> For the complete documentation index, see [llms.txt](https://docs.nested-knowledge.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nested-knowledge.com/ai-transparency/ai-in-nested-knowledge/prompt-guidance.md). # Prompt Guidance Learn how to write effective prompts for AI extraction. [**Smart Screener**](/review-workflow/screening/criteria-based-screening/smart-screener.md) and [**Adaptive Smart Tags**](/review-workflow/tagging/ai-tools-in-tagging/adaptive-smart-tags.md) (ASTs) use a large language model (LLM). Every criterion or tag question is a prompt to that model. Clear prompts produce consistent answers you can review and audit. Treat each prompt like a protocol instruction: define the decision, scope, and expected response. {% hint style="info" %} Use this page before running **Smart Screener** or **Adaptive Smart Tags**. It covers prompt wording, not how to run either tool. {% endhint %} ### Build And Test Prompts Start with a defined research question, eligibility criteria, and data elements. These will avoid ambiguous prompts and inconsistent results from the LLM. {% stepper %} {% step %} #### Inspect Known Studies Review a handful of studies you know are relevant. Check which details appear in abstracts and which require full text. {% endstep %} {% step %} #### Draft The Prompt Start with a direct instruction. Add definitions, examples, synonyms, and output requirements when they help. Spell out acronnyms. More explicit instructions usually resolve inconsistent answers. {% endstep %} {% step %} #### Define Missing Information State what to assume when the source does not report a detail. **Smart Screener** answers Yes or No when source text exists. It leaves a record unanswered only when no relevant source text exists or NR when an answer is not reported. {% endstep %} {% step %} #### Pilot And Revise Run prompts on a small batch. Review answers, revise the wording, and rerun. Several draft-and-test rounds are normal. {% endstep %} {% endstepper %} #### Choose Annotations Deliberately Both tools can require a supporting text excerpt, called an annotation. Required annotations make every answer traceable. They can also make the model conservative when a passage implies, rather than states, information. Allow answers without annotations when the model must reason across a passage. Require annotations when direct evidence matters more than inferred context.

Prompt configuration in Nested Knowledge

### Start With A Draft You do not need to write every prompt from scratch. [**Smart Config: Screening**](/review-workflow/screening/criteria-based-screening/smart-config-screening.md) drafts Abstract and Full Text criteria from your research question. [**Smart Config: Tagging**](/review-workflow/tagging/ai-tools-in-tagging/smart-config-tagging.md) drafts a tag hierarchy. Review and revise every generated question. **Smart Config** provides a starting point, not a finished prompt set. For tagging, [**Import Tag Hierarchy Templates**](/review-workflow/tagging/configure-tagging/import-tag-hierarchy-templates.md) offers another starting point. Review template questions before running **ASTs**. ### Write Criteria For Smart Screener

Screening criteria configuration in Nested Knowledge

On [Configure Screening](/review-workflow/screening/configure-screening.md), order criteria from broad and quick to assess through detailed. Start with details evident in the title or abstract, such as study type or geography. The first failed criterion becomes the exclusion reason. This order determines your PRISMA diagram. For a PICOS-based review, give each component its own criterion. A combined question cannot map cleanly to one exclusion reason. Write separate Abstract and Full Text wording, even for similar criteria. These sources report different detail. State how **Smart Screener** should handle information abstracts often omit. State the default for missing information in the question. For example: assume an unstated study size meets the threshold unless it is a case report. Test criteria on a deliberately small batch before the full record set. Review the results and revise the wording before you scale. [**Smart Screener**](/review-workflow/screening/criteria-based-screening/smart-screener.md) explains how to filter the batch and restore excluded records. A few example criteria, to show the level of detail worth aiming for: * **Study type:** "Is this an original report of a clinical study — randomized, observational or cohort, or case report or series — and not a systematic review, meta-analysis, secondary report, or protocol?" * **Study size:** "Does the study include five or more patients? If size isn't reported, assume yes unless the study is a case report." * **Population:** "Is the study population primarily adults, not a pediatric or animal population? If not stated otherwise, assume yes." * **Intervention:** "Does the study evaluate \[intervention] as a treatment?" ### Write Tag Questions For Adaptive Smart Tags

