← Hub

Super Individual

BigCat's Power Stack — Engineering depth for AI super-individuals

> Prompt · Context · Harness · Agent · Eval · Tool Use
SERIESSkill Shelf →Top-rated AI skills & best practices · 10 categories · updated daily
Super Individual Daily— Day 1–59 · the AI engineering set · complete
D01Prompt EngineeringFour-layer arch · XML vs MD · CoT limits · Prefix caching D02Context EngineeringLost-in-the-Middle · Layout · Chunk design · Compaction D03Harness EngineeringAgent runtime · Tool registry · Permission gate · Loop control D04Tool Use & Function CallingSchema as prompt · Granularity · Selection collapse · Parallel calls D05Agent Design PatternsWorkflow vs agent · ReAct/Plan/Reflexion · Multi-agent anti-patterns D06Eval EngineeringGolden set · LLM-as-judge debias · Prompt regression · Anthropic Evals D07Memory & State ManagementFour state layers · Short-term tiers · 3 long-term archs · Self-maintained profile D08Multimodal EngineeringResolution = tokens · PDF tiers · Vision prompt · Multimodal RAG D09Prompting PatternsCalibration drift · Negation fails · Examples > Rules · Sycophancy → Steelman D10RAG in ProductionChunking gens · Hybrid + RRF · Reranker · HyDE / Multi-query D11Hallucination EngineeringRLHF bias · Token risk map · Grounding stack · Calibration eval D12Fine-tuning vs PromptingROI decision tree · LoRA/QLoRA · Data quality · Decoding params D13Multi-agent SystemsAnti-patterns · Orchestrator-Worker · Debate · Coord tax / A2A D14Inference OptimizationMemory-Bound · Continuous Batching · KV Cache · Speculative Decoding D15Latency EngineeringTTFT vs TPOT · Streaming · Prompt Caching · Tail Latency / p99 D16Cost EngineeringToken Economics · Prompt Caching · Model Routing · Batch & Monitoring D17Claude Code Power UserLayered Memory · Subagents · Skills · Headless + MCP D18MCPThree Primitives · stdio vs HTTP · Build a Server · Context Tax D19Coding AgentsEdit Format · Plan Mode · Autonomy Spectrum · Closing the Loop D20Refactoring with AICharacterization Test · Strangler Fig · Non-local Reasoning · Codemod vs LLM D21AI for ResearchDeep Research Loop · Research Harness · Source Curation · Citation Tracing D22Writing EngineeringVoice Spec · Anti-Slop · Hierarchical Prompting · AI as Editor D23Personal AI InfraLLM Gateway · Cross-Model Routing · Two Cache Layers · Observability & Keys D24Prompt Injection: Attack & DefenseLethal Trifecta · Untrusted-Content Isolation · Capability Gate · Injection Eval D25Agentic IDEAsync Agents · Index vs Agentic Search · Verification Ceiling · Fan-out Parallelism D26Computer Use & Browser AgentsPixel vs DOM · Screenshot Loop · Reliability Engineering · Injection Isolation D27Synthetic Data & Self-ImprovementDistillation Filter · Constitutional/RLAIF · Rejection Sampling · Model Collapse D28Local & Edge LLMQuant Loss · Memory Math · Runtime Diffs · Hybrid Routing D29Eval Beyond BenchmarkContamination · Construct Gap · Slice-Based · Agentic Eval D30Open Source ModelsSelection & License · Quantization & VRAM · Prompt/Tool Porting · MoE & Reasoning D31Personal AI SafetyKill Switch · Blast Radius · Deception Signals · Reversibility D32The Next Five Years of AI CodingSWE-bench Limits · Time Horizon · Skill Repricing · Agent-Ready D33AI Governance for Large Legacy CodebasesStrangler Rollout · Guardrails-as-Code · The Ratchet · Review Without the Bottleneck D34Human-in-the-Loop EngineeringConfidence Routing · Async Approval · Interruptible Resume · Approval Audit D35Prompt-as-Code / Version GovernanceRegistry · Eval Gate · Canary & Rollback · Cross-Model Drift D36Governing an Org's Skills LibraryIntake Contract · Dedup · Security Gate · Usage Retirement D37AI ObservabilityTrace/Span · Cost Attribution · Drift Detection · Online Eval