--- title: Medical Writers' Handbook — Guideline Deep Dives description: >- A practical field guide to the requirements, standards, ethics, reporting guidance, procedures, and templates that govern medical-writing work. publisher: 7Scholar content_scope: >- Educational reference content; current product capability and fit are described by the core and audience pages. --- # Guideline Deep Dives This shelf extends the [Medical Writers' Handbook](/guide) into the rules and expectations that shape a medical document. It is synthesized only from the books and AMWA and EMWA journals in the source corpus. Because those sources span many publication years, this shelf teaches the method of governing a document; it does not certify that a named source, version, or regional requirement is current. At the start of live work, retrieve the current controlling material and involve the responsible specialist when interpretation exceeds the writer's authority. The central discipline is simple: > **A guideline is useful only after the writer knows why it applies, what it changes in the document, and how conformance will be checked.** ## Contents - [The governing stack](#the-governing-stack) - [Build the stack for a live project](#build-the-stack-for-a-live-project) - [Publication ethics and good practice](#publication-ethics-and-good-practice) - [Study-reporting guidelines](#study-reporting-guidelines) - [Statistical and visual reporting](#statistical-and-visual-reporting) - [Clinical and regulatory writing](#clinical-and-regulatory-writing) - [Safety and pharmacovigilance](#safety-and-pharmacovigilance) - [Medical devices and IVDs](#medical-devices-and-ivds) - [HEOR, HTA, and real-world evidence](#heor-hta-and-real-world-evidence) - [Disclosure and transparency](#disclosure-and-transparency) - [Plain language, readability, and accessibility](#plain-language-readability-and-accessibility) - [Translation and localization](#translation-and-localization) - [AI-assisted medical writing](#ai-assisted-medical-writing) - [Templates, SOPs, and house style](#templates-sops-and-house-style) - [Turn any governing source into a working tool](#turn-any-governing-source-into-a-working-tool) ## The Governing Stack Medical writers use the word _guideline_ for sources with very different force. A law is not a journal checklist. An agency guidance is not a sponsor procedure. A treatment recommendation may be evidence for the document without governing its structure. A template may reveal an expected sequence without explaining the rule that created it. Good work begins by classifying every source. ### 1. Law and regulation Law and regulation create binding duties in a defined jurisdiction and situation. They may govern authorization, clinical-trial conduct, informed consent, transparency, privacy, safety reporting, devices, labeling, accessibility, or submission. The writer records the jurisdiction, scope, status, and effective date and obtains legal or regulatory interpretation where needed. The writing consequence is often larger than a sentence-level requirement. A rule may determine which document must exist, who must approve it, when it must be submitted, what information must be included, how personal data are handled, and which public disclosures are required. ### 2. Regulatory guidance Regulatory guidance describes an authority's expectations and recommended approaches. It can shape study documents, evidence summaries, dossier modules, safety reports, public information, and submission interactions. Status matters: final and draft materials do not play the same role, and regional guidance may add to or interpret harmonized guidance. A writer should not reduce guidance to a contents list. Extract expectations about purpose, evidence, analysis, presentation, cross-references, appendices, lifecycle maintenance, and quality control. Record questions for the regulatory lead instead of converting ambiguous language into private certainty. ### 3. Consensus and technical standards Consensus standards give industries and disciplines a common technical language. The corpus repeatedly discusses ICH guidance in drug development and ISO-related standards in areas such as devices, quality, terminology, and plain-language practice. A standard can govern content, data, process, terminology, or performance. Access and copyright constraints may mean the project cannot reproduce a standard. The writer needs the controlled source or an authorized organizational interpretation, a clear applicability decision, and a checklist containing only what the team is entitled to operationalize. ### 4. Reporting guidelines Reporting guidelines help authors report a particular study design or evidence-synthesis method completely and transparently. Examples discussed throughout the corpus include CONSORT for randomized trials, PRISMA for systematic reviews and meta-analyses, STROBE for observational research, CARE for case reports, and statistical reporting recommendations such as SAMPL. Extensions address particular interventions, data sources, outcomes, populations, designs, or formats. These sources answer _what a reader needs to see to understand and appraise the work_. They do not repair poor methods, replace a protocol, turn a narrative search into a systematic review, or guarantee journal acceptance. ### 5. Publication ethics and good practice Publication practice includes authorship, contributor roles, professional medical-writing support, acknowledgments, conflicts of interest, funding, data integrity, originality, duplicate publication, trial registration, results disclosure, corrections, and interactions with journals. The corpus treats ICMJE recommendations, Good Publication Practice, COPE materials, journal policies, and author or sponsor procedures as an interlocking practice layer. This layer controls not only the manuscript's declarations but also the process used to create it. Authorship discussions, contribution records, writing support, approvals, and disclosure information must be managed before the submission screen asks for them. ### 6. Target instructions Journal, congress, authority, payer, registry, and portal instructions govern the destination. They specify article or submission type, section order, length, format, displays, references, supplementary files, declarations, forms, metadata, deadlines, and technical validation. Target instructions are project evidence. Preserve the version or capture used to plan the work. A remembered word limit or a template from another target is not sufficient. ### 7. Sponsor, client, and organizational procedures SOPs, work instructions, controlled terminology, data-handling rules, review systems, approval matrices, quality standards, and house style determine how the team works. These private requirements are not universal medical-writing rules. They are nevertheless controlling inside the project when valid and applicable. Record the owner, version, effective date, role implications, required records, and escalation route. Keep sensitive material inside approved systems. ### 8. Templates and schemas Templates, forms, data schemas, authoring tools, and submission structures express requirements in usable form. A template is a starting architecture, not permission to paste irrelevant boilerplate. The writer must know which fields are mandatory, which sections are conditional, which prompts are instructions rather than document text, and which content is inherited or generated elsewhere. ### 9. Clinical and professional guidance as evidence Treatment guidelines, consensus statements, position papers, and professional recommendations often support background, clinical context, comparator choice, or implications. They belong in the evidence dossier unless a target or process specifically makes them a governing requirement. This distinction prevents a common category error: a recommendation about clinical care does not automatically determine how a study must be reported. ## Build the Stack for a Live Project The stack is not a generic bibliography. It is a versioned set of decisions about what controls this document. ### Step 1: fix the identity of the output Write one line that names the document, subtype, evidence object, destination, jurisdiction, lifecycle stage, and primary audience. “Manuscript” is not specific enough. “Primary publication of a randomized phase III trial for a named journal” is much more useful. “Safety report” is not specific enough. Name the report, reporting period, product or study scope, and destination. ### Step 2: inventory candidate sources Search across the relevant layers: - law and regulation; - regional authority guidance; - harmonized or technical standards; - study-design reporting guidelines and extensions; - publication ethics and good practice; - target instructions and portal specifications; - organizational procedures and controlled templates; - clinical guidance needed as evidence. Use the [document-type deep dives](/guide/document-types) to identify the likely families, then verify the actual project. ### Step 3: test applicability For every candidate source, ask: 1. What kind of source is it? 2. Who owns or issues it? 3. Which jurisdiction, audience, product, study design, lifecycle stage, document, or submission does it cover? 4. Is it final, draft, superseded, transitional, or otherwise limited? 5. Which version and date apply to this project? 6. Does an extension or companion source apply? 7. Does the target explicitly require a declaration or checklist? 