rules-reference.md

Rules Reference

The Review rules panel (toolbar, or ⌘K → "Review rules") lets you select extra rules that are injected into the AI's system prompt for every review, on top of the active skill. Rules are grouped by category, and you can add your own free-text rules too. Your selection is saved locally and included in Backup & Restore.

This is the Research Scribe analog of a project conventions file: a curated catalog you opt into, so the reviewer follows your house conventions instead of guessing.

Each rule below shows its chip label and the instruction injected into the prompt. Every rule ships in both Arabic and English; the English text is shown here.

Integrity

Rule	Injected instruction
No fabrication	Never invent data, references, author names, years, DOIs, or findings; if something is missing, flag it instead of supplying it.
Causality guard	Never infer causation from correlational or cross-sectional designs; use "associated with", "predicted", or "explained variance in" unless the design supports causal claims.
Preserve hypotheses	Never flip a hypothesis outcome (supported ↔ rejected) or alter any reported result to fit a smoother narrative.
Flag, do not fix	If numbers, sample sizes, or results contradict each other, keep the original values and report the discrepancy in the notes rather than silently correcting it.
No detection evasion	Do not introduce errors, awkward phrasing, or altered data to evade AI-detection or plagiarism tools; aim for clean, honest academic writing.

Statistics

Rule	Injected instruction
Lock all numbers	Keep every number, percentage, mean, SD, correlation, p-value, R², Beta, and sample size exactly as written.
Correct stat symbols	Use correct statistical notation consistently: italic M, SD, r, B, β, R², p, and report exact values where the source gives them.
Separate result & reading	In the discussion, keep the statistical result separate from the researcher's interpretation; never present an interpretation as a proven fact.
Keep effect sizes	Preserve effect sizes and confidence intervals when present, and do not drop them in favor of p-values alone.

Citations

Rule	Injected instruction
Preserve citations	Keep every in-text citation and its attribution intact; do not merge, drop, or re-attribute sources.
Flag weak references	If a reference looks incomplete or doubtful, leave a note for review instead of fabricating the missing parts.
Synthesize, don't list	When presenting prior studies, connect them analytically by agreement, disagreement, and research gap rather than listing them one after another.

Structure

Rule	Injected instruction
Logical cohesion	Ensure each paragraph reads as part of a sequential argument with clear transitions, not a set of isolated sentences.
Journal concision	Write for a concise peer-reviewed article: foreground the research gap, framework, methods, key findings, and discussion; trim dissertation-style padding.
Declarative results	Report results in precise, direct, declarative language and move extended interpretation to the discussion.
No added claims	Do not add conclusions, recommendations, or results that are not present in the source text.
Preserve proper nouns	Keep all institution names, regions, variable names, and dimensions exactly as they appear in the study.

Tone

Rule	Injected instruction
Objective tone	Maintain a neutral, objective tone; avoid praise words like "stunning", "remarkable", or "massive significance".
No clichés	Avoid formulaic transitions and filler such as "It is worth noting", "In this context", "Needless to say", and "Based on the above".
Vary sentences	Vary sentence structure and length naturally, mixing short and longer sentences so the prose does not read mechanically.
No em-dashes	Do not use em-dashes (—); restructure with commas, parentheses, or separate sentences instead.

Reporting

Rule	Injected instruction
Conservative phrasing	Soften over-strong claims to match the evidence: prefer "the results suggest" or "the variable contributed to explaining" over "proves" or "causes".
Define acronyms	Define each acronym at first use, then use it consistently; do not introduce undefined abbreviations.
State limitations	Where the source acknowledges limitations, preserve them plainly and do not overstate generalizability beyond the sample.

---

Adding a rule to the codebase: append an entry to RULES in src/lib/rules.ts with a label, a category, and textEn + textAr. It appears in the panel automatically and is resolved into the prompt by resolveRules().