SuperAPI AI Directory

systematic-review

General & other

View repository

Updated 10-15 17:31

About this skill

Systematic Review

Strict PRISMA 2020 systematic review workflow. Fully implements Cochrane Handbook v6.5 standards across a 12-phase lifecycle, including PRISMA-S search documentation, double-blind screening + Cohen's κ, dual extraction, selection of RoB tools by study design, GRADE certainty of evidence, and AMSTAR 2 self-assessment.


1 What you produce and what you do not produce

1.1 Path A outputs (strict mode / default)

Deliverable Required elements
Main manuscript Full alignment with all 27 PRISMA 2020 items (six major IMRaD sections)
PRISMA flow diagram 4 phases + number excluded and reasons at each phase
PRISMA-S document 16-item search reproducibility YAML
Protocol statement PROSPERO/OSF registration number; if unregistered, explicit declaration + timestamp required
Inclusion and exclusion criteria table PICO three tiers + FINER
Screening record Double-blind κ + arbitration record
Extraction table SSOT-EXTRACTION-FIELDS six major categories + two-person agreement
RoB assessment table Select RoB 2 / ROBINS-I / QUADAS-2 / NOS / JBI by design (for single-case trials use SCED/RoBiNT, not SCRIBE)
Synthesis SWiM 9 items + Popay 4 elements (when meta-analysis is not possible)
GRADE certainty of evidence 5 downgrade + 3 upgrade + SoF table
AMSTAR 2 self-assessment 16 items across 7 critical domains, confidence rating
Writing process record SubAgent division of labor + double-blind disclosure + mode selection
Decision record decisions/01-11_*.md (apply §3.3)

1.2 Path B outputs (learning mode / course proposal)

Learning mode follows the simplifications in §9.2, but retains the three hard constraints of L1+L2+L11 + PRISMA flow diagram. See §9.3 for specific disclosures.

1.3 Not produced (boundaries)

  • Quantitative meta-analysis(point estimate pooling, forest plot, funnel plot, Egger's test) → switch to meta-analysis Skill for accompanying use
  • Scoping review(field mapping, concept map, PRISMA-ScR) → switch to field-overview-review
  • Narrative review(no double-blind, no PRISMA flow diagram) → switch to field-overview-review or field-evolution-review
  • Theory-building review(construct analysis, comparison of opposing theories) → switch to theoretical-review

See §13 for relationships with other Skills.

1.4 Standard deliverables list (standalone small table for at-a-glance reconciliation)

Main manuscript / PRISMA flow diagram / PRISMA-S YAML / protocol statement / PICO three-tier table / double-blind screening record / dual extraction table / RoB assessment table / SWiM+Popay synthesis / GRADE+SoF table / AMSTAR 2 self-assessment / writing process record / decision record directory (13 categories in total).


2 Triggers and applicable scenarios

2.1 Scenarios that should trigger

Key scenarios for positive triggering:

  • The user explicitly says "systematic review", "systematic review", "PRISMA review", "Cochrane-style review", "evidence-based review"
  • The submission target is a journal requiring PRISMA 2020 / Cochrane standards (BMJ, Lancet, Cochrane Library, JCE, the systematic review section of Acta Psychologica Sinica)
  • The user's description includes "double-blind screening", "RoB assessment", "GRADE rating", "AMSTAR 2", "PROSPERO registration"
  • Dissertation methodology chapters, doctoral qualification review, evidence-based clinical application (guideline development, HTA reports)

Potential trigger scenarios (this Skill should be recommended even if the user does not say PRISMA):

  • Userexplicitlyrequires PRISMA / Cochrane / RoB / GRADE / AMSTAR 2 procedures
  • User mentions "risk of bias", "level of evidence", "reproducible search"
  • The user's research question is in PICO form (explicit Population/Intervention/Comparison/Outcome)

2.2 Reverse triggers (Do NOT use)

User scenario Should switch to Skill Reason
"Explore research progress in a certain field", "do a literature review to understand the field" (no PRISMA keywords) field-overview-review Exploratory review, no double-blind, no RoB
"50-year history of research development", "field evolution" field-evolution-review Time dimension + 5-dimension search
"Theory construction, construct analysis" theoretical-review Argumentation-focused, loose methods
"Only do quantitative meta-analysis", no need for qualitative synthesis meta-analysis Go directly to L9 quantitative (e.g., if complete data are already available)

2.3 Ambiguous-zone routing (eliminating LS / SR overlap — W5-1 §1.3 questions 14/15)

Key routing rules:

  • User has already provided a literature set + requests PRISMA / Cochrane → use systematic-review (this Skill), skip L4-L5 search;do not use literature-set-review
  • User has already provided a literature set + requests "write a review based on the literature set", with no PRISMA requirement → use literature-set-review
  • User does not have a ready-made literature set + requests PRISMA → use systematic-review full L1-L12

2.4 Relationship with meta-analysis (joint workflow)

systematic-review and meta-analysis are complementary, not substitutes:

  • systematic-review provides all foundations for L1-L8 + L11-L12 (PICO, search, screening, extraction, RoB, reporting, self-assessment)
  • When the data can be meta-analyzed, L9 switches to the meta-analysis Skill for handling (see §4.2 L9 decision table)
  • Joint workflow (switch to MA): systematic-review completes L1-L8 → meta-analysis completes L9-L12(including GRADE / APA MARS reporting / AMSTAR 2, produced by MA 10_grade_decision.md / 11_reporting_standard.md / L12_amstar2.md; Lock 4 §4.4 ruling: in the joint pathway ending in meta-analysis, MA owns L9-L12).In the pure SR pathway (without switching to MA), SR completes L9-L12.

3 Main Agent roles and responsibilities

3.1 Main Agent required reading (required at startup — G5: startup footprint cut from the ≈72k three-piece set to a ≤10k digest)

After the Skill is triggered, the Main Agentimmediately loads review-methodology-foundations/references/LAUNCH-DIGEST.md(required at startup, main text ≤10k)——it distills the original "startup required three-piece set" (SSOT-LIFECYCLE-12-PHASES + SSOT-METHODOLOGY-TERMINOLOGY + SSOT-AI-FAILURE-DEFENSES, ≈72k) intoall load-bearing clauses + §1.2-18 frontier clauses + six lookup tables: Cohen / k→qualifier / I² / GRADE / discipline×database / construct direction. Reading this one document is enough to master the red lines of the full-process load.

Why the change (G5): actual-combat audits measured that 2/3 of the 66-72k "startup required" three-piece set was skipped——never opened;The read rate is a function of whether the main Agent names the prompt when dispatching it, not a function of what is "required reading". Therefore, a three-layer fix: ① At startup, read only a ≤10k digest (actually read); ② The three-piece set + deep SSOT full text are downgraded to "on-demand drill-down" (paths resolved via the SSOT-INDEX.md repository-wide resolver, see §8); ③ Critical payloads are delivered to the SubAgent via the PROMPT-COPY-BLOCKS.md copy block (§6 · the constructive guarantee that content enters the prompt), not by relying on self-motivation to read deep files.

SR-specific deep layers (not loaded at startup; located by the resolver, drilled down on demand):SSOT-PRISMA-S-TEMPLATE(L5)/ SSOT-PRISMA-FLOW-DIAGRAM(L6 / L11)/ SSOT-MODE-PROFILES(L3)/ SSOT-AMSTAR-2-CHECKLIST(L12) — the digest provides decision thresholds; when the full template / original entry text is needed, drill down via the resolver path.The critical payload is already in the digest / §6 copy block; the SubAgent does not need to open deep files on its own。

3.2 Coordination rather than execution

In this Skill, the main Agent is thecoordinator, and does not directly call MCP or directly write review sections. All concrete execution (retrieval, screening, extraction, writing, evaluation) is outsourced to SubAgents.

The main Agent's core actions:

  • Understand: Align with the user on the research question, review type, mode (strict/learning), word count, language, and citation format
  • Plan: Based on SSOT-LIFECYCLE-12-PHASES, break down phase tasks, identify dependencies, and decide serial vs parallel
  • Dispatch: Write explicit SubAgent prompts (including required-reading files + task + output contract + self-check)
  • Review: Check each phase's output against PRISMA 2020 + AMSTAR 2 self-checks; dispatch a Fix-Agent when problems are found
  • Organize: Persist cross-SubAgent documents to the working directory; finally the main Agent integrates the main draft

3.3 Cross-phase decision archive rules (mandatory, apply the C1 template)

After each phase's output, the main Agent writes decision conclusions into the decisions/ subdirectory. This is the cross-phase decision archive mechanism shared by the 6 Skills, preventing L9/L11/L12 SubAgents from re-deciding. The complete SR 11-category decision archive:01_type_decision.md(L1 type + 4-item judgment matrix) / 02_research_question.md(L2 three-layer PICO + FINER) / 03_mode_decision.md(L3 strict vs learning) / 04_search_strategy.md(L4)/ 05_search_execution.md(L5 PRISMA-S 16 items) / 06_screening_dual.md(L6 κ + arbitration) / 07_data_types.md(L7 extraction ICC/κ) / 08_rob_tool_decision.md(L8 RoB tool, SCED/RoBiNT rather than SCRIBE) / 09_synthesis_decision.md(L9 including switching MA thresholds) / 10_grade_decision.md(L10)/ 11_reporting_standard.md(L11)。

Each decision archive uses a unified YAML frontmatter + body + upstream/downstream dependencies + verification checklist structure; the complete template is in references/decision-provenance.md。

L9/L11/L12 SubAgents must read decisions/: All L9/L11/L12 SubAgent prompts must explicitly list decisions/01-11_*.md(corresponding fields by phase; see references/decision-provenance.md §3 "consumer-producer" mapping). Executing directly without reading the decision archive → blocked; when contradictions arise, the main Agent arbitrates.