8. Who can confirm an uncertain interpretation? Mark the result _applies_, _does not apply_, _applies in part_, or _decision pending_. Add the reason. A list without applicability reasons becomes dangerous when copied to another project. ### Step 4: extract document impact Translate each applicable source into observable consequences: | Impact area | What to extract | | ------------------ | ------------------------------------------------------------------------------------------------------ | | Document existence | Whether the output is required and at what event or stage | | Structure | Mandatory or expected sections, order, headings, and cross-references | | Content | Required questions, data, explanations, declarations, and appendices | | Method | Search, analysis, appraisal, conduct, or user-testing expectations that must be reported | | Terminology | Controlled names, definitions, coding, units, and preferred expressions | | Presentation | Tables, figures, flow diagrams, precision, denominators, and formats | | Process | Roles, author input, independent review, approvals, reconciliation, and sign-off | | Metadata | Registry identifiers, product or study identifiers, dates, versions, contributor data, and disclosures | | Submission | File types, forms, portal fields, validation, and timing | | Traceability | Which claims, data, decisions, and changes must be recoverable | ### Step 5: resolve overlap and conflict Do not create one global hierarchy and assume it works everywhere. Classify the sources and resolve each overlap in context. 1. Confirm that both sources truly apply. 2. Separate a direct conflict from a difference in detail or purpose. 3. Identify the controlling legal, regulatory, contractual, or organizational owner within scope. 4. Check whether the more specific source validly adds to the broader source. 5. Surface the question to the responsible decision-maker. 6. Record the decision, rationale, owner, and source versions. 7. Update the outline, checklist, and review plan. > **Never solve a governance conflict silently in the prose.** ### Step 6: make a compliance matrix The smallest useful matrix has six columns: | Requirement | Source and location | Applies because | Document location | Verification | Owner/status | | ----------------------------------------- | ------------------- | --------------- | ----------------------------------------------- | ----------------------------- | ------------------------ | | A testable content or process requirement | Exact source anchor | Scope decision | Section, table, appendix, metadata, or workflow | How a reviewer can confirm it | Decision owner and state | Keep requirements atomic. “Comply with CONSORT” is not testable. “State how the random allocation sequence was generated in Methods” is. A good matrix drives the outline and later becomes a focused QC instrument. ### Step 7: preserve currency At project start and before final delivery, confirm the version, status, links, target instructions, template, and relevant procedures. Save what governed the work. Record a review trigger when the project will run long enough for requirements to change. The corpus is a learning source, not a currency service. Any named framework in this handbook is a route to verification, not proof that the project has the current version. ## Publication Ethics and Good Practice Publication ethics begins before drafting and continues after publication. It is the operating system for honest attribution, transparent support, fair reporting, and accountable correction. ### Authorship is a role with continuing duties The corpus's discussion of ICMJE authorship criteria centers authorship on substantial contribution, participation in drafting or critical revision, approval of the final work, and accountability. Authorship therefore cannot be determined reliably by title, seniority, funding alone, data collection alone, or a promised position in the byline. At kickoff: - identify proposed authors and the publication lead; - explain the applicable authorship and contribution expectations; - record expected contributions and decision roles; - establish how drafts, comments, approval, and accountability will be documented; - identify contributors whose work should be acknowledged rather than assigned authorship; - revisit the list as actual contributions develop. Authorship order follows the relevant discipline, team agreement, and target convention. The writer should facilitate a documented decision, not invent a universal ordering rule. ### Professional medical writers are transparent contributors A professional medical writer may help plan, organize, draft, edit, coordinate, document comments, manage references, and prepare submission materials. This support does not transfer author accountability to the writer and does not erase the authors' need to engage with the content. Record the writer's name or role, affiliation where applicable, nature of support, and funding source in the place required by good publication practice and the target. Ensure authors review, revise, approve, and can defend the publication. Hidden writing support weakens transparency even when the underlying work is accurate. ### Contribution records make the process visible A contribution taxonomy or target-specific contribution statement can show who conceived the work, designed methods, collected or curated data, performed analysis, provided resources, created software or visuals, supervised the work, drafted text, or critically reviewed it. Use only roles that reflect actual work. Keep the internal record sufficiently detailed to support the published declaration. Resolve discrepancies before submission. A form completed at the last minute cannot reconstruct months of unrecorded participation. ### Conflicts of interest and funding need complete collection Disclosure is broader than a single financial relationship. The exact scope comes from the target and applicable policy. Build a collection process that asks every author for current information, tracks completion, reconciles manuscript language with forms, and retains the evidence of approval. State funding and sponsor involvement accurately, including relevant roles in study design, data handling, analysis, interpretation, writing, and the decision to submit. Avoid both concealment and vague language that leaves the reader unable to understand the relationship. ### Originality is more than avoiding copied sentences Plagiarism includes presenting another's words, ideas, structure, data, or images without appropriate attribution. Self-plagiarism and redundant publication can mislead readers about novelty and distort the evidence base. Similarity tools can identify text overlap; they cannot decide whether reuse is legitimate, cited, permitted, or scientifically misleading. When material overlaps with an abstract, registry entry, protocol paper, prior analysis, preprint, conference presentation, or another manuscript: - disclose the relationship to the target where required; - cite related publications; - distinguish the new question, data, or analysis; - avoid salami slicing that fragments one coherent contribution merely to multiply outputs; - obtain permission for copyrighted material and label adaptation or reproduction accurately; - preserve the provenance of author-supplied text and figures. ### Data integrity outranks narrative convenience Do not select only favorable analyses, switch outcomes without explanation, hide protocol deviations, invent post-hoc hypotheses as prospective, suppress relevant harms, or imply causality beyond the design. Discuss inconsistencies with the statistician, study lead, authors, or governance owner and preserve the decision. Retractions, corrections, expressions of concern, and post-publication questions are part of the literature's correction mechanisms. If an error is found after submission or publication, assess its scientific effect quickly and follow the target's correction process. Defensive wording is not a substitute for correcting the record. ### Predatory or unsuitable targets create ethical and practical risk Assess whether a journal or congress has a credible scope, transparent editorial process, recognizable ownership, appropriate peer review, clear fees, indexing claims that can be verified, and policies for ethics and corrections. A persuasive invitation is not evidence of legitimacy. Match the audience and scientific contribution before convenience or speed. ### Publication-practice control card Before submission, verify: - every listed author meets the applicable criteria and has approved the final version; - contributor roles, acknowledgments, writing support, and funding are accurate; - conflicts and target forms are complete and consistent; - registration and protocol identifiers are correct; - primary and secondary outcomes are represented honestly; - related publications and prior presentations are disclosed; - permissions and licenses are documented; - the target's ethics declarations and checklist requirements are satisfied; - the submission package and cover letter do not overstate novelty or certainty. ## Study-Reporting Guidelines Reporting guidelines turn the reader's appraisal needs into a structured set of items. The writer uses them twice: first to build the document architecture, then to verify the completed report. ### Use a reporting guideline at outline stage Select the guideline from the actual design, not from the title the team hopes to use. Retrieve the main checklist, explanatory material, flow diagram where relevant, and any extension that fits the intervention, outcome, population, data source, or format. Map every applicable item to a section, display, supplement, or explicit not-applicable decision. Do not paste checklist language into the manuscript. Convert the requirement into a question the section must answer. For example, a requirement concerning participant flow becomes an instruction to define the populations, describe movement through the study, account for exclusions and losses, and reconcile the narrative with the flow diagram and analysis sets. ### Use it again at final review Complete the target's required checklist against the final pagination or section identifiers. Check