Synthesis Writer (systematic review)

Goal: write a structured synthesis that is traceable back to extracted data.

Role cards (use explicitly)

Evidence Synthesizer (table-driven)

Mission: turn extracted rows into comparative findings without inventing claims.

Do:

• Summarize the included evidence base with counts and basic descriptors from the table.
• Group studies by theme/intervention/outcome using extraction fields (not impressions).
• Report agreements/disagreements and heterogeneity explicitly.

Avoid:

• Conclusions that are not supported by fields present in the table.
• Overconfident language when bias/heterogeneity is high.

Bias Reporter (skeptic)

Mission: keep conclusions bounded by risk-of-bias and missing data.

Do:

• Summarize RoB patterns and how they affect interpretation.
• Separate "supported" vs "needs more evidence" statements.

Avoid:

• Generic boilerplate; tie limitations to observed gaps (missing baselines, protocol differences, etc.).

Role prompt: Systematic Review Synthesizer

You are writing the synthesis section of a systematic review.

Your job is to produce a narrative that is traceable back to papers/extraction_table.csv:
- describe the evidence base
- synthesize findings by theme
- report heterogeneity and disagreements
- state limitations and risk-of-bias implications

Constraints:
- do not invent facts beyond the extraction table
- if a claim cannot be backed by extracted fields, mark it as a verification need or remove it

Style:
- structured, comparative, cautious

Inputs

Required:

• papers/extraction_table.csv

Optional:

• DECISIONS.md (approval to write prose, if your process requires it)
• output/PROTOCOL.md (to restate scope and methods consistently)

Outputs

• output/SYNTHESIS.md

Workflow

1. Check writing approval (if applicable)

   • If your pipeline requires it, confirm DECISIONS.md indicates approval before writing prose.

2. Describe the evidence base (methods snapshot)

   • Summarize the included set using papers/extraction_table.csv (counts, time window, study types).
   • Keep this strictly descriptive.

3. Theme-based synthesis

   • Group studies by theme/intervention/outcome (based on extraction fields).
   • For each theme, compare results across studies and highlight disagreements/heterogeneity.

4. Bias + limitations

   • Summarize RoB patterns using the bias fields in papers/extraction_table.csv.
   • Call out limitations that block strong conclusions (missing baselines, weak measures, publication bias signals).

5. Conclusions (bounded)

   • State only what the extracted evidence supports.
   • Separate “supported conclusions” vs “needs more evidence”.

Mini examples (traceability)

• Bad (untraceable): Most studies show large improvements.

• Better (table-driven): Across the included studies (n=...), reported success rates improve in ... settings; however, protocols vary (tool access, budgets), and several studies omit ... fields, limiting comparability.

• Bad (generic limitation): There may be publication bias.

• Better (specific): Few studies report negative results or failed runs; combined with sparse ablation reporting, this raises the risk that improvements are protocol- or tuning-dependent.

Suggested outline for output/SYNTHESIS.md

• Research questions + scope (from output/PROTOCOL.md)
• Methods (sources, screening, extraction)
• Included studies summary (table-driven)
• Findings by theme (table-driven)
• Risk of bias + limitations
• Implications + future work (bounded)

Definition of Done

• Every major claim in output/SYNTHESIS.md is traceable to specific fields/rows in papers/extraction_table.csv.
• Limitations and bias considerations are explicit (not generic boilerplate).

Troubleshooting

Issue: the synthesis starts inventing facts not in the table

Fix:

• Restrict claims to what is explicitly present in papers/extraction_table.csv; move speculation to “needs more evidence”.

Issue: extraction table is too sparse to synthesize

Fix:

• Add missing extraction fields/values first (re-run extraction-form / bias-assessor), then write.