Prompt Cards

Build an eval that scores ground-truth attribution (exact passage match), answer faithfulness, and coverage. Provide dataset schema and a nightly regression plan.

Propose a router that routes by budget and SLA: small model for recall, larger for synthesis, tools only when signal warrants. Provide policy examples and guardrails.

Create a memory layer that stores hypotheses, decisions, and unresolved questions. Provide schemas, TTL rules, and a notebook export (Markdown/CSV).

Design a UI: inline citations, hover previews, source quality badges, and ‘expand to passage’. Include keyboard flows and accessibility specs (WCAG). Output wireframes and event schema.

Specify constrained modes with scope limits, consent prompts, guideline links, and mandatory citation density. Provide refusal templates and domain-specific disclaimers.

Create a cross-lingual flow: detect language, translate queries, use multilingual embeddings, and cite in user language while preserving original quotes. Provide locale heuristics and typography rules.

Draft defenses for tool-using agents: content sanitization, domain allowlists, URL reputation, and read-only sandboxes. Provide red-team prompts and pass/fail gates.

Design a time-decay model that boosts fresh, authoritative sources. Include publisher priors, sitemap/news bias, and re-crawl triggers. Output evaluation with time-split test sets.

Create a compression stage with map-reduce summaries, selective citation carry-through, and entropy-based token pruning. Provide metrics (coverage, faithfulness) and an ablation plan.

Specify a KG-augmented RAG: entity/link extraction → KG lookup → vector recall → late fusion ranking. Provide ranking features (entity overlap, path length), and a sample Gremlin/SPARQL query pack.

Design a debate pattern: one agent argues a thesis, another refutes with sources, a judge agent scores evidence and drafts a balanced synthesis. Provide rubric, tie-breaker rules, and output template.

Build a verifier that detects agreement/contradiction between sources. Define stance labels, entailment thresholds, and escalation rules to human review. Output a confusion matrix target and fallback messaging.