formtastic

Produce a clean output/proposed-questionnaire.json or output/suggestions.json with stable linkIds, exact baseVersion, and the current typed proposal vocabulary. Use whenever you're about to suggest a structural change to the form.

Write product-facing replies and organize run outputs in the canonical Formtastic artifact layout. Always applies to every turn because reply.md is the visible chat response and all structured output depends on the run directory contract.

Create organized, copyright-safe research artifacts under output/research/. Always applies to evidence, guideline, citation, article, source-review, or literature turns.

Search Cochrane Reviews for systematic-review evidence and effect sizes. Use when the user wants the strongest evidence-base summary for a treatment, screening, or diagnostic question. Produces structured effect-size estimates with confidence intervals.

Find clinical practice guidelines from authoritative bodies (NICE, USPSTF, ACP, AHRQ, GIN). Use when the user asks "what does the guideline say" or wants to anchor a question or threshold to a recommended practice. Produces guideline-typed citations with the canonical URL of the recommendation.

Search PubMed via NCBI E-utilities to ground questionnaire items in primary literature. Use when the user asks for evidence, citations, or "what does the research say about X." Produces an organized research bundle with PubMed metadata, abstracts when available, and source-grounded notes.

Run schema + HL7 FHIR Validator + clinical-sanity validation on Questionnaire.json through Formtastic's host-managed validation endpoint. Use before publish, before a big propose-edit, or when the user asks "is this form valid?"

Rewrite summary.md so future turns don't need to re-read all of conversation.jsonl. Use periodically (every ~20 turns) or when the user explicitly asks "summarize what we've done so far."

Red-team review of a draft Questionnaire before publish. Checks for redundancy, reading-level issues, burden, missing red-flag screens, license problems, and bias. Use when the user says "review this form" or as a self-check before proposing a major change.

Convert a populated findings.json (from research skills like pubmed-search / guideline-search) into FHIR R4 Questionnaire item[]. Handles item.type selection, answerOption construction, and item.code population. Use whenever the user says "now add a question about X" after a research turn.

Common conditional-display patterns for Questionnaire items via core enableWhen. Covers red-flag triage, skip-if-not- applicable, and branch-on-severity. Use when the user wants conditional logic ("only ask this if they answered yes to X").

Find existing validated FHIR Questionnaire instruments before writing custom items. Searches Simplifier, the FHIR package registry, AHRQ's instrument repository, and PROMIS. Use whenever the user asks for a "depression screen", "pain scale", "ADL assessment", etc. — almost always there's a published instrument.

Apply the GRADE framework (Grading of Recommendations, Assessment, Development, and Evaluations) to label evidence quality and recommendation strength. Use when the user wants to weight competing recommendations or label confidence on a proposal.

Decide whether to inline a validated instrument (copy items into Questionnaire.json) versus reference it via subQuestionnaire. Surfaces license + fidelity tradeoffs and records the decision. Use when the user wants to "add PHQ-9" (or any published instrument) to a draft.

Look up LOINC codes for standardized clinical observations and panels via fhir.loinc.org. Use when populating Questionnaire.code / item.code with the right LOINC for a validated instrument item, or to map a draft question onto an existing LOINC observation.

Update decisions.md after a material accept/reject event. decisions.md is the user-facing curated rationale log (separate from auto-generated summary.md). Use whenever the user accepts a meaningful proposal, includes an instrument, or rejects something with a stated reason.

Search SNOMED CT via the public tx.fhir.org $expand endpoint. Use when LOINC isn't the right vocabulary (problems, procedures, body sites, social context). Knows when SNOMED beats LOINC and when to use both side-by-side.