A medical writer opens the enterprise model on a Wednesday and types eight words into the box: "Draft the efficacy section of the Clinical Overview." The model obliges, because the model always obliges. It produces a fluent, structurally plausible 2.5.4 that quietly assumes the trial was a two-arm superiority study when in fact it was a three-arm non-inferiority design, invents a comparator the protocol never named, and orders its subsections in a sequence that does not match the eCTD granularity the publisher expects. None of those errors are the model failing. They are the model doing exactly what an eight-word prompt instructs it to do, which is to guess the other ninety percent of the brief. The difference between that draft and a usable one is not a better model or a cleverer trick. It is a prompt with a skeleton. This lesson teaches the skeleton: the five named components every regulated-writing prompt needs, why each one closes a specific failure mode, and the five-line preamble pattern that turns a paragraph of vague intent into an instruction a model can execute against ICH E3 and eCTD M4 without improvising.

Why the Prompt Is the Specification, Not the Request

In ordinary use, a prompt feels like a request: you ask, the model answers, and if the answer is wrong you ask again. In regulated writing that framing is the source of most of the danger, because the model treats every gap in your request as permission to invent. The prompt is not a request. It is a specification, in the same sense that a Statistical Analysis Plan is a specification for an analysis: anything you do not pin down, the model gets to decide, and it will decide in the direction of the most statistically common pattern in its training data rather than the specific reality of your study.

This reframing changes what a good prompt looks like. A request is short because brevity is polite. A specification is complete because every omission is a decision delegated to a system that cannot see your protocol, your TLF package, or your house style, and that experiences a missing constraint as a silence rather than a question. When you do not name the analysis population, the model picks the one most common in oncology abstracts. When you do not forbid invented references, the model supplies the citations the genre expects. When you do not specify the output structure, the model produces the structure that appeared most often in the documents it learned from, which may be a journal article rather than a Module 2.5 section.

The practical consequence is that prompt-writing for submission content is a drafting skill with the same seriousness as drafting the document itself. The named author who signs the Clinical Overview owns every decision the prompt delegated by omission. So the discipline is to delegate nothing by accident. Every component in the skeleton that follows exists to convert one category of silent invention into an explicit, auditable instruction, so that what the model produces is a draft of your study and not a draft of the average study.

Component One: Role, the Frame That Selects the Register

The first component is role, the sentence that tells the model who it is being for the duration of this task. "You are a regulatory medical writer drafting submission content for an NDA" is not decoration. It selects a register, a vocabulary, and a set of conventions from inside the model's learned distribution, steering it away from the marketing copy, the patient-facing summary, and the journal manuscript that share the same subject matter but obey different rules. A model told it is a medical writer will reach for ICH-conformant phrasing; a model given no role will reach for whatever phrasing is most globally common, which is rarely submission-grade.

Role is also where you set the posture toward uncertainty. "You are a regulatory medical writer who flags missing data rather than filling it, and who never states a result that is not present in the provided sources" does real work, because it pushes the model toward the conservative behavior that regulated writing requires and away from the helpful completionism that is its default. The model will not perfectly obey this, and the named author still verifies, but the role sentence measurably shifts the base rate of invention, and a measurable shift on a high-volume task is worth claiming. Role is the cheapest lever in the prompt and the one writers most often skip.

Component Two: Context, the Anchor to a Named Section

The second component is context, and it is the component that separates a regulated prompt from a generic one. Context names the exact artifact and the exact standard: "You are drafting Section 5.3.5 of a Clinical Study Report aligned to ICH E3," or "You are drafting the Module 2.5.4 efficacy summary of the Clinical Overview, aligned to ICH E3 for the underlying CSR and ICH M4E for the CTD summary structure." This sentence does more than orient the model. It pins the document to a published structure that the model has seen, which means the model can reproduce the correct section ordering, the correct level of summarization, and the correct relationship between this section and its siblings.

The failure that context prevents is subtle and expensive. Without it, the model defaults to the structure most common in its training data, and for clinical efficacy content that is often a journal article, which summarizes differently than a Module 2.5 and at a different granularity than a Module 2.7.3. A writer who asks for "the efficacy section" gets a draft that reads well and is structurally wrong for the dossier, and structural wrongness is harder to catch than factual wrongness because it hides in the shape rather than the numbers. Naming "ICH E3 Section 11" or "Module 2.5.4 under ICH M4E" tells the model which shape to produce, and the shape is most of what you are buying from the first draft.

Context is also where you state the granularity you need relative to eCTD M4. A Module 2.5 is a critical, interpretive overview that cross-references rather than reproduces; a Module 2.7.3 is a detailed summary of efficacy; a CSR Section 11 is the primary analysis narrative. These are different documents about the same results, and the model can produce any of them, but only if the context sentence tells it which. The granularity instruction is the difference between a draft that slots into the dossier and one that has to be rewritten to fit.

