reframing to carry with you. The reason prompt discipline matters is not that the model is untrustworthy and you must trick it into honesty. It is that you, the discloser, are accountable for every number you publish, and the prompt is the first place you exercise that accountability. A well-formed prompt is you doing your job at the input, so that the output arrives already shaped like something you can stand behind. The model never becomes accountable. You always are. Prompting well is simply the earliest moment in the workflow where that truth shows up in what you type.

Five Sloppy Prompts and Their Disciplined Rewrites

To make the habit concrete, here are five prompts a disclosure professional might actually type in a hurry, each followed by the disciplined version. Notice that in every case the fix is not cleverness; it is restoring the framework, the boundary, the datapoint, the source, and the cite-or-refuse instruction that the rushed version dropped.

Sloppy: "Summarize this supplier's emissions performance." Disciplined: "Using only the supplier response pasted below, list the emissions figures the supplier actually reported, with the quoted line for each. If a figure is described as estimated rather than measured, say so. Do not add any figure not present in the text." The sloppy version invites the model to characterize and embellish; the disciplined version confines it to extracting what is really there and preserving the measured-versus-estimated distinction your file depends on.

Sloppy: "What is our Scope 2 emissions?" Disciplined: "From the pasted utility bills below, total the kWh for our owned sites for 2024, quote each billing line, and apply the market-based factor I provide in the next message. Do not supply a factor yourself; if a bill is missing a period, flag the gap." The sloppy version asks the model to invent both the consumption and the factor; the disciplined version makes it a reader and a calculator over your evidence, and surfaces missing data instead of papering over it.

Sloppy: "Is this target ambitious enough?" Disciplined: "Quote the company's stated target from the policy pasted below, state its base year and target year exactly as written, and do not assess ambition or compare it to any external benchmark unless I provide one." The sloppy version invites the model to opine and possibly to invent a comparison; the disciplined version keeps it tethered to the words on the page.

Sloppy: "Write the climate narrative section." Disciplined: "Draft the climate narrative using only the figures and facts in the material pasted below. For every quantitative claim, cite the figure it rests on. Do not assert any progress, target, or trend that is not supported in the provided material; where support is missing, insert a bracketed note saying so." The sloppy version produces fluent, confident, and possibly fabricated progress claims; the disciplined version produces a draft whose every claim is either supported or visibly flagged as unsupported.

Sloppy: "Give me the emission factor for natural gas." Disciplined: "From the DEFRA 2024 extract pasted below, give the natural gas factor in kgCO2e per kWh for gross calorific value, quote the exact row, and reply 'not in the provided material' if it is absent." The sloppy version yields an orphaned number; the disciplined version yields a sourced, unit-specified, quoted factor or an honest gap. Five prompts, one pattern: the discipline is always the same five elements, restored every time.

Key Takeaways

• For a regulated discloser the prompt is the first control in the evidence chain, not a search box. What you type decides whether the output can be defended or becomes a liability.
• Prompt discipline has two halves: give the context the model cannot guess (framework, boundary, datapoint, allowed source) and demand provenance or refusal (cite the source, or say it is not in the provided material).
• A vague prompt fails because the model fills every gap with the statistical average of its training data, producing a plausible number that is orphaned from any source you can cite.
• Name the framework (GHG Protocol, ESRS, CBAM), the boundary (what is in scope and at which scope), and the exact datapoint with units, because each pins the answer to your reporting reality.
• The cite-or-refuse instruction inverts the model's default: it makes an honest "not in the provided material" a success and a confident invented number a failure.
• A refusal is a feature, not a bug. It surfaces a real data gap in seconds, which you can fill with a documented estimate or a follow-up, instead of discovering it from the assurer months later.
• Same model, opposite outcomes: the vague prompt yields a number you defend from memory, the disciplined prompt yields a sourced figure or an honest gap, which are the only two acceptable outputs.
• Keep a standing preamble (cite or refuse, no silent assumptions) so the discipline survives a deadline, and remember that the prompt is where your own accountability for the number first shows up.