D38Structured OutputFour Tiers · Constrained Decoding · Schema as Prompt · Tolerance Stack D39Agent Error Recovery & ResilienceTwo-Layer Retry · Idempotency & Saga · Checkpoint Resume · Breaker & Partial Failure D40Data Pipeline for AIEmbedding ETL · Incremental Index · Index Versioning · Feature Store D41Streaming & Interruption EngineeringTTFT Perception · Partial Parsing · Streaming Tool Use · Cancel & Rollback D42Testing & CI/CD for AIRegression Assert · Eval Gate · Shadow Test · Canary Rollback D43Retrieval Quality EngineeringRecall First · RRF Fusion · Reranker · Layered Attribution D44AI-Assisted Data AnalysisSchema Linking · Code Interpreter · Statistical Traps · Reconciliation D45AI Engineering Anti-PatternsOver-Engineering · Missing Evals · Prompt Brittleness · Vendor Lock-in D46Voice AI EngineeringCascaded vs End-to-End · Turn Detection · Barge-in · Latency Budget · Watermarking D47Image Generation EngineeringControl Stack · Counter-intuitive Prompting · Generation Agent · Provenance D48Knowledge Bases & GraphRAGLayout-aware Parsing · Row-as-Record · Entity Resolution · Local/Global D49Reasoning Model EngineeringRouting · Budget Sweet Spot · Interleaved Thinking · Trace Trust D50LLM Guardrails & SandboxingDual Gate · Jailbreak Classifier · Fail-closed Policy Gate · Sever the Lethal Trifecta D51Automatic Prompt OptimizationWorth Automating? · DSPy Compile · APE/OPRO Search · Few-shot & Overfit Governance D52Agentic Commerce & Payment ProtocolsAP2/ACP/x402 Layer Selection · Verifiable Mandate · HTTP 402 Settlement · Authz/Authenticity/Accountability D53Guardrail Engineering for Learning with AIInvert Generate-Verify · Learning vs Production Mode · Letting Juniors Use AI · Measure the Ladder D54Oracle Inventory for AI WorkflowsFour Oracle Classes · Same-Source Independence Trap · Reorder by Verifiability · Fail-Closed Guardrails D55Three Proven Learning SwitchesRetrieval: feedback + delayed post-test · Spacing: optimal fraction of deadline · Interleaving: confusable only · Four-Question Tool Filter D56Determinism EngineeringReal drift = batch non-invariance · seed is only "mostly" · steady evals with distributions · determinism costs throughput D57Semantic CachingThree-tier stack exact→prefix→semantic · threshold is a risk dial · scope key stops poisoning · hit-rate + false-hit two metrics D58Durable Agent ExecutionJournal replay resume · Checkpoint≠Durable · Exactly-once + idempotency keys · sink non-determinism into activities D59Prescribe Yourself a Manual DoseSkill decay is dose-dependent, cognitive first · assisted gain masks post-removal decline · Rx = AI-free core tasks + living through failure + accountability · verify the dose actually happened D60MCP Goes Stateless, and How to MigrateHandshake and Session-Id deleted, identity in _meta · Mcp-Method/Mcp-Name lifted into HTTP headers · MRTR replaces long-lived connections with a retry · mislabelled cacheScope = cross-user leak · DCR to CIMD D61Portable Agent Runtime GovernanceACS policy as data, PEP/PDP split · five checkpoints announced, eight in the spec · escalate = liftable deny + enforced_identity against TOCTOU · the snapshot contract is the real portability bottleneck D62Specialized Cyber Models and Cheap Vulnerability Discoverycompletion rate measures willingness, not capability; the Cyber model writes shorter reports · finding is 10x cheaper than exploiting, and the gap is closing · two-lane dependencies: version updates cool down, security fixes jump the queue · the bottleneck moves from discovery to validation D63Managed Agent Harnessesrent the brain and session, choose the hands · a sentinel for harness drift · keep side effects in custom tools · self-hosted sandboxes move execution, not data flow