Audits and closes the gap between benchmark scores and production behavior — matched eval-shape vs production-shape probe pairs, per-workload delta with CIs, mandatory differential diagnosis (distribution shift / template fragility / length effects / tool availability / safety...
Post-deployment trajectory sampling and triage prompt — three-dimensional signal taxonomy (interaction / execution / environment), cheap-rules-first extractors, diversified ranking, reviewer-feedback loop, explicit privacy-redaction step; designed to lift informative traces ov...
Reliability-engineering prompt that separates reliability from capability — four-dimension scorecard (consistency, robustness, predictability, safety/fault-tolerance), 3D reliability surface R(k, ε, λ) with explicit operating envelopes, chaos-engineering plan with fault inject...
Evaluation prompt for real-world agents — task suites, noise audits, reproducibility, intervention/safety metrics, failure taxonomy; derived from Anthropic's 2026 eval guidance
Benchmark design, evaluation metrics, rubric development, failure mode analysis, continuous monitoring — regression testing, cost-effective evaluation (2026)
Side-by-Side (SxS) interleaved reasoning strategist — designs when an agent should reveal reasoning vs. keep it private in streaming interfaces; support-threshold gating, update-granularity ladders, silence-tax management, anti-filler rules, correction protocols for commitment...
Guide + templates for o1/o3/Claude thinking/Gemini — what to do, what NOT to do, effort control (2026)
Production decision framework for *structural* prompt compression (LLMLingua / LongLLMLingua / LLMLingua-2 / Selective Context / RECOMP) — workload profiling, compressor-family selection by prompt structure, per-workload ratio sweeps with slice-level accuracy budgets, end-to-e...
Minimal reasoning scratchpad — 5 words per step, 92% fewer tokens vs CoT (arXiv 2502.18600)
Parallel divergent ideation for open-ended problems — spawns N isolated reasoning branches under cognitive frames (hardware, biology, speedrunner, $0 budget), separates generator from critic, scores novelty/viability/fit, clusters by angle, deepens survivors; based on UditAkho...
Provider-neutral agent harness architect — MVP blueprint, loop design, tool/permission contracts, context/memory/compaction, planning/goals, skills/MCP connectors, prompt caching, observability/evals, safety guardrails; based on DenisSergeevitch/agents-best-practices (May 2026...
Context-window optimization architect for AI coding agents — Think-in-Code discipline (script execution vs bulk file reads), sandboxed tool-output routing, session continuity via indexed event stores, context telemetry with savings targets, and cross-platform discipline (3 OS ...
Distill any person's thinking into a reusable agent skill — six-layer extraction (mental models, decision heuristics, expression DNA, values, anti-patterns, honest limits), triple-verification gate, parallel research swarm, and calibrated uncertainty; based on alchaincyf/nuwa-...
Filesystem-as-working-memory pattern for long-horizon agents — three durable Markdown files (`task_plan.md` / `findings.md` / `progress.md`) as the single source of truth, KV-cache–stable prefixes (no timestamps, append-only), plan recitation against "lost in the middle" atten...
Verbatim, locally-stored, benchmark-driven agent memory — palace-structured index (Wings/Rooms/Drawers/Diaries), no-LLM raw recall path, pluggable backends, temporal entity-relationship graph with validity windows, MCP/auto-save host hooks, held-out R@k discipline (LongMemEval...
Auto-generates high-quality prompts from a brief description
Meta-Expert orchestrates specialist sub-agents to solve complex problems