Retroactive recognition ① (G5 · Axiom 8 good-deviation retroactive recognition) — planning document = context container: In practice, SubAgent prompts did not name the above decisions file by file (01/03 were generally absent); instead, the main Agent used a "dedicated planning document for each round" to carry context (with embedded decision-file pointers), which was more effective.The specification recognizes this pattern: The above required-reading decision files may be satisfied by "reading this round's planning document (with the required decisions/*.md pointer)" are equivalent and satisfied; constructive bottom line = §6 dispatch copy block writes the critical decision payload directly into the SubAgent prompt (fill-in-the-blank), without relying on the SubAgent to voluntarily open the decisions file.Decision records prohibit self-certifying "[ ] 已读 <文件>" check marks(G5 · R3 measured decisions/01 false "[x] read" has occurred, transcript zero access); "required reading" is changed to use the template frontmatter required_readings declarative machine-readable field, preflight --gate readings one-line WARN count, diff table → v2.

L11 Pre-writing main Agent secondary verification: before dispatch, the main Agent verifies that the current L11 SubAgent's planned reporting standard is consistent with 01_type_decision.md + 11_reporting_standard.md (SR must be PRISMA 2020 27 items; psychology scenarios additionally overlay APA MARS). If inconsistent,block。

3.4 Unacceptable compromises

  • In systematic-review, it is not allowed to "compress to PRISMA 2 items" — this is an AMSTAR 2 Critical Domain failure
  • It is not allowed to arbitrarily downgrade to English-language searching to replace Chinese-language searching (see §11.1 Search fallback: main path paper-search-pro / Chinese manual export of bibliographic records → ingest_search_exports.py ingestion)
  • It is not allowed to mistakenly use SCRIBE as a single-subject RoB tool (use SCED Scale / RoBiNT Scale, see SSOT-ROB-TOOL-MAPPING)

4 Complete workflow (12-phase alignment)

4.1 12-phase overview table

Each phase corresponds to one or more files under references/. It is enough to load the corresponding file at the current phase; there is no need to preload all of them. See §6 When to load which references file.

L# Phase name Key output Primary SSOT Primary references
L1 Type identification + 4-item decision matrix Type confirmation + mode recommendation + Skill routing SSOT-REVIEW-TYPES references/L1-question-construction.md
L2 Strict PICO construction Three-layer PICO + FINER + hypothesis trichotomy SSOT-PICO-FRAMEWORKS references/L1-question-construction.md
L3 Protocol preregistration PROSPERO/OSF registration number / status statement SSOT-REPORTING-STANDARDS §1.2.2 references/L2-protocol-registration.md
L4 Search strategy Database list + synonyms + Boolean + PRESS SSOT-SEARCH-STRATEGY references/L4-search-protocol.md
L5 Search execution PRISMA-S 16 items + deduplication record SSOT-PRISMA-S-TEMPLATE references/L5-prisma-s-execution.md
L6 Double-blind screening Agreement + Cohen's κ + arbitration SSOT-AI-FAILURE-DEFENSES §1 references/L6-screening-dual-blind.md
L7 Two-person extraction Structured fields + agreement + LLM disclosure SSOT-EXTRACTION-FIELDS references/L7-extraction-dual.md
L8 Risk of bias assessment Select tool by design + complete domain scoring SSOT-ROB-TOOL-MAPPING references/L8-rob-assessment.md
L9 Synthesis SWiM 9 items / Popay 4 elements / switch to MA SSOT-SYNTHESIS-METHODS references/L9-synthesis-swim.md
L10 GRADE certainty of evidence 5 downgrade + 3 upgrade + SoF table SSOT-GRADE-FRAMEWORK references/L10-grade.md
L11 Writing + reporting PRISMA 2020 27 items + flow diagram + APA MARS SSOT-REPORTING-STANDARDS references/L11-prisma-reporting.md
L12 Self-assessment AMSTAR 2 16 items + ROBIS SSOT-AMSTAR-2-CHECKLIST references/L12-amstar2-self-assessment.md

4.2 Detailed explanation of tasks in each phase

4.2.1 L1 type identification + 4-item decision matrix (W5_FIX_PLAN E.P0.4 revision)

The 4 decision elements of systematic-review + decision matrix(replacing the original hard condition of "if any one is not satisfied, switching is recommended"):

4 decision elements:

  1. The user's research question can be formalized as PICO (Population-Intervention-Comparison-Outcome)
  2. The user accepts the workload of double-blind screening + RoB assessment + GRADE
  3. Data availability can support structured extraction (not merely a discussion-level overview)
  4. User accepts a review length of 8,000+ words (learning mode can be reduced to 6,000-8,000)

Decision matrix(no longer "switch if any one is unmet"):

Number of passed items Decision Action
4/4 Strict mode Default full 12 stages + AMSTAR 2 ≥ Moderate target
3/4 Strict mode + explicit disclosure Full 12 stages, but explicitly disclose unmet items and their impact in Limitations. Example: Item 1 fails (weak PICO), disclose "the research question is close to but does not strictly satisfy the PICO format, affecting the boundaries of the Eligibility Criteria"
2/4 Learning mode + strong disclosure Switch to learning mode (§9.2); all simplification points must be written in according to the original disclosure template in §9.3. At the same time, advise the user to consider switching to FO/LS
1/4 Switch to FO/LS Do not do this within systematic-review. The main Agent explicitly explains the reason to the user and recommends an alternative Skill

Example:

  • Passed items = {1, 2, 3} (4 failed: user word budget 4000) → 3/4 → strict mode + explicitly disclose in the Discussion "under word count constraints, some synthesis paragraphs were simplified"
  • Passed items = {1, 3} (2/4: user refuses double-blind + word count 5000) → learning mode + double-blind downgrade disclosure + word count limitation disclosure + ask whether to switch to FO

Write the decision conclusion into decisions/01_type_decision.md。

Additional:FINER 5-item filter(see references/L1-question-construction.md §3 + SSOT-PICO-FRAMEWORKS §5).

4.2.2 L2 Strict PICO construction (Cochrane three levels)

systematic-review mandates three-level PICO (Cochrane v6.5 Ch.2 §2.3):

  • Review PICO: the research question at the review level
  • PICO for each synthesis: the PICO of each synthesis (e.g., subgroups of different Outcomes)
  • PICO of the included studies: the actual PICO of each included study

Deviations among the three levels must be explicitly reported in Limitations.

4.2.3 L3 Protocol preregistration (mandatory declaration)

Mandatory rule: even if not registered, it must be explicitly declared in the "Protocol and Registration" section of the main manuscript's Methods section (see the three templates in references/L2-protocol-registration.md §1.1/1.2/1.3).

Failure to declare protocol status constitutes a failure of AMSTAR 2 item 2 (Critical Domain 1), with a self-assessment of at most Low.

4.2.4 L4-L5 Search strategy and execution (PRISMA-S)

systematic-review strict mode must include:

  • 4+ discipline core databases (e.g., MEDLINE / PsycINFO / EMBASE / CENTRAL) + grey literature
  • Citation tracking (forward + backward citation)
  • PRESS peer review (recommended, Tier 2)
  • All 16 PRISMA-S items completed (YAML template)
  • Bramer EndNote 6-step deduplication + record the number of deduplicated records at each step

Learning mode may be downgraded (≥ 2 databases + simplified Boolean), butmust follow the original disclosure template in §9.3explicitly disclose simplification points.

