Opus Severity-Filter Trap ⚠ verify

Definition

Telling Opus to "only report high-severity issues" causes it to investigate fully but withhold sub-threshold findings, dropping recall — the fix is to separate detection from filtering.

Key points

• The trap: a prompt like "only report high-severity findings" makes Opus 4.8 do the full analysis but suppress anything below the bar, so genuinely relevant lower-severity items never surface. ⚠ verify (2026-07-08)
• The fix: "Report every issue you find; a separate step filters by severity." Detect exhaustively first, filter afterward (in a later pass or in application code).
• Mirrors the general principle: don't fuse a discretionary filter into the generation step — separate concerns.
• Pairs naturally with Multi-Pass And Multi-Instance Review (pass 1 = detect all; pass 2 = classify/filter).

Why it matters for the exam

• A named, model-specific D4 gotcha — the correct answer always decouples "find everything" from "filter by severity."
• Tests whether you recognize recall loss from severity-gated prompts vs an unfiltered detect-then-filter design.

Common gotchas

• Choosing "add stronger emphasis to only report high-severity" — wrong; that worsens recall.
• Filtering inside the same generation call rather than in a separate downstream step.

See also

• Multi-Pass And Multi-Instance Review
• Prompt Clarity And Motivation
• Validation-Retry Loops

Sources

• netlight-study-guide