the manuscript, supplement, registry, protocol, and referenced companion documents together. A page number beside an item proves only that the writer pointed somewhere; it does not prove the required information is complete. ### Randomized trials: CONSORT family For a randomized trial, readers need to understand the question, design, participants, interventions, outcomes, sample-size rationale, randomization, allocation concealment, blinding, statistical methods, participant flow, recruitment, baseline data, analyzed numbers, outcomes and estimation, harms, limitations, generalizability, interpretation, registration, protocol access, and funding. The writer should pay particular attention to distinctions that vague prose erases: - sequence generation versus allocation concealment; - who was blinded, to what, and how blinding was maintained; - prespecified versus changed or post-hoc outcomes; - enrolled, randomized, treated, completed, and analyzed populations; - effect estimates and uncertainty versus declarations based only on significance; - adverse-event reporting versus selective mention of expected events; - trial interpretation versus claims about routine practice or individuals. Use the flow diagram as a reconciliation instrument. Every count and reason should agree with the source data, tables, and text. ### Protocols: SPIRIT family A protocol reporting guideline helps make planned research understandable before results exist. It covers administrative information, rationale, objectives, design, participants, interventions, outcomes, timeline, sample size, recruitment, allocation, blinding, data collection and management, analysis, monitoring, harms, auditing, ethics, confidentiality, dissemination, access to data, and appendices such as consent materials. For a protocol manuscript, report the scientific plan and governance transparently. For an operational regulated protocol, applicable regulatory and organizational requirements also control the document. The reporting guideline does not replace the controlled protocol template or authority expectations. ### Observational research: STROBE family Observational reports must make design and selection visible enough for readers to judge bias, confounding, measurement, missing data, and generalizability. State the design early. Define setting, dates, eligibility, exposure, outcome, predictors, confounders, effect modifiers, data sources, measurement methods, efforts to address bias, study size, variable handling, and statistical methods. Report participant numbers across relevant stages, reasons for non-participation where available, descriptive characteristics, missingness, outcome data, estimates with precision, adjusted and unadjusted results when meaningful, category boundaries, sensitivity analyses, and limitations. Do not convert association into causation through verbs or headline claims. For routinely collected data and real-world evidence, document provenance, coding, linkage, quality, completeness, algorithm definitions, observation windows, and fit for the research purpose. Specialized extensions may add the details needed for databases, registries, or other data sources. ### Systematic reviews and meta-analyses: PRISMA family A systematic review is reproducible evidence research, not a long narrative with many citations. Report the question and eligibility criteria, information sources, complete search strategies where required, selection process, data collection, data items, risk-of-bias assessment, effect measures, synthesis methods, heterogeneity assessment, reporting-bias assessment, certainty evaluation where used, protocol or registration, amendments, support, competing interests, and data availability. The flow diagram must reconcile records identified, removed before screening, screened, excluded, sought, assessed, excluded with reasons, and included. Results should connect study characteristics, individual-study findings, risk of bias, syntheses, heterogeneity, sensitivity analyses, and certainty. The Discussion should distinguish the limitations of the evidence from the limitations of the review process. A meta-analysis adds statistical obligations; it does not automatically repair heterogeneity, bias, or incompatible outcomes. Explain model choices and transformations, show uncertainty, and avoid treating a pooled number as more credible than its contributing evidence. ### Case reports: CARE family A case report needs a clear clinical lesson supported by a transparent chronology. Describe patient information without unnecessary identifiers, clinical findings, timeline, diagnostic assessment, therapeutic intervention, follow-up and outcomes, adverse or unanticipated events, and the patient's perspective or consent where applicable. Explain why the case is noteworthy without claiming frequency, causality, or generalizability that one case cannot establish. Privacy is a design constraint. Removing a name may not de-identify a rare or richly described case. Follow the applicable consent, institutional, target, and legal requirements. ### Other designs and extensions Diagnostic-accuracy studies, prediction models, economic evaluations, quality-improvement work, qualitative research, animal or preclinical research, routinely collected data, pragmatic trials, cluster trials, noninferiority trials, harms, abstracts, and patient-reported outcomes may require dedicated guidance or extensions. The method remains: 1. identify the exact design and reporting claim; 2. retrieve the current main guideline and applicable extensions; 3. map items into the outline and displays; 4. document not-applicable decisions; 5. verify the final artifact and required checklist. > **Reporting checklist rule:** “Not reported” and “not applicable” are different conclusions. Both need conscious review. ## Statistical and Visual Reporting Statistics are part of the scientific argument, not a decorative language added after drafting. The writer must understand enough to report the analysis faithfully, detect mismatches, and ask the statistician precise questions. ### Start with the design and estimand-like question Before interpreting a number, ask what was compared, in whom, over what period, using which outcome definition, and under which handling of intercurrent events, missing data, and analysis population. The same numerical difference can answer different clinical questions under different analysis choices. Methods are the document's Rosetta stone: they define how the Results can be read. A statistically polished sentence cannot compensate for an unclear population, endpoint, model, or time point. ### Describe the data before testing it Use statistics that fit the variable and distribution. Means and standard deviations describe different features from medians and ranges or interquartile ranges. Counts need denominators. Percentages need clarity about the population, missing observations, and whether categories overlap. Avoid excessive decimal places that imply unsupported precision. Report units consistently. Explain transformations, derived endpoints, category thresholds, and composite outcomes. A table footnote should resolve interpretation, not hide essential methods. ### Report estimates and uncertainty A p-value does not show the size, direction, precision, or clinical relevance of an effect. Pair inferential results with the effect estimate and an interval measure where appropriate. State the exact p-value unless a justified threshold convention applies, and avoid “p = 0.000.” Distinguish a failure to reject a null hypothesis from proof of equivalence or absence of effect. Interpret confidence intervals in context. Wide intervals may include meaningfully different conclusions. Statistical significance can coexist with trivial clinical effect, and a clinically important estimate can be imprecise. Keep statistical and clinical interpretation connected but distinct. ### Name the analysis honestly Identify models and tests, variables included, adjustment factors, interaction terms, repeated-measure handling, multiplicity approaches, missing-data methods, sensitivity analyses, and software when relevant. Separate prespecified, supportive, exploratory, and post-hoc analyses. If an analysis changed after seeing the data, the prose should not imply prospective planning. In observational work, adjustment does not guarantee removal of confounding. In subgroup work, a significant result in one group and a nonsignificant result in another does not itself establish interaction. In time-to-event work, define the event, censoring, analysis time, and measures presented. In diagnostic work, provide the reference standard and the denominators behind sensitivity and specificity. ### Make harms visible Report the safety population, observation period, coding and collection approach, severity and seriousness, treatment relationship when assessed, withdrawals, deaths, and events of special interest as applicable. Use consistent denominators. Avoid comparing raw percentages without attention to exposure or study design. Keep zero-event statements within the observed sample and period. ### Tables and figures must answer reader questions Choose the display from the comparison or pattern the reader needs to see. - Use a table when exact values and multidimensional lookup matter. - Use a figure when pattern, distribution, change, relationship, or flow matters. - Use a flow diagram to explain movement through a process or study. - Use a forest plot to show effects and uncertainty across analyses or studies. - Use a survival plot only with clear risk sets, time origin, censoring context, and interpretable groups. Every display needs a purpose, standalone title or legend, defined abbreviations, units, analysis population, denominators where needed, and traceable source. Axes and scales must not create a false impression. Color must carry