PROMPT LIBRARY

Meta & Prompt Engineering: system prompts for agents.

Role prompts and system prompts maintained in awesome-prompts, grouped by field. Copy one into Claude Code, Codex or any chat model.

71 Meta & Prompt Engineering prompts

5W3H Intent Architect EN

Structured intent expansion for any request: Who/What/When/Where/Why/How/How much/How long; reduces cross-model variance and dual-inflation bias; based on "Does Structured Intent Representation Generalize?" (arXiv 2603.25379, 2026)

A2UI Agent-to-User Interface Architect EN

Prompt for designing A2UI-compliant declarative agent-generated interfaces: component catalog allowlists, surface updates, data-model bindings, action intents, sandboxed rendering, no executable code; based on Google's A2UI open protocol (github.com/google/A2UI, 2026, 15.4k+ stars, Apache-2.0)

Abstract Chain-of-Thought Architect EN

Design latent reasoning systems with discrete abstract tokens: vocabulary design, bottleneck warm-up, self-distillation under constrained decoding, RL length penalty, early-exit probes, trajectory audit; up to 11.6× fewer reasoning tokens vs. verbal CoT; based on "Thinking Without Words" (arXiv 2604.22709, April 2026; IBM Research AI)

ADK SkillToolset Designer EN

Prompt for ADK-style progressive-disclosure skills: L1 metadata, on-demand skill payloads, load/unload triggers, versioning, skill-factory tradeoffs (2026)

AG-UI Frontend Architect EN

Prompt for designing AG-UI-compliant agent-to-user frontend integrations: event sourcing, lifecycle/tool/state events, SSE/WebSocket transport, human-in-the-loop interrupts, generative UI payloads; based on the AG-UI open protocol (ag-ui-protocol/ag-ui, 2026, 14k+ stars)

Agent Context Efficiency Engineer EN

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 × 15 adapters); based on mksglu/context-mode (Feb 2026, 15.4k+ stars, Hacker News #1, used by Microsoft/Google/Meta/Amazon/NVIDIA)

Agent Eval Designer EN

Evaluation prompt for real-world agents: task suites, noise audits, reproducibility, intervention/safety metrics, failure taxonomy; derived from Anthropic's 2026 eval guidance

Agent Governance Orchestrator EN

Prompt for defining ownership, delegation, authority, approvals, and audit trails across multiple agents: governance-first orchestration design (2026)

Agent Memory Architect EN

Agent memory systems architect: STM/LTM design, extraction/storage/retrieval modules, hierarchical graph memory, context compression, reasoning-aware recall; based on 2026 memory-architecture research (2026)

Agent Reliability Engineer EN

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 injection, harness-hardening checklist (environment-coupled loops, replan triggers, snapshots, typed error contracts, confirmation gates, budgets), pass@1-overestimates-by-20-40% guardrail, unsafe-success detection; based on "Towards a Science of AI Agent Reliability" (arXiv 2602.16666, 2026) and "ReliabilityBench: Evaluating LLM Agent Reliability Under Production-Like Stress" (arXiv 2601.06112, 2026)

Agent Skill Optimizer Architect EN

Text-space skill trainer that treats natural-language skill documents as neural-network parameters: rollout (forward pass), reflect (backward pass), aggregate, select (gradient clipping), update, and gate (validation) loops; learning-rate schedules, slow-update epoch boundaries against catastrophic forgetting, meta-skill cross-epoch memory, and convergent diagnostics on frozen LLMs; produces deployable best_skill.md artifacts; based on microsoft/SkillOpt (May 2026, arXiv 2605.23904)

Agent Tool Engineer EN

Prompt for designing, evaluating, and iteratively improving agent tools: tool selection/omission (constraint collapse), namespacing, context-rich returns, token-efficient responses, description prompt-engineering, agent-driven optimization loops; based on Anthropic's 2026 "Writing effective tools for agents" guidance

Agent Trajectory Triage Specialist EN

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 over random sampling without ground-truth labels; based on "Signals: Trajectory Sampling and Triage for Agentic Interactions" (arXiv 2604.00356, April 2026, 6.2k HF likes)

Agent-Native Memory System Architect EN

Data-management-first memory system architect: designs representation/storage, extraction, retrieval/routing, and maintenance as measurable modules; workload-aware benchmarking, localized-vs-global maintenance trade-offs, update-correctness discipline; based on "Are We Ready For An Agent-Native Memory System?" (arXiv 2606.24775, June 2026; OpenDataBox/MemoryData benchmark suite)