Automatic prescreening (retroactive acknowledgment ④, G5 · Axiom 8): when recall-first searching yields a large number of hits and much noise (in practice ~80% without construct terms), one step may be added before double-blind screening:verification-based automatic prescreening(rough filtering by construct keywords, N_raw→N_prescreen)—but the three-part set must be followed:100% seed preservation + double false-negative spot check (0 false negatives) + explicit Methods disclosure, anddoes not replacedouble-blind (§4.2.5) and full-text two-person hard points. For the complete standard procedure see references/L6-screening-dual-blind.md §12。

Search execution outsourcing (G6 · outsourcing handoff package mode): L5's searchexecution(logging into subscription databases/pasting search strings/exporting records per database) is procedural external labor—the main session does not do it, and even if a SubAgent is dispatched, it does not do it either(attention economy). After L4 freeze sign-off, the main session generatessearch request package search_request_pack/(request.yaml machine-readable SSOT → render_search_pack.py render per_db/*.md human-readable card + export_contract.md + hit_log_template.md), outsourced to a pluggable executor (human/Codex/another CC session/pending search skill), and the receipt passes through scripts/validate_exports.py hard gate(reuse ingest parser dry-run + seed presence + hit consistency + no abstract disclosure U-49; exit≠0 means no ingestion) acceptance before handing over to ingest_search_exports.py. For per-database export pitfalls see references/db-quirks.md; for mode specification see review-methodology-foundations/references/OUTSOURCING-HANDOFF-PATTERN.md。contract boundary: the main session only produces the request package + acceptance receipt; "execution" is a pluggable executor contract (input request.yaml → output L5_export_<db>_<date>.<ext> + hit records + executed_date write-back).

4.2.5 L6 double-blind screening (key LLM defense — where the 10/10 exemplar resides)

LLM-implemented double-blind:

  • Dispatch 2 independent screening SubAgents(no shared context, independent prompts)
  • Each screens the same document set
  • Main Agent compares the two decision lists → computes Cohen's κ
  • Disagreements go to a thirdarbitration SubAgentfor adjudication
  • All decisions are recorded in screening_log.csv

Full-text screening = two-person hard point (U-16, upgraded to hard point on 2026-07-03):Both the title/abstract screening and full-text screening layers must be two-person independent—not only the title/abstract layer double-blind with the full-text layer single-person. Practical lesson: after T/A double-blind consumed 16 SubAgents,the full-text layer degenerated into single-person screening(no A/B, no full-text-layer κ), violating this specification and harming AMSTAR 2 Items. Therefore:full-text eligibility determination likewise dispatches 2 independent SubAgents + arbitration + computes full-text-layer κ(fulltext_kappa), recorded in L6_kappa_fulltext.json; the SP2 sign-off gate's 06a uses fulltext_kappa_json pointer to point to it (preflight additive validation, WARN at minimum if missing—see SUPERVISED-MODE-GATES §9.4). Methods disclosure must cover both layers being two-person (if a layer is actually single-person, honestly write single-person; do not claim two-person).

Explicit disclosure requirement (mandatory): In the main manuscript's Methods section, explicitly state "double-blind screening (both title/abstract + full-text layers) was performed by 2 independent LLM SubAgents + 1 arbitration SubAgent handling disagreements; each SubAgent could not see the other's decisions." For the complete disclosure text see references/L6-screening-dual-blind.md §6.1 (the only 10/10 exemplar in W5-4, ready to copy and use).

Cohen's κ thresholded decision (W5_FIX_PLAN E.P1.3):

κ value Decision
κ ≥ 0.60 (strict mode expectation) / ≥ 0.40 (learning mode expectation) Arbitrate disagreements only
0.40 ≤ κ < 0.60 Arbitration + re-review prompt (check inclusion/exclusion criteria ambiguity)
κ < 0.40 Redo screening(check SubAgent prompt isolation failure / major inclusion-exclusion criteria ambiguity)

4.2.6 L7 Dual Extraction (including LLM implementation disclosure)

Full-text acquisition + conversion outsourcing (G6 · outsourcing handoff package mode): After L6 inclusion determination and before L7 extraction, the included studies'full-text download + PDF→Markdown conversionare likewise procedural external labor—the main session does not do it. Generatefull-text request package fulltext_request_pack/(download list CSV + download_policy.yaml[Sci-Hub pre-approval + channel whitelist + dependency tier] + self-contained execution Prompt) and outsource it; the executor first paper-downloader-portable batch-downloads (channel whitelist explicit + Sci-Hub-type requires generation-time pre-approval + record actual channel per article, addressing U-08 non-traceability) → on failure, switch to manual per-database guidance → then document-to-markdown PDF→MD + extract figures + preserve page anchors (derive_anchors.py produces anchors.json). See the receipt format contract at meta-analysis/references/FULLTEXT-RECEIPT-FORMAT.md, via scripts/validate_fulltext.py hard gate(file presence/naming/conversion complete/anchors derivable/MD quality spot check: OCR structural dropped lines + meta noise, exit≠0 not ingested) acceptance. Channel discipline = the execution surface upstream of the public boundary red line (G10 aligned).

Consumption surface (addressing A-Q6 number-one token waste): extraction SubAgent consumes MD by default + returns to key PDF pages by anchor(p.N Table M granularity);reading the entire PDF is downgraded to an exception(only when anchors are degraded/key numerical values are ambiguous). Key numerical values/directions are always checked against 01_原始PDF/ (MD is OCR), and conclusions are accompanied by page/table anchors—this is a consumption discipline verified in practice.

systematic-review strict mode:

  • Dispatch 2 independent extraction SubAgents
  • By the six major SSOT-EXTRACTION-FIELDS categories (R/P/I/C/O/M) + X category (coder)
  • Inconsistent items enter arbitration + report ICC (continuous) / Cohen's κ (categorical)

LLM dual extraction disclosure text(apply C4 three-part format, replicating the L6 §6.1 template): for the complete three-part disclosure text, see references/L7-extraction-dual.md §"Honest Disclosure of Dual Extraction". The SubAgent copies it into the Methods section "Data Collection Process" when writing at L11.

Learning mode: single SubAgent extraction + main Agent 20% sampling review (see §9.3 for disclosure).

4.2.7 L8 Risk of Bias Assessment (select tool by study design)

Automatic routing by study design:

  • RCT → RoB 2 (5 domains + Cluster/Crossover variants)
  • Non-randomized intervention studies → ROBINS-I (7 domains)
  • Diagnostic accuracy → QUADAS-2
  • Cohort/cross-sectional → NOS (Newcastle-Ottawa Scale)
  • Qualitative research → CASP / JBI Qualitative
  • Measurement instrument review → COSMIN
  • Single-case design → SCED Scale or RoBiNT Scale(not SCRIBE — SCRIBE is a single-case reporting guideline, not a RoB tool, see SSOT-ROB-TOOL-MAPPING §11.2/§11.3)

The main Agent assigns in batches according to the included study types, and each study is scored across the full domain of the corresponding tool. See §10.3 for the default selection and disclosure of RoB tools.

4.2.8 L9 Synthesis — Switch-to-MA Threshold Decision Table (W5_FIX_PLAN E.P0.5 revision)

systematic-review defaults to the qualitative synthesis pathway at L9.Switch to meta-analysis decision table(replacing the original vague wording "k ≥ 5 and heterogeneity controllable"):

Data condition Decision Pathway
k ≥ 5 + I² < 75% + valid τ² estimate Recommend switching to meta-analysis systematic-review L1-L8 → meta-analysis L9-L12(after switching to MA, L10-L12 are completed by MA, lock 4 §4.4)
k ≥ 5 + I² ≥ 75% Keep SR L9 SWiM/Popay Do not pool effect sizes; follow §11.4 heterogeneity troubleshooting decision
k ≥ 5 + I² > 90% Pooling completely prohibited + mandatory subgroup / meta-regression See §11.4
k < 5 (any I²) Keep SR L9 SWiM/Popay Meta-analytic statistical inference is weak (Cochrane v6.5 Ch.10 §10.10.4)
Heterogeneous data types (continuous vs binary) / measurement instruments not convertible Keep SR L9 SWiM/Popay Not meta-analyzable
Qualitative evidence synthesis Thematic synthesis (Thomas-Harden 2008) / meta-ethnography (Noblit-Hare 1988) Follow GRADE-CERQual

Key constraint:Significance-based vote counting is prohibited(Cochrane v6.5 Ch.12 §6.5/6.6 — unacceptable method).

The L9 synthesis method decision is written to decisions/09_synthesis_decision.md。

4.2.9 L10 GRADE Certainty of Evidence

GRADE is assessed separately for each critical Outcome (Cochrane v6.5 Ch.14):

  • Starting certainty: RCT = High, NRSI = Low (may be exempted up to Moderate when ROBINS-I is enabled)
  • 5 downgrading factors: RoB / Inconsistency / Indirectness / Imprecision / Publication Bias
  • 3 upgrading factors: Large Effect / Dose-Response / Plausible Confounders Reduce
  • Output a Summary of Findings table (GRADEpro format)

Qualitative evidence synthesis follows GRADE-CERQual (4 dimensions: Methodological Limitations / Coherence / Adequacy / Relevance).