meaning consistently and not be the sole way to distinguish critical information. ### Statistical QC pass Check: - Methods name every analysis needed to understand the reported results; - populations, endpoints, time points, models, and estimands or comparison questions align; - counts, denominators, percentages, effect estimates, intervals, and p-values match the validated source; - precision and rounding are consistent and do not change interpretation; - text, tables, figures, abstract, and conclusion agree; - subgroup, sensitivity, exploratory, and post-hoc analyses are labeled; - limitations reflect design, conduct, missing data, multiplicity, power, bias, and generalizability where relevant; - the narrative does not convert association into causation or nonsignificance into equivalence; - every visual can be interpreted without deceptive scale, missing labels, or hidden denominator changes. ## Clinical and Regulatory Writing Regulatory documents form a connected evidence system across development and the product lifecycle. Their governing material comes from harmonized guidance, regional requirements, authority expectations, sponsor procedures, controlled templates, and the approved data. The writer's task is to make the system internally consistent and reviewable. ### Read ICH guidance by function The corpus organizes ICH material into broad families that help a writer search: - **Quality** guidance concerns the chemical, pharmaceutical, and manufacturing dimensions of development and lifecycle control. - **Safety** guidance concerns nonclinical evidence and safety evaluation. - **Efficacy** guidance includes clinical study conduct, design, statistics, reports, populations, safety, and related development questions. - **Multidisciplinary** guidance includes cross-cutting topics such as the Common Technical Document, terminology, data standards, and electronic information. The family name is only a locator. For live work, determine the exact topic, current document, regional implementation, effective status, and sponsor interpretation. A summary in a book or journal can teach context but cannot replace the controlled source. ### Understand the core document relationships A clinical-development writing system can be read as a chain: ```text development question ↓ protocol + amendments ──→ statistical analysis plan ↓ ↓ study conduct ───────────→ validated results and listings ↓ ↓ clinical study report + appendices ↓ investigator information, summaries, overviews, safety reports, disclosure, and publications ``` The documents have different purposes, readers, and levels of detail, but factual identity must persist. Study identifiers, objectives, design, populations, interventions, endpoints, analysis methods, dates, results, deviations, and conclusions should not mutate as information moves through the chain. Differences must be attributable to purpose, timing, scope, or approved interpretation. ### Protocol and amendments The protocol controls the planned conduct of a study. Its structure commonly includes administrative information, background and rationale, objectives and endpoints, design, participants, treatments or interventions, assessments, safety, statistics, data handling, monitoring, ethics, and supporting material. The protocol must be operational enough for sites and functions to act consistently and precise enough for later reports to distinguish plan from conduct. Use applicable protocol guidance, Good Clinical Practice expectations, regional requirements, organizational procedures, and the controlled template. Check that: - objectives, estimands or endpoint questions, endpoints, time points, and analyses form a coherent chain; - eligibility criteria match the intended population; - treatment, dose, schedule, comparators, rescue therapy, discontinuation, and follow-up are unambiguous; - schedules and prose agree; - safety collection, escalation, stopping, and oversight responsibilities are defined; - randomization, blinding, sample size, analysis sets, missing data, and interim analyses are explained at the right level; - data, records, monitoring, ethics, consent, publication, and confidentiality provisions are present where applicable; - terms and abbreviations remain stable across the synopsis, main text, schedules, appendices, and related materials. An amendment changes the controlled plan. State what changes, why, when it applies, and which related materials or systems require alignment. Do not let an amendment create a document whose synopsis, body, schedule, consent materials, registry, and operational instructions disagree. ### Statistical analysis plan The statistical analysis plan translates the protocol's questions into executable analysis detail. It defines analysis populations, endpoints and derivations, models, covariates, intercurrent-event handling, missing data, multiplicity, interim work, sensitivity and supplementary analyses, table/figure/listing shells, and decision rules as appropriate. The writer and statistician reconcile the plan with the protocol and mock outputs. Terms, analysis-set definitions, visit windows, censoring, baseline definitions, and subgroup categories should have one controlled meaning. Changes after unblinding or data access need transparent governance and description in the report. ### Clinical study report and ICH E3 The corpus repeatedly presents ICH E3 as the principal structural reference for clinical study reports. A CSR integrates protocol, conduct, analysis, results, safety, deviations, interpretation, and supporting appendices for regulatory review. It is not a manuscript expanded by adding tables; it is a detailed account designed for traceability and authority assessment. At planning stage, map the controlled template and project requirements to the E3 structure. Build the report from the final protocol and amendments, SAP, validated tables/figures/listings, data review and reconciliation records, and study decisions. Keep data-dependent text anchored to validated outputs. High-value controls include: - reconcile planned and actual study conduct; - explain protocol and analysis deviations without burying their effect; - distinguish disposition, exposure, efficacy, and safety populations; - align endpoint definitions, time points, and statistical methods with reported outputs; - report efficacy and safety comprehensively, including findings that do not favor the product; - integrate tables and figures that answer regulatory questions while preserving source traceability; - keep the synopsis a faithful compressed report rather than a separate optimistic narrative; - ensure appendices, listings, investigator information, signatures, and required artifacts are complete; - run cross-document checks against protocol, SAP, registry and disclosure information, and relevant development documents. The conclusion should follow the evidence and design. It should not broaden the population, indication, dose, duration, comparator, endpoint, or certainty beyond what the study supports. ### Investigator's brochure The investigator's brochure gives investigators and others responsible for trial conduct the clinical and nonclinical information needed to understand the investigational product and manage the study. It evolves as knowledge changes. Applicable Good Clinical Practice and sponsor procedures govern content, review, approval, distribution, and update. Create a source map across physical, chemical, pharmaceutical, nonclinical, pharmacokinetic, clinical, and safety evidence as relevant. Present a balanced summary of known and potential risks, adverse reactions, precautions, monitoring, and management. Reconcile the brochure with the current development safety profile and other controlled materials. Date-lock the data cut and record new information awaiting the next update. ### Common Technical Document summaries and overviews The CTD gives a common modular architecture for marketing-authorization material. Writers working in clinical and nonclinical summaries and overviews integrate evidence across studies rather than copy individual report conclusions. Regional administrative material and product information require separate attention, while quality, nonclinical, and clinical evidence occupy their relevant modules. An overview is an expert argument about the program. A summary is a structured, sufficiently detailed presentation of evidence. Both require: - a complete study and source inventory; - stable product, indication, population, dose, comparator, endpoint, and terminology conventions; - evidence tables that allow studies to be compared; - explicit handling of study limitations, missing evidence, conflicting findings, and data cutoffs; - benefit-risk reasoning that connects magnitude, uncertainty, harms, unmet need, and the proposed use; - cross-references that lead reviewers to the supporting location; - consistency with labeling or proposed product information and with regional submission strategy. Writers should distinguish evidence generated for the program from contextual literature. They should also distinguish an integrated conclusion from a recital of every study. Integration asks what the body of evidence means and how robustly it supports the requested decision. ### Regulatory response documents and briefing materials Authority questions, deficiency responses, meeting requests, briefing packages, and other correspondence need issue-based structure. Repeat or identify the question accurately, give a direct response, explain the evidence and reasoning, cite the exact supporting locations, and state proposed actions or commitments precisely. Build a response matrix that tracks question ownership, interpretation, contributing functions, source data, draft, review status, dependencies, commitments, and cross-response consistency. Do not answer a narrow question with a data dump. Do not hide