What problem does it solve? LLM research experiments consume GPU hours that are hard to track, and teams often discover budget overruns only after the cloud bill arrives. This Skill computes per-run cost estimates from run duration and GPU pricing, aggregates them into a project-level cost report, and warns before the compute budget is exhausted. ## Core Features & Use Cases - Per-run cost estimation: Reads each run's metrics.json, multiplies elapsed time by GPU hourly price and GPU count, and writes a cost_estimate block back into the metrics file. - Cumulative budget watch: Generates a 06_COST_REPORT.md comparing used GPU-hours against the compute_budget_gpu_h set in Phase 1, with status tiers (Safe, On track, Caution at 80%, Over budget at 100%). - User price overrides: Supports a cost_overrides.json file so actual contract or spot pricing replaces the bundled public price table (gpu_price_table.json covering A100, H100, H200, L40S, L4, RTX 4090/3090, TPU v4, Apple M-series). - Use Case: After a fine-tuning run finishes in Phase 6 of an auto-research project, dispatch this Skill to record that the 1-hour A100 run cost about $2.00, update the cumulative report, and get a warning if the project has crossed 80% of its GPU-hour budget. ## Quick Start Use the research.cost.estimate skill to estimate compute costs for all completed runs of my project and show the remaining GPU-hour budget.
