name	audit
description	Full FactHarbor quality audit — all prompts and key pipeline stages. Use when investigating analysis quality regression, before a release phase, or after major pipeline changes.
allowed-tools	Read Glob Grep Bash

ultrathink

Perform a comprehensive quality audit of the FactHarbor CB pipeline. Focus area: $ARGUMENTS (leave blank for full audit)

Step 1 — Read all prompt files: Read every file under apps/web/prompts/. Note each file's role in the pipeline.

Step 2 — Read key pipeline stages: Read these files in apps/web/src/lib/analyzer/: claim-extraction-stage.ts, research-extraction-stage.ts, boundary-clustering-stage.ts, verdict-generation-stage.ts, aggregation-stage.ts, evidence-filter.ts

Step 3 — Identify issues across 7 categories:

A. LLM Intelligence — deterministic text-analysis logic that should be LLM-powered (AGENTS.md: MANDATORY)
B. Hardcoding — domain-specific keywords, named entities, or case-specific patterns
C. Teaching-to-the-test — prompt examples using terms from known benchmark inputs (Bolsonaro, Hydrogen, Venezuela)
D. Evidence quality — probativeValue assignment, claimDirection accuracy, EvidenceScope correctness
E. Verdict calculation — direction, confidence propagation, counter-evidence weighting
F. UCM sync — hardcoded values that belong in UCM admin config
G. Contract mismatches — prompt output schema ≠ TypeScript types consuming it

Step 4 — Classify each issue:

PHASE-BLOCKER — degrades analysis quality or violates a MANDATORY AGENTS.md rule
HIGH — significant quality impact
MEDIUM — noticeable but not critical
LOW — polish / housekeeping

name	audit
description	Full FactHarbor quality audit — all prompts and key pipeline stages. Use when investigating analysis quality regression, before a release phase, or after major pipeline changes.
allowed-tools	Read Glob Grep Bash