ct-advisor

面向临床研发全生命周期的 ct 系列「总入口」，是方法学、法规证据与实操细节的总顾问：方法学/设计/合规/QC/语气类问题在内部走 A–J 工作流自行解答；统计计算转交 ct-samplesize；原始数据/竞品情报类需求通过 Skill 工具路由到 ct-registry / ct-safety / ct-literature 三个数据源；竞品情报总览由本技能自行缝合三源产出。方法学默认基于本地知识包作答，双模型交叉验证列入后续路线图。 / The ct-series TOTAL ENTRY POINT across the full clinical-development lifecycle — your overall advisor for methodology, regulatory evidence, and hands-on operational detail. Methodology / design / compliance / QC / tone questions are answered in-house through workflows A–J; sample-size computation is handed to ct-samplesize; raw-data and competitive-intel needs are routed via the Skill tool to the three sibling data skills (ct-registry / ct-safety / ct-literature); the full competitive-intel picture is stitched in-house from those three sources. Methodology runs on the local knowledge pack by default; dual-model cross-validation is on the roadmap.

Clinical Trial Advisor

Language

• English guide → README.md
• 中文指南 → README_zh-CN.md

This skill responds in the user's input language and auto-switches; runtime prompts switch by locale. SKILL.md body is English-only (agent-facing); bilingual walkthroughs live in the two READMEs.

Requirements

Item	Requirement	Notes
Runtime	Agent (LLM) reads knowledge/ directly	No mandatory dependency — pure methodology (workflows A–J) runs fully offline.
Optional CLI helpers	python3 (stdlib only)	scripts/menu.py, scripts/check_deps.py, scripts/search_refs.py use only the Python standard library — no third-party packages (PyYAML removed; scripts/*.json is loaded via json).
Sibling data/compute skills	ct-registry, ct-safety, ct-literature, ct-samplesize	Only needed for data_intel routing & layer-B data grounding; missing ones degrade gracefully (never fabricate). Methodology works without them. The full competitive-intel brief is stitched in-house from the three data skills (no separate orchestrator skill).
Coze mode (optional)	config.json backend: coze + coze.bot_id; requests at runtime	Disabled by default. The adapters/ Coze stubs are not executed in local mode.
Outbound network	None in default (local) mode	Only the routed sibling skills make network calls; this skill itself stays zero-outbound unless you invoke them.

This skill is an "orchestration front-door", not a "monolithic reasoner"

Seams pre-split for a future Coze endpoint (swappable adapter layers):

Concern	Location	Local default	After Coze plug-in
Methodology knowledge (portable)	knowledge/	system_prompt.md + scripts/workflows.json + ref-*.md	push the whole pack to the bot as its knowledge base
Reasoning exit	adapters/backend.py	local mode does not go through this layer: the agent reads knowledge/ directly to answer; this module is only for the Coze backend	CozeBackend (HTTP → bot, stub implemented)
Data grounding	adapters/data_context.py	LocalDiskDataContext (scans sibling-skill outputs)	CozeApiDataContext (stub)
Q&A persistence	adapters/qa_store.py	JsonlStore (local JSONL)	RemoteDbStore (stub)
Outbound sanitization	adapters/sanitize.py	always on	always on
Runtime selection	config.json	backend: local (zero outbound)	backend: coze + coze.bot_id

Local-mode execution flow (default, zero outbound)

In local mode the agent reads the knowledge pack under knowledge/ directly and answers; it does not call any Python in adapters/ (those modules are only for the Coze backend and are unused in local mode). All outbound / sanitization logic only truly takes effect in Coze mode.