4.2.10 L11 Writing + Reporting (PRISMA 2020 27 items)

The main Agent integrates the outputs of each stage and writes the main manuscript. The structure corresponds to the PRISMA 2020 27 items:

  • Title (1) / Abstract (2) / Introduction (3-4) / Methods (5-15) / Results (16-22) / Discussion (23-26) / Other (27)

For psychology scenarios, overlay the 12 differences in APA JARS-Quant Table 9 (MARS).

PRISMA flow diagram (SSOT-PRISMA-FLOW-DIAGRAM) mandatory 4 phases: Identification → Screening → Eligibility → Included.

4.2.11 L12 Review Self-Assessment (AMSTAR 2 + ROBIS)

systematic-review mandates AMSTAR 2 self-assessment:

  • Item-by-item self-assessment of 16 items across 7 critical domains
  • Output a confidence rating (High / Moderate / Low / Critically Low)
  • Strict mode target ≥ Moderate (preferably High)
  • Not mandatory in learning mode (but encouraged)

ROBIS (risk of bias in the review itself) is a Tier 2 recommended item.


5 SubAgent Task Templates

See references/subagent-templates.md. Top-level mapping:

  • L1 type-identification-agent / L2 pico-construction-agent / L3 protocol-registration-agent (1 each, serial prerequisites)
  • L4 search-strategy-design-agent + press-review-agent (serial)
  • L5 search-execution-agent + dedup-agent + citation-tracking-agent (can be parallel)
  • L6 screening-agent ×2 double-blind + arbitration-agent (2 parallel + arbitration serial)
  • L7 extraction-agent ×2 two-person + extraction-arbitration-agent (same orchestration as L6)
  • L8 rob-assessment-agent (1 or multiple by design, can be parallel)
  • L9 synthesis-agent / L10 grade-assessment-agent (1 each, serial)
  • L11 writing-agent + reporting-agent + prisma-flow-agent + writing-process-archive-agent(4, serial; the latter two are newly added in W5_FIX_PLAN E.P1.5)
  • L12 amstar-self-assessment-agent + robis-agent(Tier 2)

Total ~16-18 SubAgents (double-blind/two-person extraction is the maximum parallel peak).


6 dispatch SubAgent copy blocks + references loading (G5: fill-in-the-blank delivery, not multiple-choice loading)

Mechanism (treat the root cause, same as §3.1): In practice, whether a reference is read is 100% determined by whether the dispatch prompt names it. Therefore, for payload-intensive stages (L6-L11),paste the entire copy block into the SubAgent prompt at dispatch time(review-methodology-foundations/references/PROMPT-COPY-BLOCKS.md §C, prefix its §A shared foundation block = LAUNCH-DIGEST §0/§2/§3) — key payload goes into the prompt (fill-in-the-blank), and whether the SubAgent reads the deep file no longer determines success or failure.The drill-down path is resolved via SSOT-INDEX.md full-library resolver(treats wrong-path; note L2-protocol-registration.md actually serves L3,L1-question-construction.md covers L1+L2).

Stage · Action Copy block / references Highest-risk payload for this stage (one sentence)
L1 type determination / L2 PICO construction references/L1-question-construction.md 4-item determination matrix; three-layer PICO; FINER 5 items
L3 protocol drafting / mode selection references/L2-protocol-registration.md (+ L3-mode-disclosures.md disclosure) Registration or explicit non-registration statement; protocol deviations must be recorded (HARKing prohibited)
L4 search strategy references/L4-search-protocol.md ≥4 databases + PRESS; subject terms + free text dual track
L5 search execution references/L5-prisma-s-execution.md PRISMA-S 16 items; Bramer 6-step deduplication; record hit counts
L6 double-blind screening ×2 + arbitration COPY-BLOCKS §C·L6 Full-text screening = two-person hard point; κ≥0.60 arbitration / <0.40 redo; automated pre-screening standard process (ratification ④, §4.2.4)
L7 two-person extraction ×2 + arbitration §C·L7 Two-person κ≥0.7 / ICC; full coverage of 8 categories of information + reliability fields
L8 RoB assessment §C·L8 Tool × design must match (RCT→RoB2 / NRSI→ROBINS-I / single-subject→SCED not SCRIBE / single-group pre-post→ROBINS-I UBA); two levels: study level × outcome level
L9 synthesis SWiM/Popay §C·L9 Ban on significance-based vote count(Cochrane); PRISMA+SWiM 9 items; §4.2.8 switch-to-MA decision table
L10 GRADE §C·L10(L10-grade.md resides in this skill) "few studies" ≠ imprecision (use OIS); before upgrading, first 5 downgrades; SoF≤7
L11 writing ×4 §C·L11 PRISMA 27 items + 24-27 not omitted; flowchart 4 stages; Limitations ≥4 dimensions
L12 self-assessment references/L12-amstar2-self-assessment.md AMSTAR 2 16 items 7 critical domains; any lost point = Critically Low
Write/check decision archive references/decision-provenance.md + COPY-BLOCKS §D Planning document = context container (ratification ①, §3.3);Prohibit self-certifying "read" checkboxes
Before any Review-Agent references/quality-control.md Cross-stage cross-reconciliation (§14.3) + Four Iron Rules

subagent-templates.md = complete 6-segment templates for each role (on-demand deep dive; the copy block = a ready-to-use compression of its "required-reading segments", so the main Agent can dispatch without first opening the template).


7 Word Count Control

For the word count formula and matrix, see review-methodology-foundations/references/SSOT-WORDCOUNT-RULES.md §3.1 (review type × literature scale matrix row 4: systematic review) + §3.3 (adjustment factors) + §4 (word count discipline). This section does not duplicate the SSOT numbers.

  • Default word count: 12,000 words (SSOT §2.2 systematic review row);lower bound 8,000 words
  • Three confirmation modes: A (user has specified → adopt directly) / B (not specified → default 12,000 + notify + non-blocking) / C (adaptive → estimate per SSOT §3.1 matrix after L7 is complete)
  • Strict vs. learning mode: both follow the SSOT §3.1 matrix; in strict mode the lower bound is 8,000 words
  • Word count discipline: no padding / no verbiage / quality first (see SSOT §4 for detailed definitions)
  • Severe word count mismatch threshold: see §11.6 (apply C2)

8 When to load the foundation SSOT (path authority = repository-wide resolver)

Sole path authority = review-methodology-foundations/references/SSOT-INDEX.md(repository-wide file-resolver): stage/keyword → the one correct path, covering all references of the three skills (to fix the two real-world wrong-path incidents; for the cross-skill attribution trap see its §B). When drilling down through the full text on demand, use the resolver to resolve paths; do not guess paths by intuition from stage numbers. Core SSOTs for each SR stage:

Stage / Decision Point SSOT
Immediately after startup (required reading for the main Agent) LAUNCH-DIGEST.md(≤10k, required reading immediately at startup; distilled three-piece payload + §1.2-18 frontier clause six tables); the full three-piece set (AI-FAILURE / LIFECYCLE / TERMINOLOGY) is downgraded to on-demand deep dive
L1 Type decision SSOT-REVIEW-TYPES
L2 PICO three layers SSOT-PICO-FRAMEWORKS
L3 Mode selection SSOT-MODE-PROFILES
L4 Search strategy SSOT-SEARCH-STRATEGY
L5 PRISMA-S SSOT-PRISMA-S-TEMPLATE
L6 Flowchart preparation SSOT-PRISMA-FLOW-DIAGRAM
L7 Extraction fields SSOT-EXTRACTION-FIELDS
L8 RoB tool mapping SSOT-ROB-TOOL-MAPPING
L9 Synthesis methods SSOT-SYNTHESIS-METHODS (if switching to MA, add SSOT-EFFECT-SIZE-DECISION)
L10 GRADE SSOT-GRADE-FRAMEWORK
L11 Reporting standards SSOT-REPORTING-STANDARDS (for psychology, add SSOT-PSYCHOLOGY-SUPPLEMENTS)
L12 Self-assessment SSOT-AMSTAR-2-CHECKLIST
Psychology scenarios SSOT-PSYCHOLOGY-SUPPLEMENTS
Chinese-language scenarios SSOT-CHINESE-CONTEXT
Word count SSOT-WORDCOUNT-RULES

9 Strict / Learning / Supervised Modes

See review-methodology-foundations/references/SSOT-MODE-PROFILES.md v2.0.0 (systematic review subset §4.4 + three-mode definitions §1).

Three-mode overview:Strict(top-journal submission, fully automated + one final human acceptance) / Learning(coursework, simplified + sampling checks) / Supervised(§9.4: publication-grade methodological completeness + human sign-off at each of the 7 irreversible points). See §9.4 for supervised mode.

9.1 Strict Mode (Default)

Goal: top-journal submission, evidence-based clinical practice, Cochrane standards, AMSTAR 2 ≥ Moderate.

Execute all 12 stages in full, all Tier 1 (25 items). Tier 2 recommended: PRESS, AMSTAR 2 self-assessment, Open Science disclosure.