an unfavorable answer in background. If the requested evidence does not exist, state the limitation and the team's governed response. ### Regulatory quality card Before handoff, verify: - the current governing sources, regional requirements, procedures, and template are recorded; - source data and documents are approved for the intended use; - identifiers, data cutoffs, populations, endpoints, study facts, and terminology reconcile across the submission; - required sections, tables, appendices, cross-references, and metadata are complete; - analyses and interpretations have the correct functional owners; - benefit and risk are represented fairly; - open issues, assumptions, deviations, and commitments are visible; - the package passed scientific, statistical, regulatory, editorial, and technical checks appropriate to its risk. ## Safety and Pharmacovigilance Safety writing converts individual cases and aggregate data into information that supports detection, evaluation, minimization, and communication of risk. Applicable law, authority guidance, controlled terminology, product reference information, safety-system procedures, privacy rules, templates, data-lock processes, and medical governance all matter. ### Individual case narratives A case narrative reconstructs a clinically meaningful chronology from source data. It typically connects patient characteristics, medical history, indication, exposure, event onset and course, investigations, treatment, dechallenge or rechallenge, outcome, reporter or investigator assessment, company assessment, confounders, and follow-up as applicable. The narrative should be selective without being selective _about_ risk. Include facts needed to understand seriousness, expectedness, causality, clinical course, and alternative explanations. Do not fill gaps with plausible detail. Mark unknowns according to procedure and route inconsistencies for data clarification. Control points: - use the correct case, product, event, date, and coding identities; - preserve chronology and distinguish report dates from clinical dates; - reconcile structured fields, line listings, source documents, and narrative; - separate reporter opinion, investigator opinion, sponsor assessment, and writer inference; - protect personal information; - handle follow-up and version changes through the controlled safety system; - use medical review for clinical interpretation. ### Aggregate and periodic reports Aggregate safety documents integrate interval data, cumulative exposure, case patterns, studies, literature, regulatory actions, signals, risks, effectiveness of risk minimization, and benefit-risk evaluation according to the report type. The writer needs a source and responsibility matrix because contributions arrive from multiple functions and data cuts. State the reporting interval and data lock clearly. Reconcile denominators and exposure estimates. Separate new interval information from cumulative knowledge. Explain changes in signals and important identified or potential risks. Ensure actions and conclusions follow the evidence and are consistent with risk-management and reference safety information. The summary must not conceal complexity that changes the benefit-risk view. Tables and appendices should support navigation and verification rather than act as an alternative to interpretation. ### Risk-management documents Risk-management writing links the safety specification to pharmacovigilance activities and risk-minimization measures. Each important risk or missing-information category needs a clear rationale, evidence base, planned activity, objective, implementation, and evaluation approach as applicable. Maintain alignment across safety reports, product information, investigator materials, educational materials, and regulatory commitments. When the classification or understanding of a risk changes, trace the effect across the document ecosystem. ### Safety-writing QC card Check case counts, coding, event and product terms, seriousness, expectedness, causality language, dates, exposure, intervals, data cutoffs, reference information, signal status, risk classification, actions, and cross-document consistency. Run a dedicated privacy pass. Safety prose requires scientific clarity, but controlled data identity and procedural compliance come first. ## Medical Devices and IVDs Device and in vitro diagnostic writing connects the device's intended purpose, design and technology, clinical or performance evidence, risk management, usability, post-market information, and lifecycle claims. The corpus emphasizes that the governing environment differs by jurisdiction, device or IVD status, risk class, lifecycle stage, and evidence route. ### Classify before writing Record: - whether the product is a medical device or an IVD; - intended purpose and indications; - target population and intended user; - use environment and operating principle; - claims and clinical benefit; - jurisdiction and classification or risk category; - development, conformity-assessment, market-entry, or post-market stage; - applicable regulations, guidance, standards, quality procedures, and templates. The classification decision is not a writer's casual judgment. It determines the evidence and document suite and should come from the responsible regulatory and technical owners. ### Clinical evaluation for devices Clinical evaluation is a planned, continuous assessment of clinical data relevant to safety and performance. It can integrate literature, investigations, equivalence evidence where justified, post-market surveillance, complaints, vigilance, and other product data. The report must connect the intended purpose and claims to the state of the art, identified evidence, appraisal, analysis, residual risks, and clinical benefit. High-risk areas include an underdeveloped search, weak appraisal, unjustified equivalence, reliance on irrelevant predecessor data, incomplete state-of-the-art context, untraceable claims, and conclusions that exceed the evidence. Treat literature work as reproducible evidence work and preserve searches, selection decisions, appraisal, and data extraction. ### Performance evaluation for IVDs IVD performance evaluation considers the scientific validity of the analyte-marker relationship, analytical performance, and clinical performance in relation to intended purpose. The writer should make the bridge among these evidence streams explicit and keep specimen type, target population, user, setting, comparator or reference method, performance characteristics, limitations, and claims consistent. Diagnostic metrics require their denominators and population context. Sensitivity, specificity, predictive values, agreement, precision, limits, and invalid or indeterminate results answer different questions. Prevalence can change predictive values. Avoid presenting a single attractive metric as total performance. ### Clinical investigations and performance studies Plans and reports need clear objectives, design, sites, participants or specimens, device handling, comparators, endpoints, analysis, adverse events or deficiencies, deviations, and ethical and data-management controls as applicable. Reconcile investigation documents with risk management, instructions for use, and the intended claims. ### Post-market clinical follow-up and surveillance Post-market evidence confirms performance in use, identifies emerging risks, follows residual uncertainties, and supports lifecycle updates. Plans should start from defined evidence gaps and claims, not generic activities. Reports should show what was done, what was found, what it means for benefit-risk and state of the art, and what actions follow. Maintain the line of sight among surveillance, vigilance, complaints, literature, trend reporting, risk-management files, clinical or performance evaluation, public summaries, and labeling. A new signal in one document should not remain invisible in another. ### Device-writing QC card Verify intended purpose, claims, classification, device identity, variants, comparator or equivalent device, state of the art, search dates, evidence tables, appraisal, risk terminology, clinical benefit, performance measures, residual risks, post-market findings, and conclusions across the suite. Confirm that every claimed performance or benefit has traceable evidence appropriate to the device and jurisdiction. ## HEOR, HTA, and Real-World Evidence Health economics and outcomes research asks how interventions affect outcomes, resource use, costs, value, and decisions in real practice or modeled settings. Health technology assessment places clinical, economic, organizational, ethical, and social evidence into a jurisdictional decision process. Real-world evidence uses data collected outside conventional controlled-trial contexts to answer defined questions. ### Decision context controls the document Before drafting, define the decision-maker, population, intervention, comparators, perspective, jurisdiction, time horizon, outcome measures, evidence sources, model type, willingness-to-pay or decision framework where relevant, and submission instructions. A globally persuasive value story may still fail if it does not answer a local payer's specified question. ### Economic evaluations Report the type of evaluation and its analytical perspective. Explain comparators, time horizon, discounting, outcomes, resource use, cost sources and price year, currency and conversions, model structure, assumptions, input selection, extrapolation, uncertainty, validation, heterogeneity, distributional considerations where applicable, and limitations. Keep the base case distinct from scenarios and sensitivity analyses. Show which conclusions depend on uncertain assumptions. A model is an explicit representation of a decision problem, not observed reality. Avoid giving projected values the rhetorical status of measured