Snippet library
Card-based manager for all project snippets: names, descriptions, versions, and the factor levels each version realizes.
Stage 3 · Prompt Construction
Prompts as versioned, auditable building blocks.
Templates plus a reusable snippet library turn each experimental condition into exact prompt text — reviewable in a variation grid before a single trial runs, and versioned so every result traces back to the words the model actually saw.
Templates and snippets: the separation that keeps studies clean
In AIScholar, a prompt is never a loose blob of text. It is a template with placeholders, filled by named snippets — project-scoped text blocks that each carry multiple indexed versions (v0, v1, v2…). A template references a snippet by name, like {{framing}}-style tags, and each snippet version is bound to one or more levels of a design factor.
At execution time the platform resolves each condition using the snippet versions bound to its factor levels. This makes the intended mapping inspectable; review the rendered text to confirm that each condition expresses the manipulation you planned. Single-version snippets provide shared text across conditions where appropriate.
How prompt construction works
Write or generate snippets
Create named text blocks or edit versions in a grid. Each version of a multi-version snippet can represent a factor level; check that the wording preserves the intended meaning.
Bind versions to factor levels
Attach each snippet version to the (factor, level) pairs it realizes. The platform validates coverage: every level of every factor a template varies must have a bound version.
Compose templates
Build templates that reference snippets by name. Validation can flag some structural issues, but researchers still review the rendered prompts for construct validity, ambiguity, and unintended differences.
Preview the variation matrix
Inspect the resolved prompt variations and identify which text differs between conditions. Review each included instantiation before running.
Version on every edit
Saved template versions and trial references can help you trace which text was used; confirm the recorded versions when reporting the method.
Tools that do the heavy lifting
Variation grid
Spreadsheet-style bulk editing of snippet versions across factors and levels — the fastest way to author a full design's worth of wording.
Template versioning
Version records help trace template edits and the versions associated with trials.
Live preview with highlighting
Template previews highlight snippet insertions in blue and surface invalid factor-name placeholders in red, so structural errors are visible at a glance.
AI prompt wizard
Can draft template and snippet material and flag some structural issues. Review the result yourself; automated checks do not establish that wording is clear, safe, or a valid operationalization.
Prompt variants
Draft candidate wording variants across dimensions such as paraphrasing, structure, or verbosity. Check that each variant preserves meaning before treating variants as a robustness test.
A variation matrix makes prompt differences easier to inspect, but does not validate their meaning by itself. The prompt-sensitivity walkthrough illustrates how prompt variants can be examined. For execution details, see experiment execution for supported single-turn text trials.
Study LLM behavior with a guided research workflow.
Design controlled text studies, collect model responses, and explore what changes across conditions. You decide what the evidence means.