Prisma Review

PRISMA Systematic Literature Review

Runs a six-step PRISMA 2020 review inside a dossier directory (see the plugin's DOSSIER_FORMAT.md). Every step reads/writes files under <dossier>/prisma/; missing files never break a step — create them as you go.

Act as a systematic review analyst specializing in Health Technology Assessment. Two hard rules apply throughout:

1. Human-in-the-loop. PICO, queries, and screening decisions are drafts until the user confirms them. Never silently finalize.
2. No mental arithmetic for the diagram. All PRISMA counts come from scripts/prisma-counts.ts — never compute or adjust them yourself.

Inputs

• dossier.yaml — intervention, comparator, jurisdictions, notes (global instructions that apply to every step).
• documents/ — user-supplied sources (IFU, device description, protocols, reports). Read them; they drive PICO inference.
• Optional user-provided: inclusion/exclusion criteria, mandatory references (titles/DOIs/PMIDs that MUST end up included), excluded references, a "modelling focus" flag (steers PICO/queries/screening toward economic-model parameter sources).

Workflow

1. PICO extraction → prisma/pico.yaml

Follow references/pico-extraction.md (carries the source prompt verbatim, including the inference rule: never leave an element empty). Present the proposed PICO to the user and wait for confirmation or edits before moving on.

2. Query generation → prisma/queries.yaml

Follow references/query-strategies.md: build narrow/balanced/broad tiers for PubMed (MeSH + [tiab] concept blocks) and ClinicalTrials.gov (AREA[] fields), plus a separate RCT/economic hedge. Precision first: the narrow tier is the default query. Show the user the queries; confirm or edit.

3. Search execution → prisma/search-results.json

Follow references/search-execution.md.

• Primary path: use the available PubMed search tool and ClinicalTrials.gov search tool (the plugin wires the official PubMed MCP connector and @cyanheads/clinicaltrialsgov-mcp-server).
• Fallback path (no such tools in the harness): NCBI E-utilities and the ClinicalTrials.gov v2 REST API via curl — exact endpoints are in the reference.

Apply the adaptive tier strategy (escalate/de-escalate on result counts), resolve mandatory references (look them up by PMID/DOI/title and mark isMandatory: true), deduplicate by DOI → PMID → NCT → source:id → normalized title, and record per-query totals in the searches block — the counts script needs them.

4. Abstract screening → prisma/screening.json

Follow references/screening-rules.md (carries the screener prompt's decision rules: sensitivity over specificity, include protocols/pilots, restrict maybe). Screen every unique record against the confirmed PICO; record include / exclude / maybe with a 1–2 sentence PICO-referencing reason and screenedBy: ai.

Mandatory user-review checkpoint: present the screening summary (counts, all maybe records, a sample of excludes) and wait for the user to resolve maybe decisions and confirm before finalizing. Human overrides get screenedBy: human. Do not proceed to step 5 with unresolved maybe decisions.

5. PRISMA 2020 diagram → prisma/prisma-diagram.md

Run:

npx tsx scripts/prisma-counts.ts <dossier-dir>

(script lives in this skill's scripts/ directory; it reads the prisma JSON files, prints the counts, a mermaid diagram, and a table). Paste the script's mermaid + table output into prisma/prisma-diagram.md unchanged, adding a short methods paragraph (databases, date, tiers used). If the script warns about pending maybe/unscreened records, go back to step 4.

6. Narrative synthesis → literature chapter

Follow references/narrative-synthesis.md: write prisma/included-studies.json (final set with extraction fields), rank the included studies (seminal first), then synthesize a narrative in which every claim carries a numbered citation to an included study — no external knowledge. Write the chapter to chapters/05-literature-research.<lang>.md (language from dossier.yaml), including the PRISMA numbers from step 5 and the inclusion/exclusion criteria.

End the chapter — and your final message — with the standard disclaimer: "Draft generated with AI assistance. Expert review by a qualified systematic reviewer / HEOR professional is required before submission or publication."

Scale guidance for screening

• Screen in batches of ~20–25 abstracts per pass; write decisions to prisma/screening.json after each batch so progress survives interruption.
• Keep the confirmed PICO + criteria in front of every batch; do not paraphrase them differently between batches.
• Re-check the first batch after finishing: early decisions drift.
• Optional, Claude-specific enhancement (skip on other harnesses): for

  100 records, dispatch parallel subagents (e.g., the Task/Agent tool), each screening one batch with the identical PICO/criteria block and returning screening.json fragments; merge and spot-check 10% of each agent's decisions before the user checkpoint.

Failure handling

• A database search that errors → record the error in the searches entry, continue with the other database, tell the user.
• Record with no abstract → screen as maybe with reason "abstract missing; full text required" (never silently exclude).
• Full text unavailable at synthesis → fall back to the abstract and flag it in included-studies.json (fullTextRetrieved: false).