AgentAtlas Trajectory Auditor EN

Beyond-outcome agent evaluation: separates outcome success, control-decision quality, and trajectory quality using a six-state taxonomy (Act / Ask / Refuse / Stop / Confirm / Recover); identifies primary error source and downstream impact; tests for label-menu dependence; based on "AgentAtlas: Beyond Outcome Leaderboards for LLM Agents" (arXiv 2605.20530, May 2026)

Agentic Context Engineering Architect EN

Evolving-context playbook architect for self-improving agents: Generator/Reflector/Curator roles, itemized structured bullets with outcome counters, incremental delta updates (no full rewrites), grow-and-refine with semantic de-duplication, anti-collapse and anti-brevity guardrails; based on "Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models" (arXiv 2510.04618, v3 March 2026; +10.6% agent benchmarks, +8.6% finance)

agentmemory Persistent Memory Architect EN

Persistent-memory architect for coding agents: confidence-scored memory taxonomy, hybrid retrieval, temporal knowledge graph, session compression, MCP tool surface, and platform integration across Claude Code / Codex / Cursor / Gemini CLI / Hermes / OpenClaw / pi / OpenCode; based on rohitg00/agentmemory (Feb 2026, 27k+ stars)

Agents Best Practices EN

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, 654 stars)

Caveman Mode EN

Ultra-compressed agent communication: drops articles, filler, and hedging while preserving full technical accuracy; ~75% output-token reduction; supports lite/full/ultra/wenyan intensity levels; based on JuliusBrussee/caveman (Apr 2026)

Chain of Draft EN

Minimal reasoning scratchpad: 5 words per step, 92% fewer tokens vs CoT (arXiv 2502.18600)

Clarification Timing Strategist EN

Timing-aware clarification policy for long-horizon agents: empirically-derived windows for goal/input/constraint/context clarification; goal clarifications lose nearly all value after 10% execution (pass@3 drops from 0.78 to baseline), input clarifications retain value through ~50%, and deferring any clarification past mid-trajectory degrades performance below never asking; cross-model Kendall tau 0.78, 0.87 confirms task-intrinsic timing curves; based on "Ask Early, Ask Late, Ask Right" (arXiv 2605.07937, May 2026)

Cognitive Externalization Architect EN

Unified four-layer architect that decides which cognition stays in weights, which lives in the prompt, and which is externalized into memory / skills / protocols / harness: precondition check, per-layer audit (what belongs where, what does not), interface contracts between layers (no cross-layer bypass), invariants (separation of concerns / least privilege / inspectability / reversibility / versioning), test plan, and a strict output contract that forces every cognitive function to declare its location; refuses "mega-prompt" designs and "externalize everything" router-agents alike; based on "Externalization in LLM Agents: Memory, Skills, Protocols, Harness" (arXiv 2604.08224, April 2026, Shanghai Jiao Tong / UCL)

Constraint Typology Architect EN

Constraint workflow designer for LLM-based planning: hard/soft constraint typology with formal model checking vs LLM-as-judge verification, intent alignment, conflict resolution, constraint versioning; based on U-Define (arXiv 2605.02765, May 2026)

Context Engineering Maturity Architect EN

Context-engineering maturity architect: designs the full informational environment for agents across the four-level pyramid (Prompt → Context → Intent → Specification Engineering) and audits it against five quality criteria (relevance, sufficiency, isolation, economy, provenance); based on "Context Engineering: From Prompts to Corporate Multi-Agent Architecture" (arXiv 2603.09619, 2026)

ContextNest Verifiable Context Governance Architect EN

Governed knowledge-vault architect beneath RAG: typed Markdown artifacts, deterministic set-algebraic selectors, contextnest:// URI citations, SHA-256 hash-chained versions, graph checkpoints, MCP source nodes, and audit traces so every agent output is reconstructible; based on ContextNest (arXiv 2607.02116, July 2026; IBM Research / Emory)

Diffusion LM Prompt Engineer EN

Prompt engineering for non-autoregressive diffusion language models (LLaDA, Dream, MMaDA): bidirectional prefix/suffix conditioning, fill-in-the-middle design, mask scheduling, step-level intervention, test-time scaling via S³ parallel trajectories + verifier selection, CFG and temperature analog tuning; based on 2025, 2026 diffusion-LM research (2026)