facts. ### Budget-impact work Budget-impact analysis focuses on affordability and financial consequences for a defined payer or system over a defined period. Describe eligible population, uptake, displacement, treatment duration, resource consequences, costs, scenarios, and uncertainty. Keep clinical value and short-term budget effect conceptually distinct. ### Real-world data and evidence Define the data source, provenance, collection purpose, linkage, population coverage, observation period, coding, completeness, validation, and fitness for use. Make cohort construction reproducible. Explain exposure, outcome, confounder, and covariate algorithms; index dates and windows; handling of duplicates and missingness; and bias-control methods. Common risks include immortal-time bias, confounding by indication, selection bias, outcome misclassification, incomplete capture, data drift, inconsistent coding, and causal language unsupported by design. Report sensitivity analyses and negative or alternative controls when part of the method, and state residual uncertainty plainly. ### HTA and payer submissions Follow the current jurisdictional methods and submission template. These may govern comparator, evidence hierarchy, indirect comparison, economic model, population subgroup, uncertainty, dossier format, stakeholder input, and procedural timing. Use a requirement matrix to connect every requested item to the evidence and document location. Clinical, economic, and patient evidence should form one decision narrative without being blended into one kind of claim. Patient-reported outcomes and preference evidence need their own methods, instruments, interpretation thresholds, and limitations. ### HEOR quality card Check that the decision problem, population, comparators, perspective, horizon, data sources, model, assumptions, inputs, costs, outcomes, uncertainty, validation, and limitations are transparent. Every number should be classifiable as observed, derived, assumed, modeled, or sourced from literature. Reconcile narrative claims with evidence tables, model outputs, and the submitted base case. ## Disclosure and Transparency Disclosure makes study existence, methods, results, documents, and relationships visible to the people entitled or expected to see them. The corpus discusses trial registration, results posting, public disclosure of clinical documents, lay summaries, open science, data sharing, and publication transparency as related but distinct practices. ### Build one disclosure identity Maintain a cross-output record of the study title, identifiers, sponsor, design, countries, populations, interventions, comparators, objectives, endpoints, milestones, status, results, protocol versions, analysis populations, and contact information as applicable. Registry entries, protocols, CSRs, public summaries, journal articles, and congress materials have different purposes, but they should not tell incompatible factual stories. ### Registration and results posting Identify the applicable registries, timing, data fields, result requirements, responsible owner, and update process. Compare structured fields with controlled source documents. Pay particular attention to outcome definitions, time frames, enrollment, arms, analysis populations, adverse events, completion dates, and publication references. Do not wait until posting to discover that the registry endpoint no longer matches the protocol or publication. Treat registry reconciliation as part of study-document governance. ### Public disclosure of clinical documents Public disclosure can require redaction, anonymization, de-identification, and management of confidential commercial information while preserving scientific usefulness. This is a multidisciplinary process involving privacy, legal, regulatory, clinical, statistics, and disclosure expertise. The writer should maintain meaning across the disclosed artifact, redaction decisions, anonymization changes, and related public information. Excessive removal can make the evidence unintelligible; insufficient protection can expose personal or legitimately confidential information. Use the governing process and qualified reviewers. ### Transparency QC card Verify that: - the correct disclosure duties, jurisdictions, registries, documents, and timelines are identified; - public facts reconcile with controlled study sources; - outcome and result changes are explained through the authorized process; - authorship, funding, conflicts, writing support, registration, and data-sharing statements agree across materials; - privacy and confidential-information controls have specialist approval; - plain-language content remains balanced after redaction or simplification; - links and identifiers lead readers to the intended records. ## Plain Language, Readability, and Accessibility Plain language enables the intended reader to find, understand, and use information. It is not childish language, a shorter technical abstract, or a readability score. In medicine, plain language must preserve evidence, uncertainty, risk, and the boundary between general information and individual advice. ### Begin with the reader's decision Define the audience more precisely than “the public.” A trial participant, person considering consent, patient living with a condition, caregiver, journalist, policy audience, and reader of a publication summary have different questions and consequences. Ask: - What does this reader already know? - What do they need to decide, understand, remember, or do? - What may frighten, confuse, stigmatize, or falsely reassure them? - Which languages, literacy levels, disabilities, devices, and settings shape access? - Which information is essential for a safe action? - How will understanding be tested with intended users? Organize around those questions. Put the conclusion or action close to the information that supports it. Build navigation that lets a reader find purpose, benefits, risks, choices, practical steps, uncertainty, and sources. ### Translate meaning, not terminology alone For each technical idea, decide whether to: - use an everyday term; - use the technical term after a plain explanation; - define it in place; - show it with an example or visual; - remove it because it does not serve the reader's task. Keep necessary medical names where they help readers talk with clinicians or locate reliable information. Define them on first use and keep the same term thereafter. Avoid rotating through synonyms merely for style; in patient material, consistency supports comprehension. Use familiar words and concrete verbs. Prefer “take” to “administer” when the meaning allows, and “high blood pressure” with or before “hypertension” when the audience needs both. Explain nominalized phrases as actions. Replace vague references such as “this may occur” with the event that may occur. ### Control sentence and paragraph load Shorter sentences can help, but sentence length is only one part of difficulty. A short sentence can still contain an unfamiliar concept, hidden condition, ambiguous pronoun, or dense number. Keep one main idea per sentence when possible. Place conditions next to the action they control. Use active voice when the actor matters and passive voice when the actor is unknown, irrelevant, or appropriately backgrounded. Use headings that answer reader questions. Keep paragraphs focused. Lists help with steps, choices, symptoms, or parallel facts; they do not automatically clarify a complicated argument. A document with many bullets can still hide priority and causality. ### Explain numbers as decisions, not ornaments Many readers struggle with probability, percentages, ratios, and changing denominators. Good risk communication: - uses a consistent denominator when comparing groups; - gives natural frequencies where useful, such as “3 out of 100”; - distinguishes absolute and relative change; - states the time period; - identifies the population; - presents benefit and harm with comparable framing; - avoids switching between survival and mortality frames to make one option look better; - explains uncertainty and the limits of prediction for an individual. A change from 1 in 100 to 2 in 100 is both a one-percentage-point absolute increase and a doubling in relative terms. Choosing only the relative description can exaggerate the practical impression. Give the reader enough context to understand both magnitude and uncertainty. ### Write benefits and harms in fair balance Use parallel prominence, detail, time horizon, and numerical framing. Do not describe benefit vividly and risk vaguely, or put limitations where readers are unlikely to see them. Separate what happened in the study population from what may happen to an individual. Terms such as _common_, _rare_, _serious_, _mild_, _positive_, and _significant_ can carry technical and everyday meanings. Explain the intended meaning and, when possible, accompany frequency labels with numbers. ### Readability formulas are screens, not verdicts Readability formulas usually respond to visible features such as sentence length and word length. They can flag dense text and help compare drafts, but they do not measure scientific accuracy, organization, visual usability, cultural appropriateness, emotional response, or actual comprehension. A low score can be gamed by replacing precise terms or chopping logic into fragments. Use a formula as one diagnostic. Then review vocabulary, concepts, navigation, actionability, numbers, visuals, tone, and intended-user performance. ### Design carries meaning Typography, spacing, line length, contrast, hierarchy, white space, and responsive behavior affect whether content can be used. Headings and summaries should expose the route through the information. Tables need plain labels and should not force mobile readers to compare distant columns. Icons require labels when their meaning is not