protocol-writer
将系统评价方案写入`output/PROTOCOL.md`(数据库、查询、包含/排除、时间窗口、提取字段)。**触发**:方案、PRISMA、系统评价、包含/排除、检索式、纳入排除。**使用时机**:系统评价管道的起点(C1),需要先锁定方案再开始筛选/提取。**如果**跳过:不是做系统评价(或方案已经锁定且不允许修改)。**网络**: none.**Guardrail**: 协议必须包含进行的检索与筛选规则;需要 HUMAN 签字后才能进入筛选。
📖 中文导读:将系统评价方案写入`output/PROTOCOL.md`(数据库、查询、包含/排除、时间窗口、提取字段)。**触发**:方案、PRISMA、系统评价、包含/排除、检索式、纳入排除。**使用时机**:系统评价管道的起点(C1),需要先锁定方案再开始筛选/提取。**如果**跳过:不是做系统评价(或方案已经锁定且不允许修改)。**网络**: none.**Guardrail**: 协议必须包含进行的检索与筛选规则;需要 HUMAN 签字后才能进入筛选。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。
name: protocol-writer
description: |
Write a systematic review protocol into output/PROTOCOL.md (databases, queries, inclusion/exclusion, time window, extraction fields).
Trigger: protocol, PRISMA, systematic review, inclusion/exclusion, 检索式, 纳入排除.
Use when: systematic review pipeline 的起点(C1),需要先锁定 protocol 再开始 screening/extraction。
Skip if: 不是做 systematic review(或 protocol 已经锁定且不允许修改)。
Network: none.
Guardrail: protocol 必须包含可执行的检索与筛选规则;需要 HUMAN 签字后才能进入 screening。
Protocol Writer (systematic review, PRISMA-style)
Goal: produce an executable protocol that a different reviewer could follow and reproduce.
Inputs
Required:
STATUS.md(context + scope notes)
Optional:
GOAL.md(topic phrasing)DECISIONS.md(any pre-agreed constraints)
Outputs
output/PROTOCOL.md
Workflow
Scope + research questions
- Translate the goal in
GOAL.md(if present) into 1–3 review questions. - State what is in-scope / out-of-scope (keep consistent with
STATUS.md). - If
DECISIONS.mdexists, treat it as authoritative for any pre-agreed constraints.
- Translate the goal in
Sources
- List databases/sources you will search (e.g., arXiv, ACL Anthology, IEEE Xplore, ACM DL, PubMed).
- Specify any manual routes (snowballing: references/cited-by).
Search strategy (copy/paste executable)
- For each source, write a concrete query string.
- Define the time window (from/to year) and language constraints.
- Record “search date” so the run is auditable.
Inclusion / exclusion criteria (operational, not vague)
- Write MUST-HAVE criteria (study type, domain, outcomes).
- Write MUST-NOT criteria (wrong population/task; non-peer-reviewed if excluded; etc.).
- Define how you handle duplicates and near-duplicates.
Screening plan
- Define the screening stages (title/abstract → full text if applicable).
- Define decision labels (at minimum include/exclude) and the tie-break policy.
- Specify what gets recorded into
papers/screening_log.csv.
Extraction schema (downstream contract)
- Define the columns that will appear in
papers/extraction_table.csv. - Ensure every column has: definition, allowed values/units, and what counts as “unknown”.
- Define the columns that will appear in
Bias / risk-of-bias plan
- Define the bias domains you will use (simple scales are OK).
- Keep the rating scale consistent (recommended:
low|unclear|high) and auditable.
Write
output/PROTOCOL.md- Use clear headings; avoid prose that cannot be operationalized.
- End with an explicit “HUMAN approval required before screening” note.
Definition of Done
-
output/PROTOCOL.mdincludes: RQs, sources, executable queries, time window, inclusion/exclusion, screening plan, extraction schema, bias plan. - A human can read
output/PROTOCOL.mdand run screening without asking “what do you mean by X?”.
Troubleshooting
Issue: queries are too broad / too narrow
Fix:
- Add exclusions for common false positives; add missing synonyms/acronyms; restrict fields (title/abstract) where supported.
Issue: screening/extraction criteria are vague (“relevant”, “state-of-the-art”)
Fix:
- Replace with observable rules (task/domain, metrics, dataset requirements, intervention/controls).