Component Three: Source Documents, the Ground the Claims Stand On

The third component is the source documents, the actual text the model is allowed to draw facts from, loaded into the context window and labeled clearly enough that the model can cite them. This is the component that most directly governs whether the draft is grounded or invented, because, as the prior level established, the model reasons only over what is in the window and treats a missing source as a silence rather than a gap. If the synopsis, the integrated efficacy section, and the final TLF package are in the window and labeled, the model has something true to transcribe. If they are not, the model has only its training distribution, and it will produce plausible values that wear the costume of measured ones.

Labeling matters more than writers expect. "Source A: CSR synopsis; Source B: integrated efficacy section 11.4.2; Source C: TLF package, tables 14.2.1.x" gives the model named handles it can reference in its citations, which is the precondition for the source-link discipline that a later lesson builds into a full pattern. A model that knows the loaded tables are "Source C" can be instructed to cite "Source C, Table 14.2.1.4" for every numeric claim, and a citation that points at a named loaded source is one a human can check in seconds. The source-documents component and the constraint component work together: loading the sources makes grounding possible, and the constraints make grounding mandatory.

Component Four: Constraints, the Rules That Bound the Output

The fourth component is constraints, the explicit prohibitions and requirements that bound what the model may and may not do. The single most important constraint in regulated writing is "do not invent references," stated in a form the model can act on: "Cite only tables, sections, and references present in the provided sources. If a claim cannot be supported by a provided source, write the claim and append the flag NEEDS SOURCE rather than inventing a citation." This converts the model's most dangerous default, the fabrication of plausible cross-references, into a visible flag the writer can resolve, which is far safer than a fabricated citation that reads as authoritative and survives a hasty review.

Constraints also encode the negative space of the document. "Do not state a hazard ratio, confidence interval, or p-value that is not present in the provided TLF tables." "Do not characterize the result as statistically significant unless the provided source states the threshold was met." "Do not introduce a comparator, an endpoint, or an analysis population not named in the provided protocol synopsis." Each of these closes a specific, observed failure mode, and each is worth more than a paragraph of polite instruction, because the model obeys a sharp prohibition far more reliably than a vague preference. The art of the constraint block is naming the inventions you most fear by name, so the model cannot reach for them.

There is a discipline to keeping the constraint block tight. A prompt with forty constraints dilutes the few that matter, and the model's adherence to any single constraint weakens as the list grows. The high-value constraints for submission content are few and stable: cite only provided sources, flag rather than fill missing data, do not invent statistics or comparators, and match the named output structure. Those four carry most of the safety, and a writer who states them sharply and stops has a prompt that performs better than one buried under thirty stylistic preferences.

Component Five: Output Format, the Shape the Publisher Can Use

The fifth component is output format, the specification of how the draft is structured on the page so that it slots into the downstream workflow rather than requiring reformatting. For submission content this means Markdown with section numbering that matches the target eCTD M4 granularity: a Module 2.5.4 draft should come back with the 2.5.4 heading and its conventional sub-structure, not as a wall of prose, and a CSR Section 11 draft should carry the E3 sub-numbering. Specifying the format is not cosmetic; it is what lets the draft be reviewed against the structure it is supposed to have, because a reviewer comparing a numbered draft to ICH E3 can see at a glance whether a required sub-section is missing.

The output-format component is also where you can request the scaffolding that makes verification fast. Asking the model to produce, alongside the prose, a claim-to-source table that lists each factual claim with its supporting source and locator turns the draft into a self-documenting artifact: the writer reads the table, checks each row against the loaded source, and resolves any NEEDS SOURCE flag before reading the prose for style. This is the structural form of the show-your-sources pattern, and specifying it in the output-format component is how you get the model to build its own verification harness. A draft that arrives with its claims already mapped to sources is a draft you can reconcile in a fraction of the time.

The Five-Line Preamble Pattern

The five components assemble into a compact, reusable preamble that a writer can keep as a template and adapt per task in under a minute. The pattern is five lines, each one a component, in a fixed order so that nothing is dropped by accident: Role, Context, Task, Constraint, and a Source clause that names the reason-code behavior for grounding. Written out for a real task it reads like this: "Role: You are a regulatory medical writer drafting NDA submission content who flags missing data rather than filling it. Context: You are drafting the Module 2.5.4 efficacy summary aligned to ICH E3 and ICH M4E at eCTD M4 granularity. Task: Summarize the primary and key secondary efficacy results from the provided sources into a Module 2.5.4 draft. Constraint: Cite only the provided sources; do not invent tables, statistics, or comparators; flag any unsupported claim as NEEDS SOURCE. Source clause: For every numeric claim, append the source label and locator, for example (Source C, Table 14.2.1.4); produce a claim-to-source table after the prose."