1. Clarification gate (gate 0) — triage first (friendly menu policy): classify the user's first message as simple (specific, single-intent, answerable directly → answer in one pass, no menu), complex (multi-decision / needs a workflow choice → present the clarification menu, confirm step by step), or vague (need unclear / user undecided → enter grill-me clarify mode, Workflow K, to scope step by step). Default to the friendliest path: when in doubt between simple and complex, give a short direct answer + an optional deeper-menu offer (clarify.triage_simple) rather than forcing a menu. Only open the full menu when step-by-step confirmation genuinely helps. For a vague ask present the clarification menu (scripts/menu.json, rendered via canonical strings in scripts/i18n.py / knowledge/prompts.md) or invite grill-me (clarify.vague_invite): Tier 0 profile (role / stage / input) → intent area → workflow A–J → within-workflow sub-intent → output format. For a still-vague ask use AskUserQuestion to ask only 1–3 conclusion-changing high-value questions per round (e.g. comparison type / estimand strategy / primary endpoint / population) until the problem profile is complete; for high-risk formal answers confirm the profile with one sentence before answering (rule in scripts/workflows.json gate).
2. Route: pick workflow A–J or a composite route from scripts/workflows.json (e.g. "confirmatory trial design" = 0 → B → C → D → E).
3. Assemble payload (documentation only, no outbound): question + workflow + jurisdiction + data_refs + constraints; local mode sends no network request, this step is only a self-check for completeness.
4. Answer: combine knowledge/system_prompt.md methodology rules + scripts/workflows.json routing + ref-*.md topic notes to produce advice; never fill factual gaps with fluent wording.
5. Data grounding (layer B, on demand): when advice needs real-data support, read sibling-skill outputs per scripts/workflows.json integration.data_grounding (CDE trial count, FAERS signals, literature, competitor brief), explicitly label "Data source: ct-xxx on ". Pure methodology / design / compliance questions with no data need may skip this step and state "no data grounding performed".
6. Handoff (workflow C): when sample-size / power parameters are complete → hand off to ct-samplesize for actual computation. Handoff payload template:

• Design: two independent samples / paired / survival etc.; comparator type (placebo / active); superiority / non-inferiority / equivalence; allocation ratio
• Test: one- / two-sided, α, power
• Effect size: continuous → Δ + pooled SD; binary → two-group rates; survival → HR
• Add: dropout / unevaluable rate dropout → n_adj = n / (1 − dropout)
• computed by ct-samplesize; this skill does not compute n in-house.

7. Persist (workflow J): when the user explicitly asks to remember a preference / decision → use the existing WorkBuddy memory mechanism (explicit authorization required), do not create files and do not call qa_store.py.

Boundaries with other ct skills

Skill	What it does	Relationship with this skill (ct-advisor)
ct-registry	Search trial registries & normalize	It "finds trials"; this skill "explains / designs / ensures compliance", reads its output for data grounding, does not re-search.
ct-safety	FAERS signal detection	It "computes signals"; this skill does qualitative classification & report-path judgment, reads its signals to support monitoring advice.
ct-literature	Public-literature search & normalization	It "finds literature"; this skill reads its output to argue methodology precedent, does not re-search.
ct-samplesize	Sample-size & power computation	It "computes n"; this skill "gives the parameter framework in workflow C", hands off once complete.
ct-base	Internal base (D-tier)	This skill is a B-tier published skill; reuses ct-base's i18n.py / excel_style.py for generic & Excel strings, and ships its own scripts/i18n.py for advisor-specific user prompts.

Routing & total-entry (absorbs the ct console router)

ct-advisor is now the single entry point for the whole ct series. It absorbed the former ct console: you no longer open a separate dispatcher — say "ct console / 临床试验技能入口" (or just ask) and this skill routes. It re-implements no retrieval/analysis logic; it delegates real-data and compute work to the sibling skills via the Skill tool, exactly as the old console did.

Three entry capabilities (see scripts/menu.json flows:; the clarification menu asks which one):

Need	Route	How
Methodology / design / statistics / estimand / GCP / DSUR / CSR / methodology QC / tone writing / local memory	Answered in-house (workflows A–J)	No Skill-tool handoff; this skill reasons from knowledge/.
Not sure what I need / want help scoping the question	Clarify mode (grill-me)	In-house; agent asks 1–3 branching questions per round (each with a recommended default), builds a needs portrait + recommends a route (methodology workflow A–J or data_intel skill). No Skill-tool handoff, no network.
A drug's / indication's registered trials (planned / recruiting / completed, competitor landscape)	ct-registry	Skill tool skill="ct-registry"; cross-source (CT.gov / CDE / WHO ICTRP / EU-CTR / ChiCTR / ISRCTN / DRKS).
Safety signals for a drug–event pair (PRR / ROR / IC on FAERS)	ct-safety	Skill tool skill="ct-safety"; FDA FAERS via openFDA.
Published evidence (papers / systematic reviews / RCTs)	ct-literature	Skill tool skill="ct-literature"; OpenAlex + Europe PMC + Semantic Scholar.
Full competitive-intel picture of a drug / indication (one consolidated Strategic Brief)	ct-registry + ct-safety + ct-literature (stitched in-house)	Call all three data skills (skill="ct-registry" / skill="ct-safety" / skill="ct-literature") once each, then stitch the Strategic Brief yourself. Recommended default for broad asks.
Sample-size / power computation	ct-samplesize	Handoff from workflow C once parameters are complete.

