Clinical Trial Chief Advisor

Problem

Clinical trial work scatters methodology, design, compliance, QC, and tone guidance across SOPs, templates, and project documents, while sample size, registry, safety, and literature evidence depend on separate data sources. A direct LLM answer can over-search local files, skip required remote tools, handle attachment templates inconsistently, or return the wrong language.

How it works

ct-advisor acts as the single entry point for the ct-series. It first runs scripts/route.py as a deterministic difficulty gate: vague questions enter a local clarification loop, while simple/middle/complex questions proceed to remote analysis. Every question is forwarded through refine_answer.py --ship or orchestrate.py to Coze; local knowledge is treated as a fallback rather than a shortcut. When Coze returns a need_tool card, code executes sibling skills such as ct-samplesize, ct-registry, ct-safety, or ct-literature, then stitches the result back into the draft. docx, pdf, and ppt attachments are converted to Markdown and used for full-specification generation.

Boundaries

The skill expects python3 and an available Coze service. Missing sibling skills should degrade gracefully without inventing data. Local retrieval is limited to one lookup per turn, and misses go to Coze rather than chained reads. Language alignment translates narrative to match the question language, but leaves numbers, proper nouns, regulation names, and structured fields unchanged.