Appraises a systematic review or meta-analysis with AMSTAR 2 (AMSTAR-2; Shea et al., BMJ 2017): answers the 16 checklist items from the review's report, flags flaws in the seven critical domains and derives the overall confidence rating — High, Moderate, Low or Critically low — never a summed score. Use when deciding whether a published review can be relied on, or ranking reviews for an umbrella review: "appraise this systematic review with AMSTAR 2", "how trustworthy is this meta-analysis?", "which AMSTAR 2 items are critical?". Not for one trial (use `assess-study-bias`), a body of evidence per outcome (use `evidence-appraisal`) or conducting a review (use `systematic-review`).
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Appraises a systematic review or meta-analysis with AMSTAR 2 (AMSTAR-2; Shea et al., BMJ 2017): answers the 16 checklist items from the review's report, flags flaws in the seven critical domains and derives the overall confidence rating — High, Moderate, Low or Critically low — never a summed score. Use when deciding whether a published review can be relied on, or ranking reviews for an umbrella review: "appraise this systematic review with AMSTAR 2", "how trustworthy is this meta-analysis?", "which AMSTAR 2 items are critical?". Not for one trial (use `assess-study-bias`), a body of evidence per outcome (use `evidence-appraisal`) or conducting a review (use `systematic-review`).
{"category":"evidence-verification","method":"AMSTAR 2 (A MeaSurement Tool to Assess systematic Reviews, version 2)","origin":"Beverley J. Shea et al., BMJ 2017 (original AMSTAR — Shea et al., 2007)","version":"1.0.0"}
AMSTAR 2 review appraisal
AMSTAR 2 (Shea et al., BMJ 2017) appraises systematic reviews of interventions that include randomised trials, non-randomised studies or both: 16 items answered from the review's report, seven of them critical domains, combined not into a score but into an overall confidence rating — High, Moderate, Low or Critically low. Its principle: one flaw in a critical domain — no protocol, thin search, unlisted exclusions, no risk-of-bias assessment, poor pooling, ignored bias, unexamined publication bias — can invalidate a review's conclusions however polished the rest looks.
When to invoke
Invoke when:
A decision is about to lean on a published review or meta-analysis: "appraise this systematic review with AMSTAR 2 before we rely on it", "how trustworthy is this meta-analysis?".
Several reviews on one question disagree or must be ranked (umbrella review, overview).
The question is which AMSTAR 2 items are critical, and what rating a set of answers yields. Non-clinical interventions (workplace, education, software) are appraised by adapting the items.
Do NOT invoke when:
The object is a single trial or study — use assess-study-bias (RoB 2 / ROBINS-I).
The question is the certainty a body of evidence deserves per outcome — use evidence-appraisal (GRADE); AMSTAR 2 rates methods, not certainty.
The task is to produce a review or pool studies — use systematic-review, meta-analysis.
The source is not a systematic review — use rate-source-admiralty.
Procedure
1 — Identify the review and fix its question
Record the citation, the review type (RCTs only / NRSI only / both; meta-analysis or narrative synthesis) and its PICO — population, intervention, comparator, outcomes. Agree first, as the paper advises, how the items apply in the practice or policy context at hand.
2 — Pre-specify the critical domains and the Partial Yes convention
Default critical domains: items 2, 4, 7, 9, 11, 13, 15 (Box 1); the paper lets appraisers add or substitute — e.g. item 3 for a harms review that excluded non-randomised studies. Fix the list before answering. Fix the Partial Yes convention too: the paper adds it "to identify partial adherence" but never weights it, so the team pre-specifies (De Santis et al., 2023). Default here — a Partial Yes is a non-critical weakness; strict teams make it a flaw in a critical domain.
3 — Answer the 16 items from the report, quoting the evidence
Read the report, appendices and any registered protocol. Record Yes / Partial Yes / No per item — 11, 12, 15 take "No meta-analysis conducted" when nothing was pooled — plus where the evidence is. Partial Yes exists only for items 2, 4, 7, 8, 9; elsewhere anything short of the "For Yes" criteria is No. Never give the benefit of the doubt: "If no information is provided to rate an item … the item should be rated as a 'No'." Items 9 and 11 carry separate RCT and NRSI panels — rate both, take the weaker. Wording, options and criteria for all 16: references/items.md, or python3 scripts/amstar2.py items.
4 — Classify flaws and weaknesses
A No in a critical domain is a critical flaw; a No elsewhere a non-critical weakness; a Partial Yes counts per step 2; "No meta-analysis conducted" is neither — the item does not apply. Add nothing up: the authors "strongly recommend that individual item ratings are not combined to create an overall score".
5 — Derive the overall rating mechanically (Box 2)
High — at most one non-critical weakness, and no critical flaw.
Moderate — two or more non-critical weaknesses, still no critical flaw.
Low — exactly one critical flaw, whatever the number of non-critical weaknesses.