Disclosure Policy Designer EN

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 bias; based on "When to Think, When to Speak" (arXiv 2605.03314, ICML 2026)

Divergent Ideation Architect EN

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 UditAkhourii/adhd (May 2026, 502 stars, preprint + The New Stack)

Elastic Context Orchestrator EN

Elastic context orchestration architect for long-horizon agents: Context-ReAct loop with five atomic operations (Skip, Compress, Rollback, Snippet, Delete), adaptive relevance scoring, hot/warm/cold context layers, expressive-completeness verification for compression, rollback checkpointing, and horizon-specific failure mitigation; based on LongSeeker (arXiv:2605.05191, May 2026)

Eval & Benchmark Architect EN

Benchmark design, evaluation metrics, rubric development, failure mode analysis, continuous monitoring: regression testing, cost-effective evaluation (2026)

Eval Awareness Auditor EN

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-cue) before attributing residual to eval awareness, both-direction audit (capability and safety, over- and understatement), probe rotation as a leak control, layered mitigations (report-the-gap → parallel CI → paraphrase rewrites → post-training only on held-out probes), production drift monitoring; based on Anthropic's "Eval Awareness in Claude Opus 4.6's BrowseComp Performance" (anthropic.com/engineering/eval-awareness-browsecomp, March 2026)

Headroom Context Compression Architect EN

Context compression layer architect for AI agents: 60, 95% token reduction via SmartCrusher / CodeCompressor / Kompress-base / CacheAligner; reversible CCR cache, cross-agent memory, library/proxy/wrap/MCP integration modes; based on headroomlabs-ai/headroom (Apache-2.0, ~50k stars, 2026)

HyperAgents Designer EN

Self-referential meta-agent designer: task and meta layer unified in a single editable program, evidence-grounded self-edits, recursion bounds, regression-gated commits, immutable kill switch and eval harness; based on Meta FAIR's "Hyperagents: Self-Referential Meta-Agents" (arXiv 2603.19461, Mar 2026, 2.1k HF likes; open source `facebookresearch/HyperAgents`)

Instruction Bleed Auditor EN

Cross-module interference audit for prompt-composed agentic systems: detects Compositional Behavioral Leakage (CBL) where one prompt module silently shifts the behavior of another sharing the same context window; three-channel perturbation protocol (volume / content / form), effect-size reporting, leakage classification (positional / semantic / format / compound), critical-boundary escalation, and isolation-first mitigation plan; based on "Instruction Bleed: Cross-Module Interference in Prompt-Composed Agentic Systems" (arXiv 2606.26356, June 2026; ICML 2026 FAGEN workshop)

Interruptible Agent Planner EN

Prompt for multi-step agents that must absorb mid-task user changes safely: state snapshot, stop/preserve decisions, re-plan, irreversible-risk tracking (2026)

LLM-as-a-Judge Routing Strategist EN

Cost-efficient routing strategist for LLM-as-a-Judge: per-query decisions between reasoning and non-reasoning judges under a hard budget, task-class decomposition (VERIFICATION / PREFERENCE / AMBIGUOUS), leakage-safe routing signals, KL-ball distributionally-robust optimization, budget accounting with end-of-window carve-out, production drift monitoring with rho-widening, "reasoning theater" detection on simple items, mandatory pre-promotion Pareto-dominance check against always-reason and never-reason baselines; refuses to ship policies without held-out shift evaluation or cost numbers; based on "Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge" (arXiv 2605.10805, ICML 2026; reasoning helps on structured-verification tasks like math/code but yields limited or negative gains on simpler evaluations at multiples of the cost)

Local-First Memory Engineer EN

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/LoCoMo/ConvoMem/MemBench); refuses summarization-as-storage and global-scope searches by default; based on MemPalace/mempalace (Apr 2026, 51k+ stars)

Lookahead Planning Specialist EN

Replaces stepwise-greedy CoT with explicit forward planning for long-horizon agents: plan tree (branching × depth), reward-estimation strategy (self-eval / learned verifier / env proxy / retrieval / hybrid), explicit replan triggers, optimal-vs-satisficing decision, K×D compute budgeting, planner/executor separation, irreversibility gates; based on FLARE: Why Reasoning Fails to Plan (arXiv 2601.22311, 2026) and Google DeepMind's Optimality of LLMs on Planning Problems (arXiv 2604.02910, April 2026)

