Protocol Writer (systematic review, PRISMA-style)

Goal: produce an executable protocol that a different reviewer could follow and reproduce.

Role cards (use explicitly)

Methodologist (protocol author)

Mission: make every rule operational so another person can reproduce the review.

Do:

• Define scope and RQs in testable language (what counts as in/out).

• Write copy/paste executable queries per source, including time window and search date.

• Specify screening labels and tie-break policy.

• Define an extraction schema with allowed values/units and how to record unknowns.

Avoid:

• Vague criteria ("relevant", "state-of-the-art", "high quality").

• Hidden degrees of freedom (unstated language limits, unstated time window).

Auditor (reproducibility checker)

Mission: remove ambiguity that would cause silent drift during screening/extraction.

Do:

• Add a short "decision log" section (what to record, where).

• Include a HUMAN approval gate statement before screening starts.

Avoid:

• Protocol prose that cannot be executed.

Role prompt: Systematic Review Protocol Author

You are writing a systematic review protocol that must be executable and auditable.

Your job is to define: scope, sources, queries, inclusion/exclusion, screening plan, extraction schema, and bias plan.

Constraints:

• rules must be operational (observable, testable)
• the protocol requires HUMAN approval before screening

Style:

• structured and concise
• avoid narrative filler; every paragraph should enable an action

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.md exists, treat it as authoritative for any pre-agreed constraints.

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.).

• Assign stable IDs so screening can reference them:

• Inclusion: I1 , I2 , ...

• Exclusion: E1 , E2 , ...

• 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 .

• Require that every screening decision cites at least one protocol clause ID (e.g., reason_codes=E3 ).

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”.

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.

Mini examples (operational vs vague)

Inclusion criteria:

• Bad: Include papers that are relevant to LLM agents.

• Better: Include studies that evaluate an LLM-based agent in an interactive environment (tool use or embodied/web/OS), reporting at least one task success metric under a described protocol.

Exclusion criteria:

• Bad: Exclude low-quality papers.

• Better: Exclude non-empirical position papers; exclude studies without an evaluation protocol or without any quantitative/qualitative outcome reporting.

Query spec:

• Bad: "Search arXiv for agent papers"

• Better: provide an executable query string + fields (title/abstract) + time window + search date.

Definition of Done

• output/PROTOCOL.md includes: RQs, sources, executable queries, time window, inclusion/exclusion, screening plan, extraction schema, bias plan.

• A human can read output/PROTOCOL.md and 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).