Skill UI

Pattern for controlling LLM API costs while preserving quality. Combines complexity-based model routing, immutable budget tracking, retry logic limited to transient failures, and prompt caching. Use when building LLM API apps, processing batches with varying difficulty, or enforcing budget guardrails in production.

This skill monitors operational costs against defined budgets by reading accumulated spending and computing utilization. It emits a detailed 4-tier alert ladder (INFO, WARNING, CRITICAL, HARD_STOP) based on predefined thresholds (50%, 75%, 90%, 100%). It is essential for cost governance, ensuring resources are managed within budget limits, and preventing system overspending.

RouterBase acts as a unified, OpenAI-compatible API gateway, allowing applications to route requests across diverse large language models (LLMs) like GPT, Claude, and Gemini. It helps developers migrate existing OpenAI SDK calls, implement sophisticated primary/fallback model selection based on cost, latency, and quality, and manage multimodal payloads (image, audio, video) from a single, standardized endpoint.

This guide helps researchers determine the most appropriate conference venue (e.g., HPCA, ISCA, MICRO, ASPLOS) for their microarchitecture or systems research paper. It outlines key criteria, such as whether the contribution is a measured microarchitectural cost, a buildable mechanism, or a measurable characterization, and how to strategically route the project based on its novelty and evidence.

A detailed guide on how to write a rigorous and comparative 'Related Work' section for top-tier microarchitecture conferences (e.g., MICRO). It teaches authors to structure the section as a 'mechanism taxonomy,' differentiating designs based on structural and cost differences rather than merely listing citations. Key areas covered include sweeping academic venues, integrating industry disclosures, and verifying all venue attributions using tools like dblp.

A rigorous framework to prevent confident fabrication (hallucination) by treating competence as an evidence boundary. It guides users to classify claims—Grounded, Partial, or Ungrounded—based on verifiable sources. It mandates assessing the cost of wrongness before answering, fetching, or abstaining, ensuring that assertions are always backed by sufficient evidence.

This methodology provides a comprehensive framework for estimating buffered numbers—such as capacity limits, timeframes, or budgets—when making commitments under uncertainty. It guides users to calculate a buffer based on both residual estimation error and the potential cost of failure (breach cost), ensuring resilience against costly underestimation in critical systems or plans.

A rigorous decision-making methodology that quantifies the trade-offs when allocating scarce resources (time, budget, attention). Instead of merely evaluating a chosen path, this framework forces the user to identify the single best alternative forgone, calculating the net value delta (chosen value minus best forgone value). It is essential for feature prioritization, major commitments, and strategic planning to prevent decisions based solely on immediate perceived gains.

This framework guides decision-makers to structure their thinking by classifying decisions based on the cost and feasibility of reversal. It emphasizes treating low-cost, reversible choices (two-way doors) quickly, while reserving deep deliberation for high-cost, irreversible commitments (one-way doors). It helps mitigate analysis paralysis and premature over-engineering.

Read-only review of Caveman Cloud evidence covering cost, Cave Score, workflows, traces, latency, errors, routing, and savings. Use when asked what Caveman found or where LLM spend goes. Conclusions are based on current data and cite trace IDs.

A sophisticated model-routing advisor for generating, redlining, or rewriting contract language. It determines the optimal LLM based on user constraints (stakes, cost, speed, jurisdiction) and grounds its recommendations in legal benchmarking evidence. It provides a primary model, fallback, and critical warnings.

This tool serves as a model-routing advisor for high-stakes legal translation tasks, covering contracts, statutes, and case law. It does not perform the translation itself, but instead provides expert decision support by recommending the optimal primary and fallback LLMs based on language pair, required privacy level, cost, and speed. Crucially, it emphasizes that human legal-linguist review is mandatory, as no single LLM can guarantee certified legal accuracy.