MASPO Joint Prompt Optimizer EN

Joint prompt optimizer for LLM-based multi-agent systems: Local Validity + Lookahead Potential + Global Alignment evaluation, misalignment-case hard-negative mining, evolutionary beam search with Beam Refresh, trace-guided mutation, Gauss-Seidel synchronization; no ground-truth labels needed for intermediate agents; based on MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems (arXiv 2605.06623, ICML 2026)

MCP Apps UI Architect EN

Prompt for designing interactive UI extensions for MCP servers: `ui://` resources, `_meta.ui` tool bindings, sandboxed iframe bridge, JSON-RPC over postMessage, permissions/CSP; based on the MCP Apps open standard (Anthropic/OpenAI, 2026)

MCP Server Architect EN

Prompt for designing secure, interoperable Model Context Protocol servers: flat schemas, error contracts, transport guidance, testing strategy (2026)

Meta Context Engineering Architect EN

Bi-level architect that co-evolves context-engineering skills and context artifacts: meta-level agentic crossover over a skill library, base-level execution that produces files/code/retrieval queries, dynamic context sizing, and feedback-driven skill promotion; based on "Meta Context Engineering via Agentic Skill Evolution" (arXiv 2601.21557, ICML 2026; 16.9% mean improvement, 13.6× faster training)

Meta Prompt EN

Meta-Expert orchestrates specialist sub-agents to solve complex problems

Meta-Cognitive Tool Use Specialist EN

Prompt for deciding *whether* to invoke a tool: self-knowledge probing, cost-benefit gating, confidence calibration, tool-budget tracking, redundant-call detection; addresses the meta-cognitive deficit where naive agents over-tool 98% of the time; based on Alibaba's "Act Wisely" / HDPO research (April 2026)

Modular Prompt Transpilation Architect EN

Design scalable, build-system-native prompt programs: modular skill files, deterministic transpilation, static validation (missing imports / undefined variables / circular dependencies), golden-file drift checks, progressive skill disclosure, and agent-self-maintenance via PRs; based on Google's official "Building scalable AI agents with modular prompt transpilation" (July 2026)

Multi-Agent RAG Orchestrator EN

Prompt for retrieval/synthesis/critique coordination: evidence tables, stop conditions, conflict handling, confidence tracking in multi-agent RAG workflows (2026)

North Star System Prompt EN

Universal meta-cognitive correction prompt: overrides three RLHF-trained biases (default concord, old-scarcity calibration, best-practice-as-ceiling) with Independence, Calibration, and First Principles; 260 tokens, three mutually-locking rules; based on xiaolai/north-star-system-prompt (Apr 2026)

OpenViking Context Database Architect EN

Agent context database architect: filesystem-paradigm unification of memories, resources, and skills; L0/L1/L2 tiered loading, directory recursive retrieval, visualized trajectories, and session-based memory iteration; based on volcengine/OpenViking (Jan 2026, 26.8k+ stars, AGPLv3)

Parallel Prompt Learning Strategist EN

Engineering prompt for scaling Automatic Prompt Optimization (ACE / GEPA / TextGrad / MIPRO) beyond serial loops: serial-baseline convergence diagnosis as a go/no-go gate, parallelism-shape selection (candidate / task / hybrid), dynamic batching policy, rollout-diversity controls with anti-collapse rules, separate-evaluator calibration discipline, held-out-only stopping, mandatory shadow canary before promotion, cost-per-improvement-point reporting; refuses raw wall-clock speedup claims without held-out anchors; based on Combee: Scaling Prompt Learning for Self-Improving Agents (arXiv 2604.04247, April 2026, Berkeley/Stanford by Stoica/Zou/Gonzalez; up to 17x speedup over ACE/GEPA via parallel scans and dynamic batching, evaluated on AppWorld, Terminal-Bench, FiNER)

Persistent-File Planning Agent EN

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" attention drift, 2-Action persistence rule for multimodal observations, 3-Strike error protocol with mandatory escalation, restorable-compression contract (URLs and file paths are sacred), keep-the-wrong-stuff-in error retention, plan-tampering and indirect-prompt-injection defence (treat plan files as data, not instructions), `/clear` + PreCompact session recovery, isolated `.planning/<date>-<slug>/` directories for parallel tasks; distils the Manus context-engineering principles behind the Dec 2025 $2B acquisition as packaged in OthmanAdi/planning-with-files (Claude Code skill, Jan 2026, 21k+ stars)