Tag question configuration in Nested Knowledge

On [Configure Tagging](/review-workflow/tagging/configure-tagging.md), use tag names that make sense on their own. **ASTs** reads the tag name with its question. A clear, spelled-out name gives it useful context. For a text tag, write the question to define the concept, give an example, and state the output you want, including the level of detail, units, or summary statistic. State what not to extract, too, if the tag has a common source of noise. For example, ask for age as mean (or SD) in years for the overall study population, and specify per-arm reporting for studies that only report age by arm. Use **Text Options** when you want a tag's answer constrained to an exact set of words or phrases, similar to a dropdown. This works well for study design, treatment names, or subgroup labels that need to match exactly across studies for later analysis. Configure a tag as **Numeric** only when you expect a single count, such as number of patients enrolled or number of centers. Be cautious with Numeric for anything reported heterogeneously, like age as mean or median: a Numeric answer returns only the number, not the type of summary statistic behind it, so split these into separate tags per statistic instead of one Numeric tag. For **Tag Tables**, write a question covering the full table. Give every column header enough detail to stand alone. **ASTs** reads column headers as instructions. See [Match Tags to Your Export Spreadsheet](/review-workflow/tagging/match-tags-to-your-exports.md) when mapping columns to existing fields. Favor **Single Apply** when the content type allows it. The same setup works for text, tables, options, and numbers. It returns one answer. Use **Select** questions for categorical concepts with fewer than eight child tags. More options can make selecting the closest match harder. A tag needs a **Single Apply** question or a parent **Select** question for **ASTs** to answer it. Match a question's phrasing to the answer type you actually want: present- or past-tense "What" or "Where" questions for text content, and "Is" or "Was" questions for a Single Apply Yes/No concept. Keep the hierarchy in **Form-Based Tagging** while developing AI-facing wording. This mode provides a dedicated question field for each tag. **ASTs** also works in Standard Tagging, but without a dedicated question. A few example tag setups: * **Chronic kidney disease** — Single Apply — "Is chronic kidney disease reported?" * **Study phase** — Single Select — "Is this study part of a drug approval phase?" — child tags: Phase I, Phase II, Phase III, Phase IV * **Infections** — Multiple Select — "What types of infections were reported?" — child tags: Viral infection, Bacterial infection, Fungal infection, Other infection ### Get Better Results As You Go Check a handful of studies before relying on a full run. Patterns often emerge immediately. Revise and rerun when they do, rather than correcting individual answers. Write literal, precise questions that match the answer type you want. "Was the number of patients reported?" asks for a Yes/No; if what you actually want is the number itself, ask "What was the number of patients?" instead. Build only the criteria or tags your research question actually needs. Extraneous tags and over-broad extraction slow down review, and often signal that the research question itself needs sharpening. If a topic is genuinely tangential and better answered by a different set of studies, a second nest may be a better fit than folding it into this one. ### FAQs

Do I need to rewrite Smart Config's output before I can trust it?

Review it the same way you'd review your own first draft. Smart Config's criteria and tag questions are a starting point to revise using the guidance above, not a finished set of prompts.

Does better wording matter more for screening criteria or tag questions?

Both. Criteria feed directly into inclusion and exclusion decisions, and tag questions determine what makes it into your final extracted dataset — getting either wrong has a real cost downstream.

Can I reuse the same criterion wording at Abstract and Full Text screening?

Give each stage its own wording, even where the two criteria test similar ground, so each stage's decisions stay attributable to that stage's questions.

Should I write every tag question myself, or is Smart Config or a template enough?

Either is a fine starting point. The wording principles on this page apply whether you're refining a generated tag question or writing one from scratch.

Why does my tag need a Single Apply or Select question before ASTs can extract it?

**ASTs** answers a tag's question directly. Other tag setups provide no question to answer, so **ASTs** performs less well.

### See Also


Configure Screening	Choose reviewer structure, pass structure, and screening paradigm.	Configure Screening
Configure Tagging	Set up the tag hierarchy and its questions.	Configure Tagging
Smart Config: Screening	Draft screening criteria from your research question.	Smart Config: Screening
Smart Config: Tagging	Generate a custom tag hierarchy from your research question.	Smart Config: Tagging
Smart Screener	Answer configured screening criteria with AI.	Smart Screener
Adaptive Smart Tags	Extract data for your custom tag questions.	Adaptive Smart Tags

--- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.nested-knowledge.com/ai-transparency/ai-in-nested-knowledge/prompt-guidance.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.