Critically low — two or more critical flaws, whatever the number of non-critical weaknesses.
Box 2's footnote — accumulating non-critical weaknesses may erode confidence enough to warrant moving Moderate down to Low — is advisory; state it if applied. Run python3 scripts/amstar2.py rate --file answers.json for the rating and rule trace.
6 — Write the appraisal
Fill the template: rating, each critical flaw with its quoted evidence, the non-critical weaknesses, and the implications — what the conclusions can still be used for, and what needs an independent check.
Mandatory: every item answered (no blanks, no "not applicable"), the conventions line, the rule applied, the implications. A rating without per-item evidence is not an appraisal.
Worked example
Review (illustrative): "Digital cognitive behavioural therapy for insomnia in adults: systematic review and meta-analysis of 12 randomised trials" (2024). PICO: adults with chronic insomnia; app-delivered CBT-I; sleep-hygiene education; sleep-onset latency and Insomnia Severity Index. From the report: items 1, 3, 5, 8, 9, 11, 12, 13, 14, 16 = Yes; 2 = Partial Yes (protocol agreed, not registered); 4 = Partial Yes (MEDLINE, Embase, PsycINFO, CENTRAL, but no registry or grey-literature search and a search 30 months old); 6 = No (one extractor); 7 = No (PRISMA counts only); 10 = No; 15 = No (12 trials pooled, no funnel plot). python3 scripts/amstar2.py rate --demo:
Implications: unlisted exclusions and unexamined small-study bias could each shift the pooled effect, so treat it as a lead and check the trials with assess-study-bias. The rating does not hinge on the convention — under --partial-yes flaw items 2 and 4 also become flaws and it stays Critically low; say so whenever it would change.
Verification
All 16 answered, each citing its evidence (unreported means No); Partial Yes only on 2, 4, 7, 8, 9; "No meta-analysis conducted" only on 11, 12, 15, consistently.
Critical domains and the Partial Yes convention were fixed before answering, and appear in the output.
Recompute with scripts/amstar2.py rate — flaws, weaknesses and the Box 2 rule must match the write-up; item 15 answered, not skipped, for any review with a meta-analysis; no numeric score anywhere.
Companion tool
scripts/amstar2.py (stdlib only) applies Box 1 / Box 2 mechanically: a per-item table (response, critical?, counts as), counted critical flaws and non-critical weaknesses, the rating with its rule trace, and warnings (inconsistent no-meta-analysis answers, lenient conventions). Judgements stay yours; it removes counting slips.
Do not rate a review by its journal, authors or "systematic" label — only by the 16 answers.
Do not treat "Partial Yes" as "Yes" — it marks partial adherence, a weakness by default.
Do not skip item 15 for a review with a meta-analysis — uninvestigated publication bias is a critical flaw.
Do not answer "not applicable" or "cannot tell"; AMSTAR 2 removed both — unreported means No.
Do not sum items into a score or percentage; the instrument's authors reject this explicitly.
Do not appraise primary studies with AMSTAR 2, report this rating as GRADE certainty, or change the critical domains after seeing the answers.
Reference
B. J. Shea, B. C. Reeves, G. Wells, M. Thuku, C. Hamel, J. Moran, D. Moher, P. Tugwell, V. Welch, E. Kristjansson, D. A. Henry, "AMSTAR 2: a critical appraisal tool for systematic reviews that include randomised or non-randomised studies of healthcare interventions, or both," BMJ, vol. 358, j4008, 2017. doi:10.1136/bmj.j4008 — item wording, Box 1 (critical domains), Box 2 (ratings); licensed CC BY 4.0.
AMSTAR 2 checklist and guidance document, 2017, https://amstar.ca/Amstar_Checklist.php — the authoritative response options and scoring criteria. Separately copyrighted and not reproduced here; consult it when an appraisal must be defended verbatim.
K. De Santis, D. Pieper, R. Lorenz, U. Wegewitz, W. Siemens, K. Matthias, "User experience of applying AMSTAR 2 …: a commentary," BMC Med Res Methodol, vol. 23, 63, 2023. doi:10.1186/s12874-023-01879-8 — Partial Yes weighting is a team decision.
Attribution and licence
The 16 item questions and the Box 1 critical-domain list used here and in references/items.md are reused from Shea et al., BMJ 2017;358:j4008 (doi:10.1136/bmj.j4008), licensed CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/. Changes made: the item text is condensed and reformatted, and the short labels, the Y/PY/N/NMA response codes, the per-item criteria, the overall-confidence descriptions and all surrounding guidance are this skill's own wording, not the original authors'. The criteria and rating descriptions in particular are restatements, not reproductions of the separately copyrighted AMSTAR 2 checklist. Neither the AMSTAR 2 authors nor BMJ endorse this skill; everything in it other than the CC BY material is MIT-licensed.