dvp-composer（dvp 作曲家）

DVP Composer

Purpose

Guide the user through a structured 6-phase interactive workflow to compose a Data Validation Plan (DVP) for clinical trial Data Management in CRO context. Each phase builds on the previous one, with user confirmation required between phases.

Workflow Overview

After completing each phase, present a summary to the user and ask for confirmation before proceeding to the next phase. If the user requests changes, iterate within the current phase before moving forward.

Task Tracking

When the workflow starts, create a task list to give the user a clear roadmap and granular progress tracking.

Phase-Level Tasks

At workflow startup, use TaskCreate to create 6 phase-level tasks in order. Each task must include addBlockedBy pointing to the previous phase's task ID (except Phase 1). Mark Phase 1 as in_progress immediately.

Example:

TaskCreate subject="Phase 1: Collection" description="Gather all input materials and understand study context" activeForm="Collecting study materials"
→ returns task_id (use the returned value, not a hardcoded number)

TaskCreate subject="Phase 2: Scope & Strategy" description="Define validation scope, key data, and risk points" activeForm="Defining validation scope & strategy" addBlockedBy=[<Phase 1 task_id>]

...and so on for all 6 phases.

TaskUpdate taskId=<Phase 1 task_id> status="in_progress"

Sub-Tasks

When entering a phase, read its reference file and create sub-tasks for that phase's steps. Each sub-task should addBlockedBy the current phase's task ID. Mark each sub-task completed as its step finishes. When all sub-tasks are done, mark the phase-level task as completed and the next phase as in_progress.

Dynamic Task Decomposition

When a sub-task involves work across multiple distinct items whose count depends on runtime context, decompose it into leaf-level work items tracked with the task system.

When to decompose — apply when a sub-task:

• Involves iterating over a list whose length is determined at runtime (modules, dimensions, findings, materials)
• Has distinct work items that can be completed independently
• Risks losing progress or missing items without individual tracking

How to decompose:

1. Before starting the sub-task's actual work, identify the items (e.g., read data-modules.md to get the module list)
2. Use TaskCreate to create one leaf task per item — each MUST addBlockedBy the parent sub-task's task ID
3. Set the first leaf task to in_progress
4. Complete each leaf task in order: mark completed, then set next in_progress
5. When all leaf tasks are completed, mark the parent sub-task as completed

Naming: Leaf task subjects use the pattern [Sub-task subject]: [item], e.g. Design by Module: AE/SAE.

When NOT to decompose — skip decomposition when the sub-task has 2 or fewer items, or the items are trivial (single-action each). Use judgment — decomposition should help tracking, not add overhead.

Phase Transition

At each [Done] confirmation:

1. Mark the current phase task as completed
2. Mark the next phase task as in_progress
3. Load the next phase file and create its sub-tasks

Input Methods

Support three input methods simultaneously:

1. Paste text: User pastes protocol summaries, CRF descriptions, etc. directly into the conversation
2. Read files: User provides file paths (PDF, Word, Markdown); use the Read tool to access content
3. Q&A: Claude asks targeted questions, user answers verbally

Adapt questioning strategy based on what the user has readily available. Do not demand all materials upfront — work with what is provided and ask targeted questions for gaps.

Output

Generate an Excel (.xlsx) file as the final DVP document.

Template handling: If the user provides a DVP Excel template during Phase 1, use that template's format and sheet structure. Otherwise, generate with the default 4-sheet format (see references/excel-spec.md).

Default sheets:

• Check List — All validation rules (Check ID, module, description, logic, trigger, query wording, priority, method)
• Summary — Document overview, scope, roles & responsibilities, review strategy
• Revision History — Version change log
• Ext Data Recon — External data reconciliation checks

Workspace

All phase deliverables are written to dvp_workspace/ under the user's working directory.

Rules

1. Create dvp_workspace/ when the workflow starts
2. At each phase's [Done] step, write all deliverable files to dvp_workspace/
3. At the start of each phase (except Phase 1), read the previous phase's deliverable files first
4. File names are hard-coded in each phase instruction — do not rename them
5. Overwrite existing files without asking for confirmation
6. All files use Markdown format. The only exceptions are: dvp_content.json (JSON), DVP_<ProtocolNumber>_v1.0.xlsx (Excel), and any user-provided template files.

Cross-Phase File: assumptions-and-gaps.md

Phase 1 creates this file with the initial table structure: | # | Category | Description | Status | Phase |. Phases 2 and 3 MUST append new rows (with Status = Open) when they encounter information gaps or make assumptions not confirmed by the user. Phase 4 MUST update the Status column from Open to Resolved for each item it resolves. Before appending, MUST read the existing file to avoid duplicate entries.