The power of the fixed order is that it is a checklist disguised as a sentence. A writer who has internalized Role, Context, Task, Constraint, Source clause will notice, at draft time, when one is missing, the way a pilot notices a skipped item on a pre-flight list. The five-line preamble is not a magic incantation; it is the minimum specification that closes the five most common categories of silent invention, and it is short enough to write every time, which is the property that actually matters. A perfect prompt you write once and abandon is worth less than a good prompt you write on every task, because the failure modes recur on every task.

The preamble is also the unit that belongs in the audit trail. When the time comes to document how AI assisted a section, the five-line preamble, together with the system prompt, the model and version, the temperature, the loaded sources, and the human-verified artifact, is the record that survives an inspection. A vague paragraph of intent cannot be reconstructed or defended; a structured five-line preamble can be read back, understood, and shown to have constrained the model in the ways the regulation expects. The skeleton makes the prompt not only safer to run but defensible after the fact, which in this domain is the same thing.

Adapting the Skeleton Across Artifacts Without Losing It

The five-component skeleton is not a Module 2.5 device; it is the same frame for every regulated artifact, and what changes from task to task is the content of each component, not the components themselves. Drafting a CSR Section 12 safety narrative, the Role stays "regulatory medical writer who flags missing data," but the Context shifts to "CSR Section 12 under ICH E3," the Sources become the safety TLF package and the narrative listings, and the Constraints sharpen to forbid inventing adverse-event frequencies or recoding a preferred term the source did not use. Drafting a Module 3.2.P.5 control-of-drug-product section, the Context becomes "Module 3.2.P.5 under ICH Q6A," the Sources become the specification and the analytical validation report, and the highest-value Constraint becomes "state no acceptance criterion not present in the loaded specification." The skeleton is portable precisely because the failure mode it guards against, silent invention of the part of the document you did not specify, is universal across artifacts, and only the specifics of what must not be invented change with the document.

Learning to adapt the skeleton fast is mostly a matter of learning the high-value Constraint for each artifact class, because the Constraint block is where the document-specific danger lives. For an efficacy summary, the danger is invented statistics and comparators. For a safety narrative, it is invented frequencies and miscoded terms. For a CMC section, it is invented acceptance criteria and methods. For a protocol section, it is invented eligibility criteria or endpoints not in the target product profile. A writer who keeps a short, per-artifact list of the inventions they most fear can drop the right Constraint into the skeleton in seconds, and that named prohibition does more to make the draft safe than any amount of stylistic guidance, because the model obeys a sharp, specific prohibition far more reliably than a general instruction to be careful. The skeleton is the reusable container; the per-artifact Constraint is the payload that makes it fit the document in front of you.

There is a final adaptation that separates a competent prompt from a defensible one, which is matching the Output format to the actual downstream tool rather than to a generic idea of tidiness. A draft destined for Certara CoAuthor or Veeva Vault should come back in the structure that tool ingests cleanly, with the section numbering the publisher expects and the claim-to-source table in a form that can be lifted into a reconciliation log without retyping. The same content, formatted for the wrong destination, costs the writer a reformatting pass that the format component existed to prevent. The discipline is to write the Output line for the next human and the next system that will touch the draft, not for the screen it first appears on, so that the draft slots into the workflow rather than interrupting it. Across all five components, the through-line is the same: the prompt is a specification of everything the model would otherwise invent, and adapting it well means knowing, for each artifact, exactly what that everything is.

Key Takeaways

• A regulated-writing prompt is a specification, not a request, and every omission is a decision delegated to a model that cannot see your study. The model fills gaps with the most common pattern in its training data, so an eight-word prompt produces a draft of the average study rather than yours. The named author owns every decision the prompt delegated by accident.
• The five components are Role, Context, Source documents, Constraints, and Output format, and each closes a specific failure mode. Role selects the register, Context anchors the section to ICH E3 and eCTD M4 granularity, Sources ground the claims, Constraints forbid invented references and statistics, and Output format makes the draft slot into the dossier and self-document its claims.
• Context is the component that separates a regulated prompt from a generic one. Naming "Module 2.5.4 under ICH M4E" or "CSR Section 11 under ICH E3" tells the model which structure to produce, preventing the silent default to a journal-article shape that reads well and is wrong for the dossier.
• The highest-value constraints are few and sharp: cite only provided sources, flag rather than fill missing data with NEEDS SOURCE, do not invent statistics or comparators, and match the named output structure. A tight constraint block outperforms a long one because adherence to any single constraint weakens as the list grows.
• The five-line preamble, Role, Context, Task, Constraint, Source clause, is a checklist disguised as a sentence. It is short enough to write on every task, closes the five common categories of silent invention, and becomes the defensible record in the audit trail. A good prompt written every time beats a perfect prompt written once.