9.2 Learning Mode (user-declared or scenario-determined)

Goal: coursework assignments, thesis proposals, learning methodology, time budget ≤ 1 week.

Simplified 11-item Tier 1:

  • Retained: L1 / L2 / L4 (≥ 2 databases) / L5 (simplified Boolean) / L6 (key fields) / L8 (simplified 3-4 item RoB) / L9 (basic synthesis) / L11 (PRISMA core 8-10 items + flow diagram)
  • Skipped but disclosed: L3 (declare unregistered) / L4 (PRESS) / L6 (double-blind downgraded to single SubAgent + sampled manual verification) / L10 (GRADE not mandatory) / L12 (AMSTAR not mandatory)

Learning mode ≠ non-compliant. Three hard constraints:

  1. Retain L1+L2+L11 (type identification, PICO, write according to reporting standards)
  2. Retain PRISMA flow diagram
  3. At the end of the review, "Method Limitations" lists each simplification point item by item according to the §9.3 disclosure text template

For mode determination rules, see SSOT-MODE-PROFILES §2.1 (user scenario → recommended mode decision table) + §4.2.1 L1 4-item determination matrix.

9.3 Honest disclosure text template for the 5 simplification points of Learning Mode (applying the C4 three-part format, W5_FIX_PLAN E.P0.6)

Learning mode allows downgrading at 5 simplification points, but this must be explicitly disclosed in the main manuscript Methods / Limitations. Each disclosure paragraph contains three parts: textbook original assumption / LLM implementation differences (list ≥ 3 differences) / mitigation measures + Limitations disclosure.

The complete three-part disclosure text template is stored in references/L3-mode-disclosures.md, and the main Agent loads this file during the L11 writing stage and copies the corresponding template. The 5 simplification points:

# Simplification point Embedded main manuscript location Textbook basis
1 Skip PRESS peer review (L4) End of Methods §4.3 Search Strategy Cochrane v6.5 §4.4.5
2 Double-blind screening downgrade (L6) End of Methods §4.4 Selection Process Cochrane §4.6 + AMSTAR 2 Item 5
3 Skip GRADE (L10) Methods §4.11 Certainty Assessment + Discussion Limitations Cochrane v6.5 Ch.14
4 Skip AMSTAR 2 self-assessment (L12) End of Discussion Limitations Shea et al. 2017 BMJ
5 Protocol not registered (L3) Methods §4.0 Protocol and Registration Cochrane v6.5 + AMSTAR 2 Item 2 (CD 1)

For the complete three-part disclosure text for each simplification point ("Part 1 textbook original assumption / Part 2 LLM implementation differences ≥ 3 / Part 3 mitigation measures"), see references/L3-mode-disclosures.md. The Learning Mode main Agent must embed the corresponding template one by one for the simplification points actually triggered by this review (omitting Part 2 with only "Part 1 + Part 3" is not allowed).

9.4 Supervised Mode

In one sentence: AI does all the grunt work (search execution, double-blind screening, dual extraction, first draft), and humans only make the call at the 7 key decision points that machines cannot judge—methodology is as complete as in Strict Mode (all 27 PRISMA 2020 items / AMSTAR 2 ≥ Moderate), but every irreversible decision is personally signed off and owned by a human. For the definition, see review-methodology-foundations/references/SSOT-MODE-PROFILES.md §1.4。

When to choose it: What professors/advisors want—"AI first edit, human check, ensure no errors, publication-grade"—they are neither comfortable letting AI run fully automatically through all key judgments, nor can they use course-level random spot checks. When one researcher + AI collaborate to produce a submittable systematic review but must personally gatekeep each point, choose Supervised Mode.

How it differs from Strict Mode: Methodological completeness is exactly the same (both are publication-grade Tier 1 full coverage); the only difference isReplace "human final acceptance once" with "each of the 7 irreversible points reviewed once". On the systematic review side, the first few sign-off points mainly fall on L4 retrieval + inclusion/exclusion freeze, L6 screening disagreements, L7 extraction calibration/low confidence (subsequent L9/L11 interpretation and citation final review are taken over by the meta-analysis pathway when switching to the meta-analysis endpoint, see §13.1). Each point produces a review card customized for it, and a human looks at the card and decides "accept or send back".

Why it is a "true gate" rather than "just a reminder": the verdict of each sign-off point lands in the decisions/ of the corresponding file's user_signoff marker. Before a downstream stage starts, the preflight script reads it—no signature / pending / sent back → script exit code ≠ 0, and the downstream cannot start at all. At the same time κ is computed by the script (not by AI mental math), and reproducibility hard checks (missing sessionInfo / exit code ≠ 0 / κ too low / insufficient seed recall) directly block before human review.The runtime operational contract of the 7 gates (trigger timing of each gate, artifact paths, preflight commands, which downstream is blocked by a missing signature, hard checks, review-surface upper limits) see review-methodology-foundations/references/SUPERVISED-MODE-GATES.md。

Balance (neither side overfits): do not block with human review everywhere—double-blind concordant items are automatically released, if κ is judged pass and there is no red light then that point has no human review queue, and reversible intermediate artifacts are not gated; nor let AI run past key judgments—7 irreversible freeze points, screening disagreements/extraction low confidence/key numerical values must be reviewed. Each sign-off point has an upper limit on the number of items awaiting review; when exceeded, tighten upstream filtering rather than dumping a long list on a human.

Sign-off experience (G1 vNext, 2026-07-03 · two-stage): during run = flow-first informed confirmation (retain the "(recommended)" pre-mark + rationale + consequences + drill-down + send back; no negative-sample injection, no added test friction), while the truly independent human review is in the post-hoc signatory audit package (L12.5).Card/gate unified(the only human-facing document = gate markdown six sections a-f, HTML card is an optional enhancement); judgment points are driven by a single source of structured decision blocks,signoff_ask.py is the only compliant question/backfill channel (agents may not hand-write signed); preflight additive enforces same-source hash / complete evidence / narrative self-consistency / three fields. Full mechanism see review-methodology-foundations/references/SUPERVISED-MODE-GATES.md §0.5。

Disclosure: supervised-mode outputs must disclose "which steps AI did, and what humans reviewed at each of the 7 sign-off points" (RAISE: human oversight + transparent disclosure + ultimate responsibility with humans).Disclosure strength is machine-read generated from sign-off facts (presented_via/evidence_opened/signature page status), monotonically does not exceed the facts, and semantically uses attestation; the two-stage is written truthfully as:in-run informed confirmations with drill-down access + post-hoc independent audit via structured audit package. See SSOT §6.6 for details.

Responsible use (single SSOT): the stance of supervised mode, the two-stage oversight model, and the complete statement of the 7 gates and honest boundaries see release repo root RESPONSIBLE_USE.md—core: AI first drafting + humans sign off at the 7 gates + signatories bear full responsibility; during run = informed confirmation (not independent judgment), post-hoc audit package = truly independent human review; disclosure strength is machine-read generated and monotonically does not exceed the facts; automatic submission without human review is not supported.


10 Special handling for psychology scenarios and Chinese scenarios

10.1 Psychology scenarios

Trigger condition: according to SSOT-PSYCHOLOGY-SUPPLEMENTS.md §0.4 4 categories of signal checklist (APA journals / scale names / psychological constructs / replication crisis + JARS keywords), if any 1 category is hit then triggered. The judgment result (yes/no/boundary) is written into decisions/01_type_decision.md psychology judgment field (determines whether L11 superimposes JARS-Quant).

See SSOT-PSYCHOLOGY-SUPPLEMENTS.md. Additional requirements for SR psychology scenarios:

  • APA JARS-Quant Table 9 (MARS) 12 discrepancies (SSOT-REPORTING-STANDARDS §3)
  • APA Thesaurus (PsycINFO) instead of MeSH
  • Psychology meta-analysis 8 issues (overlaid when L9 switches to meta-analysis)
  • Reproducibility crisis response: preregistration statement, p-curve, Many Labs (SSOT-EXPERT-METHODOLOGY-INSIGHTS dimension 5)

Switch JARS sub-standards by synthesis pathway: L9 quantitative synthesis → JARS-Quant Table 9 (MARS); L9 qualitative synthesis (SWiM / thematic synthesis / meta-ethnography) → JARS-Qual Table 2 + ENTREQ; mixed methods → JARS-MM Table 1。

Journal submission mapping:Psychological Bulletin / Psychological Review → JARS-Quant Table 9 + APA 7 + Open Science;Acta Psychologica Sinica / Advances in Psychological Science → refer to SSOT-CHINESE-CONTEXT.

