Public handling: selected full-text prompt with secret filtering.
Prompt Text
你是 Codex/GPT-5.5 targeted audit worker B。只读审计 DoneBench 的这些任务：crm_workflow_030, crm_workflow_037, crm_workflow_038, crm_workflow_039, crm_workflow_040, email_022, email_031, email_032, email_037, email_038, email_039。对每个任务读取 data/tasks/<domain>/<task_id>.json，并按 annotation/annotation_guide.md 的五个检查输出一条 compact JSON 对象，字段必须包括 task_id, domain, difficulty, task_pattern, audit_source="model", prompt_version="codex-gpt55-targeted-v1", model="codex_gpt55_session", provider="codex_session", provider_model="gpt-5.5", risk_labels(list), check_opinions(dict of criteria_complete,donespec_matches_criteria,near_misses_are_valid,reference_trace_is_plausible,not_too_templated each with verdict pass|warn|fail confidence number rationale string), overall_risk low|medium|high, needs_adjudication boolean, adjudication_reasons(list), suggestions(list), model_metadata with parse_status="model", audit_mode="codex_session_manual_targeted", no_api_usage=true. Important: this is AI advisory audit, not human annotation. Be rigorous but not performative: use needs_adjudication=true only for high risk/fail or unresolved ambiguity requiring human decision; warn-level issues can still be needs_adjudication=false if suggestions are clear. Return JSONL only, no markdown fences.
Reuse Notes
EXTRACTED: This is a selected Codex prompt or automation prompt from the local Codex corpus.
INFERRED: Future agents can reuse its structure, constraints, and acceptance criteria when creating similar Codex workflows.