universal. Accessibility includes more than visual design. Plan for semantic headings, keyboard navigation, meaningful link text, text alternatives for informative images, captions or transcripts for media, sufficient contrast, scalable text, and information that does not depend on color alone when the eventual web implementation is built. These details must be checked in the implemented page, not assumed from the Markdown. ### Visuals must clarify the evidence Choose a visual for a defined reader question. Use human-scale denominators and comparable icons when showing frequencies. Use timelines for study flow or treatment steps. Label axes, groups, time, units, and uncertainty. Do not use decorative area, truncated axes, emotionally loaded imagery, or unequal icon arrays to dramatize an effect. Test whether the intended audience interprets the visual correctly without the writer explaining it. A beautiful graphic that generates the wrong conclusion is a writing defect. ### User testing closes the loop Ask intended users to find and explain critical information and describe what they would do. Observe navigation and hesitation. Do not ask only whether they “like” the document. Test high-risk content such as dosing, consent choices, symptoms requiring action, benefit-risk comparisons, withdrawal, contact routes, and the distinction between group evidence and individual care. Revise and retest. Document the audience, tasks, findings, changes, and unresolved limitations. Expert review and readability scoring cannot substitute for intended-user evidence. ### Plain-language QC card Check: - audience and decision are explicit; - essential information appears early and is easy to find; - technical concepts are explained without changing meaning; - benefit, harm, limitation, and uncertainty receive fair balance; - numbers use consistent denominators, populations, and time frames; - every action states who should do what, when, and how; - headings, lists, tables, links, and visuals support navigation; - tone is respectful and avoids blame, stigma, false reassurance, or alarm; - general information is not framed as individual medical advice; - intended users have tested the high-risk content where the project permits. ## Translation and Localization Medical translation is a controlled transfer of meaning into another language and culture. It must preserve scientific identity, instructions, risks, uncertainty, and legal or regulatory function while producing text that the target audience can understand and use. ### Prepare the source before translation Ambiguous source language produces expensive multilingual ambiguity. Finalize and control the source as far as the process allows. Resolve abbreviations, undefined terms, inconsistent product names, unclear pronouns, mixed units, and unstable phrasing. Mark text that must remain verbatim or use controlled terminology. Give the translator: - document purpose and status; - audience, literacy context, and country or locale; - source files with layout and visuals; - reference documents and approved terminology; - product, condition, and study context; - a glossary and style instructions; - contact routes for questions; - version and change history. ### Use qualified people and defined roles Medical and regulatory content needs linguistic skill, domain knowledge, and understanding of the intended reader. Define translator, reviewer, medical reviewer, in-country reviewer, patient or user tester, project manager, and approver roles as appropriate. Independence between translation and review can expose errors that familiarity misses. ### Back translation is one tool Back translation can reveal shifts in meaning, omissions, or additions, especially in instruments and high-risk content. It can also reward literal language that reads poorly in the target language. Use it within a broader process that includes reconciliation, target-language expert review, cognitive or user testing when relevant, and final-layout QC. ### Localize deliberately Check differences in clinical practice, health-system routes, emergency numbers, product names, dose forms, calendars, decimal conventions, measurement units, personal names, examples, images, symbols, family roles, and cultural assumptions. Adapt only through the approved process; localization should not silently alter scientific or legal meaning. ### Control versions and layout Maintain a language matrix with source version, translation version, locale, status, translator, reviewer, approval, and implementation location. When source text changes, identify which languages and artifacts are affected and translate the change in context. Run numeric and linguistic checks after layout. Text expansion can break tables, hide warnings, separate qualifiers, or push labels away from visuals. Right-to-left scripts and non-Latin characters require implementation testing. Compare the final rendered artifact, not only the translation memory or bilingual file. ### Translation QC card Verify names, identifiers, numbers, units, ranges, dates, time, dosage, frequency, route, negation, conditional language, risk terms, headings, cross-references, links, legends, footnotes, images, and actions. Confirm that the correct source version was translated and that target-language users can understand critical information. ## AI-Assisted Medical Writing The corpus presents AI as useful but fallible support within a controlled workflow. Generative systems can help with ideation, search support, extraction, classification, outlining, transformation, editing, summarization, and quality checks. They can also fabricate facts and citations, leak confidential material, reproduce bias, lose context, obscure provenance, and produce fluent inconsistency. Safe use begins with a governed document task and workflow. ### AI literacy for medical writers A writer does not need to become a machine-learning engineer, but should understand that model output is generated from patterns and context rather than guaranteed retrieval of truth. Output varies with prompts, settings, model, tool configuration, and supplied material. A confident tone is not an evidence signal. Before relying on an AI-enabled workflow, establish whether it: - works from authorized source material rather than unspecified background information; - retrieves and exposes the source passages used for the task; - applies deterministic checks where a repeatable rule is appropriate; and - operates with the access, logging, retention, security, and review controls required for the document. Retrieval can reduce unsupported answers and improve source grounding, but retrieval itself can miss, mis-rank, or supply the wrong passage. Verification remains human work. ### Define the use case and workflow first Write down: - the exact task; - the permitted inputs; - the expected output; - who will review it; - which error types are plausible; - what harm an error or disclosure could cause; - what evidence or test will verify the result; - whether the system and workflow are approved. “Help write the CSR” is too broad. “Compare the abbreviation list with defined terms in this approved, non-confidential test document and return possible mismatches for human review” is bounded and testable. ### Risk-assess data and task Classify the information before entering it. Patient information, personal data, confidential study data, unpublished results, product strategy, privileged communications, licensed content, and controlled documents may be prohibited or require an approved environment. Follow policy, contracts, consent, privacy, security, intellectual-property, and records-management requirements. Then classify the consequence of output error. Brainstorming neutral headings is lower risk than generating a benefit-risk conclusion, safety narrative, statistical interpretation, medical advice, consent language, or regulatory response. High-consequence tasks need stronger controls and may be unsuitable for generation. A practical risk screen asks: | Dimension | Lower-risk pattern | Higher-risk pattern | | ----------------- | --------------------------------------- | --------------------------------------------------------------------------------- | | Input | Public, non-sensitive, authorized text | Personal, confidential, proprietary, unpublished, or controlled information | | Output use | Internal idea or candidate for review | Final evidence, clinical advice, regulated content, or external decision material | | Verifiability | Deterministic or easy source comparison | Judgmental, predictive, or difficult to trace | | Error consequence | Minor editing rework | Patient, scientific, legal, regulatory, privacy, or reputational harm | | Human control | Qualified reviewer with source access | Automated acceptance or reviewer unable to verify | ### Evaluate the system and workflow, not only the output Ask about intended use, model or service identity, access controls, data retention, model training on inputs, location and subprocessors where relevant, logging, reproducibility, version changes, source citation, retrieval boundaries, export, validation, incident handling, and organizational approval. The workflow must match the source material, destination, and risk of the document; work involving a controlled submission dossier needs the corresponding controls. Performance claims should be tested on representative tasks. Create a benchmark set with known answers and meaningful failure cases. Measure factual accuracy, completeness, citation validity, sensitivity to instruction changes, consistency, privacy behavior, usability, and the time required for expert verification. Compare total controlled effort, not raw generation speed. ### Ground the task in authorized sources Provide a closed source set when the task requires source-based work. Tell the model to distinguish supplied evidence from inference, identify