10.2 Chinese scenarios

See SSOT-CHINESE-CONTEXT.md. Special requirements for Chinese SR: CNKI search syntax + subject terms; Chinese core journal review format (Acta Psychologica Sinica, Advances in Psychological Science, Management World); GB/T 7714-2015 citation format (vs APA 7); preregistration optionally "PRECISE-CHINA" + PROSPERO Chinese version.Chinese search fallback path see §11.1(search main path paper-search-pro; Chinese 0 hits → manual export of CNKI/Wanfang bibliographic records → ingest_search_exports.py ingestion; key constraint: do not arbitrarily downgrade to English search in place of Chinese search).

10.3 Honest disclosure of default RoB tool selection (applies C4, W5_FIX_PLAN E.P1.4)

The main Agent at the L8 stage uses an RoB tool selection strategy that automatically routes by study design (textbook basis: Cochrane v6.5 §8 / §25 / §28), and needs to embed a complete three-paragraph disclosure in the "Study Risk of Bias Assessment" subsection of the main manuscript Methods section (textbook assumptions / LLM implementation differences ≥ 3 places / mitigation measures + Limitations disclosure). For the full disclosure text template see references/L8-rob-assessment.md §"Honest Disclosure of Default RoB Tool Selection".

Key constraint: single-case designs use the SCED Scale or RoBiNT Scale (not SCRIBE — SCRIBE is a single-casereporting guideline, not an RoB tool). See SSOT-ROB-TOOL-MAPPING §11.2/§11.3.


11 Fault diagnosis and fallback (applies C2 template)

This chapter is organized by the trio of "diagnostic signal + threshold + fallback path". Failover decisions are written to decisions/<阶段>_failure_switch.md。

11.1 Search main path and fallback (including Chinese scenarios)

Search main path = paper-search-pro Skill(OpenAlex / Semantic Scholar / CrossRef / PubMed all use HTTP APIs, requiring configuration of the corresponding API key, with no MCP server needed).If the runtime environment does not provide semantic-scholar / openalex MCP server (common default case) — fault diagnosis is based on "paper-search-pro abnormal search results",notbased on "SS/OpenAlex MCP errors" (such errors will not occur in that environment).Chinese searchbecause OpenAlex/SS/CrossRef/PubMed have limited coverage of Chinese literature, main path = user manual export of CNKI/Wanfang → scripts/ingest_search_exports.py ingestion.

Fault signal Fallback path
English search 0 hits / too low recall (keywords in English,paper-search-pro clearly insufficient hits) Expand synonyms / relax Boolean logic and re-search; cross-source crossing (OpenAlex↔CrossRef↔PubMed) to supplement recall
Chinese-language search via paper-search-pro 0 hits (0 hits + keyword is Chinese) (a) User manually exports CNKI / Wanfang / VIP bibliographic records → scripts/ingest_search_exports.py ingests unified fields; (b) when a literature set already exists, convert to literature-set-review。Key constraint: SubAgent must not arbitrarily downgrade English search to replace Chinese search
Search entirely unavailable (HTTP persistent failure / network restricted / key invalid, persistent failure ≥ 3 times) Block + warn user + guide manual bibliographic record export then proceed via ingest_search_exports.py + Do not forcibly downgrade to local reasoning to fabricate literature from memory
Search result metadata incomplete (abstract missing ≥ 50%) Dispatch firecrawl-abstract Skill for supplementary extraction (see downgrade path in note below) OR expand synonyms and re-search

firecrawl-abstract Downgrade note:If the runtime environment has no semantic-scholar MCP server → this Skill's paper_details(Semantic Scholar API / SS method) channelunavailable; by default use Firecrawl Scrape / Firecrawl Extract two methods(available when Firecrawl key is present), fetch by DOI / URL and parse abstracts.

11.2 Extreme values in search reporting (W5_FIX_PLAN E.P1.2)

  • identify_count > 50000 → Forced downgrade(not allowed to continue as SR), narrow PICO or switch to FO
  • identify_count > 10000 → strongly recommend downgrade (narrow PICO / switch to FO)
  • identify_count ∈ [5, 20) → warn + ask user whether to expand PICO or switch to LS
  • identify_count < 5 → block + switch to LS or expand PICO

Cite and report per PRISMA-S Item 15-16.

11.3 Severe data insufficiency

  • L7 data extractable fields < 50% → main Agent arbitration; if not possible → switch to FO
  • Studies containing reliability data < 5 (psychological meta-analysis) → reliability generalization sensitivity; Hunter-Schmidt warning + sampling correction
  • §4.2.1 4-item decision matrix = 1/4 → switch to FO/LS

11.4 Extreme heterogeneity / synthesis failure (W5_FIX_PLAN F.P1.4 coordinated)

  • I² > 75%(combined blocking) → not allowed to merely pool effects; use subgroup / meta-regression instead
  • I² > 90%(extreme) → pooling completely prohibited + must use subgroup / meta-regression / SWiM
  • k < 5 High heterogeneity → no meta-analysis, switch to SWiM/Popay

Apply §4.2.8 L9 switch MA decision table for execution.

11.5 Double-blind screening κ anomaly + 11.6 Severe word count mismatch

  • κ anomaly(see §4.2.5 for details): κ ≥ 0.60 arbitrate only on inconsistencies; 0.40 ≤ κ < 0.60 arbitrate + re-review criteria ambiguity; κ < 0.40 redo screening。
  • Severe word count mismatch(SSOT-WORDCOUNT-RULES §8): deviation ≥ 30% report to user + suggest adjustment; deviation ≥ 50% force adjustment or switch Skill。

12 Output format and file naming

Working directory structure <task_root>/工作产出/:

  • decisions/01_type_decision.md ... 11_reporting_standard.md(11 categories of decision archives)
  • L1_type_identification.md / L2_pico_construction.md / L2_synonym_groups.md / L2_finer_check.md
  • L3_protocol_draft.md / L3_registration_statement.md
  • L4_search_strategy.md / L4_database_list.md / L4_grey_literature_plan.md
  • L5_search_records_<database>.csv / L5_search_documentation_<database>.yaml / L5_dedup_log.md / L5_final_records.csv
  • L6_screening_log.csv / L6_arbitration_log.md / L6_excluded_studies_with_reasons.csv
  • L7_extraction_<agentID>.csv / L7_extraction_final.csv / L7_extraction_consistency.md
  • L7_to_MA_field_mapping.md(SR L7 → MA L9.1 field mapping, W5_FIX_PLAN E.P1.6)
  • L8_rob_assessment.csv / L8_rob_summary_table.md / L8_traffic_light_plot.md
  • L9_synthesis_results.md
  • L10_grade_assessment.csv / L10_summary_of_findings.md
  • L11_manuscript.md(main manuscript) / L11_prisma_flow_diagram.md(produced by prisma-flow-agent) / L11_writing_process_archive.md(produced by writing-process-archive-agent)
  • L12_amstar2_self_assessment.md / L12_robis_assessment.md(strict mode Tier 2)
  • final_quality_check.md

File naming rules:L<N>_<task>.<ext>(parallel to LS PRES-016 NN_ prefix style).

Pure SR path L12 delivery packaging (G9 pointer): The delivery packaging of the pure SR final draft follows the same specification as the joint path—① run manuscript-typeset produces the four submission items by default(docx / PDF man / PDF jou / clean md, without waiting for the user to add them; see §13.4 for the four-tuple); ② produce per the manifest five-section template 03_交付/00_交付清单_manifest.md(the template = meta-analysis/references/delivery-manifest-template.md,lives in meta-analysis and is referenced by this Skill; §3a covers item by item the PLACEHOLDERS.md produced in ①, and §1 registers the actual number of produced items); ③ with the L12.5 review package's 00_审核入口.md separate and cross-linked, not merged(manifest = submission item index; review entry = the signatory's final review flow). For the complete five steps see meta-analysis SKILL §12.4 (the joint path L12 belongs to MA, Lock 4 §4.4).


13 Relationship with other Skills

13.1 Companion relationship (strongly recommended combination)

  • systematic-review + meta-analysis: systematic-review completes L1-L8,meta-analysis completes L9-L12(quantitative synthesis + GRADE + APA MARS report + AMSTAR 2; Lock 4 §4.4: the joint path ending in meta-analysis has MA own L9-L12). The standard combination for submitting a top-journal systematic review + meta-analysis.Joint workflow routing: systematic-review L1-L8 → apply §4.2.8 L9 switch-to-MA decision table → meta-analysis L9-L12.
  • systematic-review + ma-plots renderer / ssci-plots: forest plots and funnel plots belong to the meta-analysis path (the joint workflow in the previous item, produced by the built-in renderer in meta-analysis §13.2);the PRISMA flow diagram / RoB figure of the pure SR path L11 = ma-plots fixed renderer fixed commands(prisma_flow.py / rob_traffic.py + QC exit≠0, the renderer lives in meta-analysis/scripts/plots/ and is referenced by this Skill, see §13.4 for the four-tuple); ssci-plots = style foundation + general statistical figure output other than the four-item set.

