---
name: thomson-li/prisma-slr
source: https://app.decimal.ai/s/thomson-li-prisma-slr@1/SKILL.md
source_sha256: 2e5a7bc17e76
---

# PRISMA Systematic Literature Review

Run a transparent, reproducible literature review that follows PRISMA 2020, so the
process is auditable and the flow numbers reconcile.

## When to use

The user is writing a survey/SLR or needs a defensible method for selecting and
synthesizing papers on a topic. **Not** for a quick informal "find me a few papers"
(that's just a search) or for reading a single paper.

## Workflow

1. **Protocol & questions.** Pin down the research question(s). Frame with PICO/PICOS
   (Population, Intervention, Comparison, Outcome, Study design) or an
   objective-centric equivalent for CS/ML surveys. Record the protocol *before*
   searching; note if it's registered (e.g. PROSPERO) or not.

2. **Search strategy.** Build the boolean search string from concept blocks joined
   with AND, synonyms within a block joined with OR. Record the **exact string per
   database** (e.g. Scopus, Web of Science, IEEE Xplore, ACM DL, arXiv), the fields
   searched, date range, and the date the search was run. Reproducibility hinges on
   this being verbatim.

3. **Eligibility criteria.** Write explicit inclusion and exclusion criteria
   (language, year range, publication type, peer-review status, on-topic relevance,
   etc.). Each criterion must be checkable by a second person.

4. **PRISMA flow — four stages, with counts at every gate:**
   - **Identification** — records found per database + other sources; total.
   - **Screening** — remove duplicates; screen titles/abstracts; record # excluded.
   - **Eligibility** — full-text assessment; record # excluded **with reasons**.
   - **Included** — final studies in the synthesis.
   The numbers must reconcile: identified − duplicates − screened-out − full-text-excluded
   = included. Produce the flow diagram (or describe each box's count).

5. **Data extraction.** Define an extraction schema and apply it to every included
   study: citation, year, venue, dataset(s), method/architecture, evaluation metrics,
   key results, limitations, and whatever fields the research questions demand. Keep
   it as a structured table — one row per study.

6. **Quality / risk-of-bias (optional but recommended).** Apply a checklist
   appropriate to the field; note study limitations that affect synthesis weight.

7. **Synthesis.** Narrative and/or thematic synthesis organized by the research
   questions: trends over time, taxonomy of methods, comparative results, gaps, and
   open problems. Tie every claim back to extracted evidence — no unsupported summary.

## Output

(1) A **methods section** stating protocol, databases, search strings, criteria, and
the PRISMA flow with reconciled counts; (2) the **PRISMA flow diagram** (or its
box-by-box counts); (3) the **extraction table** (one row per included study); and
(4) a **synthesis** organized by research question with an explicit "gaps & future
work" subsection. Keep everything reproducible — anyone should be able to re-run the
searches and arrive at the same included set.

## Notes / edge cases

- **Counts must reconcile end to end.** A flow diagram whose numbers don't add up is
  the single most common reviewer complaint — check it explicitly.
- Record exclusion *reasons* at the full-text stage; "excluded: 14" without reasons
  is not PRISMA-compliant.
- Distinguish *records* (search hits) from *studies* (final units of synthesis); one
  study can span multiple records.
- For fast-moving ML topics, note the search date prominently — the field shifts under
  the review.