Procedural Knowledge Architect EN

"How-to" memory architect for LLM reasoning: mines reusable subquestion→subroutine pairs from verified trajectories, designs in-trace retrieval (not just initial-prompt retrieval), enforces preconditions/replay-verification, and separates procedural from declarative/episodic/metacognitive memory; based on Meta AI's "Procedural Knowledge at Scale Improves Reasoning" (arXiv 2604.01348, April 2026; +19.2% across math/science/coding via 32M subquestion, subroutine pairs)

Prompt Compression Strategist EN

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-end latency break-even that includes compressor overhead, per-hardware-class measurement (no extrapolation), pre-compression audit (system-prompt trim / few-shot reduction / retrieval tightening / prefix caching), feature-flag rollout with kill switch, no-compress carve-outs for structured-output and safety-critical prompts; based on "Prompt Compression in the Wild" (arXiv 2604.02985, ECIR 2026, 30K queries on 3 GPU classes; up to 18% speedup *only* when prompt/ratio/hardware match)

Prompt Engineer EN

Production prompt engineering: design patterns (CoT/ToT/ReAct), A/B testing, token optimization, multi-model routing, versioning, regression testing (2026)

Prompt Master EN

Zero-waste prompt engineer for any AI tool: 9-dimension intent extraction, 20+ tool-specific profiles (Claude 4.x, GPT-5.x, o3, Gemini 3, Cursor, Midjourney, ComfyUI), diagnostic checklist, token-efficiency audit; based on nidhinjs/prompt-master (Mar 2026)

Proprioceptive Context Dashboard Architect EN

Self-managed context architect: restructures the transcript into typed, addressable blocks and exposes a runtime dashboard (token usage, recency, access history, context pressure) so the agent can KEEP / ARCHIVE / RECOVER / MERGE / PIN / DROP blocks before acting; full-fidelity recoverable archive, training-free, model-agnostic; based on VISTA (arXiv 2606.30005, revised July 2026; Gemini-3-Flash 22.7% → 50.7% on LOCA-Bench)

Reasoning Drift Auditor EN

Multi-turn agent reasoning-stability auditor: fixed hard-probe baselines, CoT length/depth instrumentation, drift vs intentional-compression discrimination, tiered mitigations (reasoning-budget directives → InftyThink-style checkpoints → fresh-context handoff → model routing), differential diagnosis vs template collapse; based on Reasoning Shift: How Context Silently Shortens LLM Reasoning (arXiv 2604.01161, April 2026)

Reasoning Model Prompting EN

Guide + templates for o1/o3/Claude thinking/Gemini: what to do, what NOT to do, effort control (2026)

Reasoning Theater Diagnostician EN

Per-workload audit of whether chain-of-thought is *substance* (genuinely changes the answer) or *theater* (decorative tokens around an answer that was already fixed before reasoning began): pre-declared probe battery (ablation / length sensitivity / trace perturbation / silence probe / logit-lens), SUBSTANCE / THEATER / MIXED / INCONCLUSIVE verdicts with confidence intervals, escape-hatched router design, weekly canary against verdict drift, differential diagnosis against memorisation and template anchoring, both-directions auditing (forcing CoT on theater workloads AND suppressing CoT on substance workloads are both bugs); refuses bare savings numbers without accuracy CIs and refuses to inherit verdicts across model versions; based on Reasoning Theater: Disentangling Model Beliefs from CoT (arXiv 2603.05488, 2026; probe-guided early-exit reduces token generation by up to 80% on simple tasks at no accuracy cost)

ReContext Recursive Evidence Replay Architect EN

Training-free long-context reasoning harness: uses model-internal attention traces to build a query-conditioned evidence pool, recursively replays it near the question, and generates from the full original context plus the replayed evidence; full-context preservation, no compression or summarization by default; based on ReContext (arXiv 2607.02509, July 2026; github.com/Yanjun-Zhao/ReContext)

REprompt Requirements Engineering Prompt Architect EN

Requirements-engineering-driven prompt architect: elicitation, analysis, specification, validation pipeline with Interviewee/Interviewer/CoTer/Critic agents; turns vague intent into complete, consistent, verifiable system or user prompts; based on REprompt: Prompt Generation for Intelligent Software Development Guided by Requirements Engineering (arXiv 2601.16507, Jan 2026)