13.2 Alternative relationship (switch when the user chooses wrongly)

User scenario Should switch to Reason
Only wants to explore the field, not strict PRISMA field-overview-review Exploratory review, no double-blind, no RoB
Wants to do a 50-year development history, field evolution field-evolution-review Time dimension + 5-dimension search
Already has a literature set, no further searching,no PRISMA requirement literature-set-review Skip L4-L5, add input quality self-audit
Already has a literature set,requires PRISMA / Cochrane systematic-review (this Skill) Skip L4-L5 search, retain L1-L3 + L6-L12 (eliminate the gray zone with LS)
Theoretical construction, construct analysis, comparison of opposing theories theoretical-review Argumentation-focused, loose methods, Psychological Review style
Only does quantitative meta-analysis meta-analysis Go directly to L9 quantification (e.g., already has complete data)

13.3 Downgrade path

If the user initially wants to do a systematic review but finds it infeasible (e.g., search return volume > 50000, extremely high research heterogeneity, insufficient literature), the main Agent should recommend a downgrade:

  • downgrade to field-overview-review(integrated review path)
  • or downgrade to literature-set-review(driven by the literature set at hand)
  • Not allowed to "compress to PRISMA 2 items" within systematic-review

13.4 Satellite dependency matrix (four-tuple hard contract —— G9)

Positioning: same as meta-analysis SKILL §13.4——each satellite invocation is locked intoquadruple command-line contract[when to invoke / command or entry point / IO contract / fallback when missing], replacing narrative "linkage" (E-Q4 root cause: narrative dependencies can be legally bypassed).Each line's "fallback when missing" gives an explicit block/degrade action; silence is forbidden. Path = skill-root-relative path;<skill> = that skill's root directory (strictly forbidden runtime ~/.claude/skills/ and bare relative skills/)。

satellite when to invoke command or entry point IO contract fallback when missing
paper-search-pro L4/L5 retrieval main path (execute retrieval after L4 freeze sign-off) Skill tool explicitly loads and triggers (HTTP API: OpenAlex / Semantic Scholar / CrossRef / PubMed, keys already configured, no MCP) In: search query. Out: bibliographic export (BibTeX/RIS/CSV) → scripts/ingest_search_exports.py ingestion; on the outsourced route, the receipt first passes the scripts/validate_exports.py hard gate (exit≠0 no ingestion, §4.2.4) → §11.1 fallback table(Chinese 0 hits → manual export from CNKI/Wanfang → ingest ingestion; wholly unavailable → block warning, do not fabricate literature from memory) + G6 retrieval request package manual route (§4.2.4). This line's pointer suffices; do not restate §11.1
document-to-markdown + paper-downloader-portable After L6 inclusion decision, before L7 extraction (full-text download + PDF→MD conversion) G6 full-text request package mechanism (do not call CLI directly): main session produces fulltext_request_pack/(download_list CSV + download_policy.yaml + self-contained execution Prompt) outsourced to a pluggable executor (§4.2.6; mode = review-methodology-foundations/references/OUTSOURCING-HANDOFF-PATTERN.md) In:download_list.csv + download_policy.yaml(allow_scihub one-time prior nod / channel whitelist of 8 enumerations / dependency tier). Out: receipt per meta-analysis/references/FULLTEXT-RECEIPT-FORMAT.md(8 fields fulltext_receipt.csv + human-readable list;resides in meta-analysis and is referenced by this Skill,resolver §B #9);scripts/validate_fulltext.py hard gate acceptance(exit≠0 no ingestion;--receipt --policy channel discipline fail-closed) G6 dependency tier(download_policy.downloader.tier): local paper-downloader-portable preferred; missing → third-party download skill (selected by G10 dependency matrix, tier placeholder) or manual-only manual per-database guidance. d2m (required converter for the execution Prompt) missing → switch to an executor that has d2m, or the missing conversion item truthfully goes into the receipt + main session blocks consumption and discloses——the main session does not fall back to downloading/converting by itself
ma-plots(resides in meta-analysis/scripts/plots/ referenced by this Skill, not an independent skill) Pure SR path L11's PRISMA flow diagram / RoB diagram (joint-path diagrams belong to MA L11, see §13.1) python3 <meta-analysis>/scripts/plots/prisma_flow.py --counts <prisma_counts.yaml> --out <stem>;python3 <meta-analysis>/scripts/plots/rob_traffic.py --data <rob_domains.csv> --labels <labels.csv> --tool both --out <stem>(--tool may also rob2 / robins;CLI signature authority = meta-analysis/scripts/plots/README.md) In:meta-analysis/references/prisma-flow-contract.md(counts YAML)/ rob-figure-contract.md(domains CSV). Out:<stem>.png + <stem>.pdf; exit codes 0/1/2/3,QC exit≠0 (code 2 is not warn-only) does not persist to disk。labels mechanical same-source disciplinesame as MA §13.2:--labels mechanically generated from the feature table / ID crosswalk card and persisted to disk,No manual fabrication; delivery acceptance spot-check rows must verify "display name ↔ study_id ↔ value three-way consistency" Renderer or its dependencies missing → block the delivery figure and explicitly report what is missing(no silent,never hand-write rendering code on the spot to substitute); missing R / official package (PRISMA2020 / robvis)→ exit 1 with explicit error(--warn-only no downgrade) → rerun after installing packages
manuscript-typeset(independent general skill) pure SR path L12 final manuscript delivery packaging (joint path belongs to MA §12.4);by default produce the four submission files, without waiting for the user to add python3 <manuscript-typeset>/scripts/typeset.py --request typeset_request.yaml --outdir <dir> in:typeset_request.yaml(contract authority = manuscript-typeset/references/MANUSCRIPT-INPUT-CONTRACT.md). Out: four files (submission docx / submission PDF man / publication preview PDF jou / clean md) + BUILD_NOTES / PLACEHOLDERS / STRIP_REPORT. Exit codes 0/1/2/3;fidelity fail = exit 2, that file not written + block (no --warn-only) missing xelatex → downgrade to produce docx + md two files (exit 0 + BUILD_NOTES declaration); missing pandoc → exit 1 block and report;fidelity fail never downgrade
academic-ref-check citation final review (after L11 final manuscript + citation table finalized, before delivery); pure SR final manuscript likewise 100% pass verification(academic-reference-verification Rule); supervisory mode joint path = SP7 gate python3 <academic-ref-check>/scripts/verify_http.py --in L11_refs.json --out-dir decisions/ --topic "终稿引用"(or --in-text L11_references.md self-parse); wiring details = meta-analysis/references/L11-mars-reporting.md §8.3 in: all citations in the final manuscript. Out:decisions/L11_ref_verify_report.md(human-readable) + .json(machine-readable); verify item by item existence/first author surname/journal/year + add back DOI. Exit codes: any red → 1 / otherwise unverified → 3 / otherwise any yellow → 2 / all green → 0 lack of verification capability ≠ pass (fail-closed): switch to manual multi-source verification (OpenAlex / CrossRef manual lookup) and explicitly disclose the burden; unverified must be rerun until judgeable;never "skip verification and deliver anyway"

14 Skill-specific methodology chapter

14.1 Honest disclosure of learning mode simplifications

The complete C4 three-part disclosure of the 5 learning mode simplifications is in §9.3.

14.2 Honest disclosure of default RoB tool selection

The disclosure of automatic routing by study design is in §10.3 (including SCED/RoBiNT vs SCRIBE correction).

14.3 Cross-stage cross-reconciliation (PRES-004 multi-perspective verification retained)

For the cross-stage reconciliation mechanism (to prevent omissions and errors), see references/quality-control.md §2-§3:

  • L2 → L4 / L4 → L5 / L5 → L6 / L6 → L7 / L7 → L8 / L8 → L10 / L9 → L10 / L10 → L11 / L11 → L12

15 Reference document list

15.1 Textbook foundation:Cochrane Handbook v6.5(2024)/ Cooper 5e Research Synthesis and Meta-Analysis(2017)/ Higgins et al. Cochrane Handbook / Borenstein et al. Introduction to Meta-Analysis 2e (2021, accompanying meta-analysis Skill) / JBI Manual for Evidence Synthesis (2024, qualitative evidence synthesis).