Phase Instructions

For detailed instructions on each phase, load the corresponding reference file from the skill's references/stages/ directory (relative to the skill root at skills/dvp-composer/):

• Phase 1: references/stages/phase-1-collection.md
• Phase 2: references/stages/phase-2-strategy.md
• Phase 3: references/stages/phase-3-design.md
• Phase 4: references/stages/phase-4-alignment.md
• Phase 5: references/stages/phase-5-draft.md
• Phase 6: references/stages/phase-6-review.md

Load each phase file at the start of that phase. Do not load all phases upfront. After loading the phase file, read the previous phase's deliverable files from dvp_workspace/ (except Phase 1, which has no predecessor). Each phase instruction lists exactly which files to read.

Additional References

• references/section-catalog.md — Standard DVP document structure. Load during Phase 5 to verify all required document sections are included.
• references/excel-spec.md — Detailed Excel output format specification. Load during Phase 5 when generating output.
• references/example-output.md — Example DVP output with sample check rules. Load during Phase 3 if the user wants to see examples.

Script

Use scripts/generate_xlsx.py to generate the final Excel file. The script accepts a JSON input containing all DVP content and produces the formatted .xlsx output.

python3 scripts/generate_xlsx.py --input dvp_workspace/dvp_content.json --output dvp_workspace/DVP_<ProtocolNumber>_v1.0.xlsx [--template dvp_workspace/template.xlsx]

All file paths in the command MUST use the dvp_workspace/ prefix. If the script exits with a non-zero code, read the error message, fix the JSON input, and re-run. MUST NOT proceed to the [Done] step until the script succeeds.

Interaction Protocol

All phases must follow this unified interaction protocol for when and how to ask the user questions.

Decision Framework — When to Ask

At each decision point, determine the appropriate level:

Question Format — How to Ask

Use one of four question types, each with a text label:

[Collect] Information gathering — used when required input is missing:

[Collect] [topic] needs [specific information]
  Background: [why this information is needed]
  Please provide: [what exactly is needed]

[Confirm] Recommendation — used when presenting a default for confirmation:

[Confirm] [decision point]
  Recommendation: [option]
  Rationale: [why this option is recommended]
  Alternative: [other option(s)]
  Please confirm whether to adopt the recommendation, or choose an alternative.

[Conflict] Conflict resolution — used when checks conflict or overlap:

[Conflict] [Check ID] [issue description]
  Finding: [specific conflict content]
  Suggested resolution: [keep/merge/remove] — [reason]
  Please confirm: adopt suggestion / other approach

[Done] Phase confirmation — used at every phase transition:

[Done] Phase N: [phase name]
  Deliverables written to dvp_workspace/:
  - [file1.md] — [one-line description]
  - [file2.md] — [one-line description]

  Output summary:
  - [key output 1]
  - [key output 2]
  - Assumptions made: [if any]

  Next: Phase N+1: [phase name]
  Will proceed after your confirmation. Let me know if adjustments are needed.

Batching Rules

• Combine multiple questions from the same phase into one batch rather than asking one by one.
• Order by priority: Must-ask first, Recommend-confirmation second.
• No more than 5 questions per batch.
• Recommendation-type questions must include rationale; user must explicitly confirm.

Per-Phase Question Guide

Key Principles

1. Interactive pacing: MUST present a [Done] summary at every phase transition and MUST wait for explicit user confirmation before loading the next phase file. MUST NOT proceed past a phase boundary without user approval.
2. Domain accuracy: MUST use standard clinical DM domain abbreviations (AE, LB, VS, DM, IE, CM, EX, PE, SC, MH, DS). MUST use standard check-type labels (Completeness, Consistency, Range, Cross-Module, Timeline). MUST NOT invent non-standard terminology.
3. Practical focus: Every check rule MUST include a Logic Rule that evaluates to a boolean (true/false) outcome. MUST NOT produce checks without a testable condition.
4. Risk-based: Checks for modules rated High risk in risk-assessment.md MUST have at least 5 checks per module. Safety-critical modules (AE/SAE, Exposure/IP) MUST have at least one Critical-severity check.
5. No hallucination: MUST NOT generate check rules for protocol details not present in the user-provided materials. If a check requires a protocol value that is missing, MUST ask the user via [Collect] before creating the check.
6. Language adaptation: MUST write all query wording, [Done] summaries, and user-facing messages in the same language the user is using in conversation. If the user writes in Chinese, output MUST be in Chinese. If in English, output MUST be in English.