missing support, preserve identifiers and numbers, and point to source locations. Ask for structured output that exposes verification fields. For example: ```text Task: extract candidate endpoint definitions from the supplied approved protocol. Rules: use only the supplied text; do not complete missing information; preserve wording for names, time points, and units. Output: endpoint name | definition | time point | source section | ambiguity for reviewer. ``` This does not make the output correct. It makes errors easier to find. ### Verify every output according to its risk For source-based content, compare every claim, number, citation, quote, identifier, and conclusion with the source. Open cited references; generated citations can look plausible while being nonexistent or mismatched. Check whether a true source actually supports the claim in context. For edits, use tracked comparison or a diff. Look specifically for altered meaning, lost negation, changed time, softened uncertainty, broadened population, swapped numerator or denominator, expanded causality, and inconsistent terminology. For summaries, compare omissions as carefully as included facts. For code, calculations, or structured transformations, run independent tests on expected, boundary, and adversarial cases. For classification, review false positives and false negatives. For repeated production, monitor drift after model or configuration changes. ### Keep a human decision gate Qualified humans remain responsible for scientific interpretation, medical judgment, statistical meaning, ethical decisions, authorship, regulatory strategy, benefit-risk conclusions, patient-facing advice, and final approval. The reviewer needs the source, task, tool context, and enough time to perform genuine verification. Automation bias makes fluent output feel finished. Require reviewers to record material corrections, unresolved uncertainty, and acceptance. A signature without a source-based review is not oversight. ### Be transparent about AI involvement Follow target, organizational, client, and applicable policy for disclosure. AI systems do not meet human authorship or accountability requirements. Humans remain responsible for originality, accuracy, attribution, confidentiality, permissions, and the final document. Maintain an internal use record proportionate to risk: task, system and version where available, date, operator, data class, source set, instruction or workflow identifier, output location, reviewer, verification, and disposition. Preserve it only in approved systems and according to records requirements. ### AI use patterns Potentially useful patterns, subject to approval and control, include: - generating neutral questions for an outline; - classifying documents for a human-reviewed evidence inventory; - extracting candidate entities with source anchors; - comparing terminology, abbreviations, or formatting; - proposing plain-language alternatives that are checked against the source and user-tested; - locating internal inconsistencies for human confirmation; - creating test data or examples that contain no real confidential information; - converting approved content into a structured draft that undergoes full review. High-risk patterns include unsupervised literature conclusions, invented references, direct use of patient or confidential data in an unapproved tool, autonomous medical advice, automatic authorship decisions, generation of final safety or benefit-risk conclusions, and acceptance of rewritten numbers without source comparison. ### AI quality card Before use: - the tool, task, data, and environment are approved; - the intended output and forbidden uses are explicit; - risks and reviewer qualifications are recorded; - the authorized source set and verification method exist. Before acceptance: - every material fact, number, citation, quote, and inference has been checked; - omissions, bias, causality, uncertainty, and audience harm have been reviewed; - confidential or personal data were handled correctly; - output is original or appropriately attributed and permitted; - human contributors and approvers understand and accept responsibility; - required use records and disclosures are complete. ## Templates, SOPs, and House Style Templates, procedures, and style controls turn general expectations into repeatable project work. Use them as governed instruments, not as inherited text that escapes scrutiny. ### Read a template in four layers 1. **Fixed structure:** required headings, fields, metadata, and document architecture. 2. **Conditional structure:** sections that appear only for certain designs, products, regions, or outcomes. 3. **Instructions:** prompts, examples, comments, and authoring notes that should not remain in the deliverable. 4. **Reusable content:** approved boilerplate or standard language whose applicability and current status must still be checked. At setup, cleanly distinguish these layers. Track deviations from the template and obtain the required approval. A section can be formally present yet substantively empty; QC must test the requirement, not the heading. ### Convert procedures into role and evidence maps For each applicable SOP or work instruction, extract triggers, inputs, roles, review steps, approvals, required records, systems, timelines, escalation, and archival duties. Add these to the project plan and definition of done. Do not reproduce private procedures in a public guide. Link or reference the controlled source inside the authorized project environment and preserve version identity. ### Use house style to protect consistency House style controls spelling, capitalization, punctuation, abbreviations, units, numerals, drug and device names, gene and protein nomenclature, tables, references, and document conventions. Create a project style sheet for decisions not already owned by a controlling guide. Scientific fidelity outranks cosmetic consistency. Do not “correct” a validated endpoint name, quoted scale, controlled term, or cited article title merely to match general style. Record the exception. ### Make automation visible Reference managers, macros, scripts, content-management systems, structured authoring, and automated checks can improve consistency. Validate them for their intended use and retain human review. Automation can spread one error across an entire suite faster than manual work. ## Turn Any Governing Source into a Working Tool Use this page pattern when the web implementation opens a dedicated guideline or standard window. ### Identity Record full title, common name, issuing body, source type, version, date, status, jurisdiction, official location, project copy, and last verification. ### Scope State the documents, study designs, products, populations, lifecycle stages, audiences, and situations it governs. State common adjacent situations it does not govern. ### Applicability test Write a small set of yes/no or conditional questions that a project lead can answer. Include extensions, transitional status, regional adoption, target mandates, and specialist confirmation. ### Thirty-second gist Summarize the source in five or fewer operational statements without pretending the summary replaces it. ### Document impact Map atomic requirements into structure, content, method, terminology, data presentation, declarations, appendices, metadata, workflow, and submission. ### Dossier impact List the evidence and metadata needed to satisfy or demonstrate each expectation. Include source anchors and owners. ### Verification Turn the source into focused questions that can be checked against the final artifact. Keep interpretation separate from quoted or closely paraphrased source requirements. ### Relationships Identify parent, companion, regional, target, procedural, template, and reporting sources. Explain overlap and known decision points without copying the same rule onto every page. ### Worked application Trace one requirement from source to applicability decision, outline instruction, drafted content, source anchor, review, and QC evidence. A worked trace teaches more than an abstract compliance claim. ### Common misapplications Capture recurrent scope errors, obsolete versions, missed extensions, checklist-only use, copied boilerplate, confused source types, and conclusions that require escalation. ### Currency trigger Specify when the page must be rechecked: new project, target change, jurisdiction change, product-stage change, revised source, transition date, inspection finding, or process update. ## Governing-Source Checklist Before the outline: - the output, design, destination, jurisdiction, lifecycle stage, and audience are precise; - candidate source types have been inventoried; - current controlling originals and controlled project copies are available; - applicability and version decisions have owners; - extensions, target instructions, procedures, and templates have been checked; - atomic requirements have been mapped into the outline and workflow. Before handoff: - the document and submission package satisfy the mapped requirements; - source, section, display, declaration, metadata, appendix, and portal checks are complete; - conflicts, deviations, not-applicable items, and specialist interpretations are documented; - scientific, statistical, ethical, regulatory, privacy, patient, editorial, and technical reviewers have completed the reviews assigned to them; - the governing versions and final compliance evidence can be recovered. ## Sources This deep-dive shelf is synthesized only from: - Strategic Scientific and Medical Writing - Writing High-Quality Medical Publications - Medical Writing: A Guide for Clinicians, Educators, and Researchers - The Complete Guide to Medical Writing - EMWA Career Guide - Ultimate Guide to Becoming a Medical Writer - AMWA journal corpus - EMWA journal corpus [Return to the Medical Writers' Handbook](/guide)