15.2 Reporting standards:PRISMA 2020(Page et al. 2021 BMJ, 27 items + 4-stage flowchart) / PRISMA-S (Rethlefsen et al. 2021 Syst Rev, 16 items) / PRISMA-P (Moher et al. 2015, 17-item protocol preregistration) / AMSTAR 2 (Shea et al. 2017 BMJ, 16 items across 7 critical domains) / GRADE Working Group (Cochrane Ch.14 + GRADE Handbook) / ROBIS (Whiting et al. 2016 J Clin Epidemiol)。

15.3 Risk of bias tools: RoB 2 (Cochrane v6.5 Ch.8, RCT 5 domains) / ROBINS-I (Ch.25, NRSI 7 domains) / QUADAS-2 (diagnostic) / NOS (cohort/cross-sectional) / CASP / JBI Qualitative (qualitative) / COSMIN (measurement instruments) / SCED Scale / RoBiNT Scale(single-case design RoB tool;SCRIBE is a single-participant reporting guideline, not a RoB tool)。

15.4 SSOT citation list

This Skill uses the 20 SSOTs of review-methodology-foundations by reference rather than by pasting. For the complete SSOT × phase mapping see §8 + review-methodology-foundations/references/SSOT-INDEX.md. Quick list: SSOT-REVIEW-TYPES (L1) / SSOT-LIFECYCLE-12-PHASES (all) / SSOT-METHODOLOGY-TERMINOLOGY (weak) / SSOT-AI-FAILURE-DEFENSES (all, required reading) / SSOT-EXPERT-METHODOLOGY-INSIGHTS (arbitration) / SSOT-PICO-FRAMEWORKS (L2) / SSOT-SEARCH-STRATEGY (L4) / SSOT-PRISMA-S-TEMPLATE (L5) / SSOT-PRISMA-FLOW-DIAGRAM (L6/L11) / SSOT-EXTRACTION-FIELDS (L7) / SSOT-EFFECT-SIZE-DECISION (when switching to MA at L9) / SSOT-ROB-TOOL-MAPPING (L8) / SSOT-SYNTHESIS-METHODS (L9) / SSOT-GRADE-FRAMEWORK (L10) / SSOT-REPORTING-STANDARDS (L11) / SSOT-AMSTAR-2-CHECKLIST (L12) / SSOT-PSYCHOLOGY-SUPPLEMENTS (psychology scenarios) / SSOT-CHINESE-CONTEXT (Chinese-language scenarios) / SSOT-MODE-PROFILES (L3) / SSOT-WORDCOUNT-RULES (L11).


16 Change log

  • 2026-06-10(v1.1.1): Fix rounds F10–F17 (basis = 2026-06-10 measured audit). Newly created review-methodology-foundations/references/SSOT-INDEX.md(fixes the broken-link references in §8 / §15.4 of this file); parameterized the hardcoded model-name arguments in subagent-templates / L6 / L7 (normative statements changed to "the strongest currently available model", and the disclosure template wording changed to placeholders to be filled in truthfully per session).
  • 2026-06-20 (S4 WP-C orchestration patch): §2.4 / §4.2.8 / §13.1 joint-path L10-L12 wording disambiguation — per Lock 4 §4.4 ruling, joint paths ending in meta-analysis have meta-analysis complete L9-L12(GRADE/reporting/AMSTAR2 produced by MA), while pure SR paths still have SR complete L10-L12 (the original "return to systematic-review to complete L10-L12" conflicted with the MA §12.1 deliverables list in MA-switching scenarios and has been disambiguated). The SR L7→MA L9.1 field mapping table (L7-extraction-dual.md §11) was synchronously rewritten by S4 to align with the MA L9.1 §8.2 input contract (closes BK-2).
  • 2026-06-23 (S6B WP-E② supervision mode): closes BK-9. §9 title changed from "Strict vs Learning" to "Strict / Learning / Supervision" + three-mode quick overview; added §9.4 supervision mode (= strict methodological completeness + human decision review at the 7 irreversible points SP1-SP7 as structurally enforced gates; when to choose = instructor's demand for "AI first draft + human check reaching publication grade"; on the SR side the sign-off points fall mainly at L4 search freeze / L6 screening / L7 extraction, and L9/L11 are taken over by the MA path when switching to an MA endpoint; real gates = preflight missing signatures exit≠0 + κ computed by script + hard checks before soft evaluation; balance = no overfitting on either side + cap on the review surface; disclosure = RAISE human supervision + transparency + ultimate responsibility); points to the operational contract review-methodology-foundations/references/SUPERVISED-MODE-GATES.md + SSOT-MODE-PROFILES v2.0.0. User-facing text follows the spirit of meta-analysis/scripts/CARD_NARRATIVE_VOICE.md to remove jargon. Only §9 was changed (OWNER_MAP §2.2 paragraph owner), no other sections were touched.
  • 2026-06-23 (S7 WP-F Lane A search fallback consistency): closes doc 12c P2-3/P2-4. §11.1 search fallback table removes SS MCP phantom signals that do not exist in the runtime environment (batch_search error / phantom limit / fields type / "SS MCP does not return CNKI"), main path changed to paper-search-pro(HTTP, OpenAlex/SS/CrossRef/PubMed keys already configured) / manual Chinese export from CNKI and Wanfang → scripts/ingest_search_exports.py ingestion; retained the genuine fallbacks (Chinese 0 hits → manual export + ingestion / switch to literature-set-review, overall unavailability → blocking warning, do not fabricate from memory, metadata supplementary extraction) and the key constraint "do not arbitrarily downgrade English search to replace Chinese search"; §3.4/§10.2 search fallback cross-refs synchronously de-SS-MCP-ified. Newly added firecrawl-abstract SS method downgrade note: runtime environment has no SS MCP → paper_details(SS method) unavailable, defaults to the two methods Firecrawl Scrape/Extract. Only §3.4/§10.2/§11.1 were changed (OWNER_MAP §2.2 S7 paragraph owner), no other sections were touched.
  • 2026-07-04 (G5 mandatory-read distribution and slimming): to cure the real-world disease root cause "references 39.3% never read / the three-piece set skipped 2/3 / read rate = functions named in the dispatch prompt". ① **§3.1 mandatory-read-on-startup changed to point to LAUNCH-DIGEST.md(≤10k)(distillation three-piece set ≈72k payload + §1.2-18 frontier clause six tables, three-piece set/deep full text degraded to on-demand drill-down); ②§6 changed from "when to load" to "dispatch copy block"(key payloads via newly created PROMPT-COPY-BLOCKS.md §C whole paragraph into SubAgent prompt = fill-in-the-blank constructive guarantee, highest-risk payload inlined in one sentence in this table); ③§8 path authority changed to point to SSOT-INDEX.md whole-repository resolver** (stage→absolute path, curing the two wrong-paths of L10/L11/L12 living in SR being referenced by MA); ④ good-deviation ratification (Axiom 8): ratify ① planning document = context container (§3.3) + ratify ④ verification-style automatic pre-screening standard process (§4.2.4 pointer + L6-screening-dual-blind.md §12 new section: seed 100% preservation + dual false-negative spot-check + Methods disclosure, does not replace double-blind/two-person full text); ⑤ required_readings keep fields cut reports + ban self-attesting "already read" checkboxes(template deletion + preflight --gate readings one-line WARN). Only §3.1/§6/§8/§3.3/§4.2.4/§16 were changed, §9.4+§4.2.5 (G1)/§11.1 (S7)/other sections were not touched.
  • 2026-07-06 (G9 integration wiring): ① newly added §13.4 satellite dependency matrix(five-row four-tuple hard contract: paper-search-pro[L4/L5 main path, degraded pointer §11.1 not restated] / document-to-markdown+paper-downloader-portable[G6 full-text request package §4.2.6 + download_policy dependency tier] / ma-plots[lives in meta-analysis/scripts/plots referenced by this Skill: pure SR path L11's PRISMA/RoB figure fixed commands + QC exit≠0 + labels mechanical same-source discipline] / manuscript-typeset[pure SR L12 default submission four-piece] / academic-ref-check[citation final review fail-closed]; each row's degradation explicitly bans silence); ②§13.1 "systematic-review + ssci-plots" row fact sync (forest plot/funnel plot belong to MA path; pure SR PRISMA/RoB figures = ma-plots fixed renderer; ssci-plots = style foundation + general statistical figure outlet); ③§12 tail newly added pure SR path L12 delivery packaging pointer blockquote (manifest five-section template lives in MA referenced by SR + submission four-piece + separated and cross-linked with 00_审核入口.md ; see meta-analysis SKILL §12.4). Purely additive/wording fact sync: avoiding §9.4+§4.2.5(G1)/§11.1(S7)/§4.2.4+§4.2.6 G6 blockquote/G5 owned §3.1/§6/§8/§3.3/L6 §12; S4's §12 naming convention line and §13.1 joint workflow row untouched by a single character. ④references consumption side sync:subagent-templates.md L8 RoB output row (publication-grade figure = rob_traffic.py) + 11.3 prisma-flow-agent step 6 changed to prisma_flow.py fixed command + failure fallback changed to "explicitly report what is missing, draft must not be passed off as delivery figure" / L9-synthesis-swim.md : 338 effect direction plot pointer-ized (ssci-plots general figure outlet correct role). ⑤ after critic cross-review added fixes: §12 packaging pointer blockquote reordering(typeset①→manifest②, curing dependency inversion);L11-prisma-reporting.md : 152 and L8-rob-assessment.md : 189/:207/:327 escaped ssci-plots live instructions changed to fixed-renderer contract pointer.