When the ask is broad ("竞争情报 / 格局 / 某药全貌 / strategic brief"), call ct-registry + ct-safety + ct-literature once each and stitch the brief in-house; when narrow (one specific dimension), route to the matching focused skill. When the ask is methodology / design / compliance / QC / tone, answer in-house via the clarification menu → workflow A–J. Skip the menu when the user's first message already names a clear target.

🚫 HARD RULE — trio for broad asks, single atomic for narrow, never redundant: there is no one-stop orchestrator skill anymore. For a broad ask, call ct-registry + ct-safety + ct-literature once each and stitch the Strategic Brief yourself; for a narrow ask, call only the one matching skill. Do not call all three when the user asked a single dimension (that is redundant retrieval and double counting against any usage quota).

The data/compute skills stay modular and independent; this skill only routes to them and reads their REAL outputs for data grounding. Deleting the old ct console is safe — its routing table is reproduced above.

A sibling data/compute skill is missing? This skill routes but does not re-implement retrieval or computation logic — so a target skill must be installed to actually fetch data. Handle it gracefully (never fabricate, never fail silently):

• If you already know it is missing (the user told you, or python3 scripts/check_deps.py reports it missing), skip the Skill call. Tell the user: (1) which skill is required; (2) how to install it — same source as ct-advisor (SkillHub / GitHub / local copy); (3) the methodology prep you can still do (draft the query, list the registries / fields that matter, outline the analysis framework); (4) explicitly label the reply "未实际取数 / data not retrieved".
• If you already called the Skill tool and it errored (skill not found), catch it and degrade the same way — never invent trials, signals, literature or sample size to fill the gap.
• ct-samplesize missing → workflow C still outputs the sample-size framework + information gap; tell the user to install it to compute n.

To see which sibling skills are installed right now, run python3 scripts/check_deps.py (local-only probe; it installs nothing and makes no network calls). Methodology (workflows A–J) works fully offline regardless.

China regulatory depth (C-layer)

CTA / IND 60-day tacit approval, communication meetings (Type A / B / C), registration ≠ tacit approval, etc. — see knowledge/ref-regulatory-versions.md (controlled version quick-reference, with ⚠️ and check fields) + ref-regulatory-statistical.md. Any version / status / deadline conclusion must be verified in real time against the official original per the snapshot's "fields to check".

Switching to Coze (future, no methodology rewrite needed)

adapters.build_backend(config.json) reads the backend field; once coze.bot_id is ready, take CozeBackend, push the same knowledge/ to the bot as its knowledge base, and fill in the Coze implementations of DataContextProvider / QASessionStore. Outbound payload always goes through sanitize() first — never carry PII / confidential fields / token (per ct-base §11).

Quality gate & stop rules

Pre-delivery checks and stop conditions are in knowledge/system_prompt.md "Quality gate & stop rules". Core red line: never expose in user-visible content personal info, subject info, unpublished project data, private path or access credential.

Topic reference routing

• Cross-domain synthesis / full lifecycle / comprehensive QC: knowledge/ref-clinical-operations.md
• ICH E2 / E3 / E6 / E9, CTD / M4, NMPA / CDE, DSUR / CSR: knowledge/ref-regulatory-statistical.md + ref-regulatory-versions.md
• Clarification / tone writing / local memory: knowledge/ref-interaction-style.md
• Locate topic content: python3 scripts/search_refs.py "estimand|intercurrent event|sensitivity"

---

Changelog

Full version history (0.1.0 → 0.7.11, 30+ entries) lives in CHANGELOG.md.