histai/skillsets
4 skills0 installs24 starsGitHub
Install
npx skills add https://github.com/histai/skillsetsSkills in this repo
1Ai Model TrainerAI Model Trainer is an agent skill for solo builders and small research teams who need to train slide classifiers on HistAI CellDX without confusing training with whole-slide image purchasing. It documents how to submit classification jobs against the ML Jobs API, including full-pipeline runs for stronger results and quick mode for single-model iteration. The skill’s critical guardrails explain that training consumes pre-extracted feature vectors from roughly sixty-six thousand H&E slides, while IHC slides are not in the feature store—so agents must never route training requests through cohort creation, Datahub cohort APIs, payment, or WSI export flows. Authentication uses an API key header; costs are GPU-session hours on Azure rather than per-slide WSI fees. Use it when the user’s goal is a classifier trained on CellDX infrastructure; use cohort_builder only when they explicitly want to buy or download WSIs for external pipelines.0installs2Cohort BuilderCohort Builder gives coding agents secure, procedural access to the HistAI Pathology Datahub so solo researchers and indie ML teams can discover cases, apply clinical and technical filters, purchase cohorts, and export whole-slide images with metadata. It is aimed at builders who need downloadable WSIs for manual review, custom preprocessing, or pipelines outside CellDX’s feature store. The skill stresses a hard boundary from AI Model Trainer: cohort checkout incurs per-slide fees for H&E and IHC, while training on pre-extracted vectors is a separate GPU-compute workflow without WSI purchase. Agents should invoke this skill only when the user explicitly wants to search, buy, or export slide files—not when they ask to train a classifier on CellDX infrastructure. Proper use reduces costly mistakes where training APIs are skipped in favor of unnecessary WSI purchases, or vice versa.0installs3Slide Analyzerslide-analyzer is an agent skill that automates the CellDX inference path for a builder's own whole-slide images (WSIs). It uploads from local paths or URLs into My Cases, navigates workspaces through slides, manages the public and custom AI widget store, runs single-slide or batch inference, polls until outputs including segmentation masks are ready, and deletes slides or cases to control storage spend. The documentation stresses picking the correct HistAI skill: cohort_builder buys public slides with per-slide pricing, ai_model_trainer burns GPU hours on server-side training, while this skill only touches user-uploaded cases billed via storage and analysis credits. Solo builders shipping digital pathology or research tooling use it so agents do not confuse Datahub purchases with private-case inference. Requires a valid CELLDX_API_KEY and careful cost awareness before large batch runs.0installs4Training Monitortraining-monitor is an agent skill for HistAI CellDX ML jobs at `https://prod.celldx.net/v1/ml-jobs`, authenticated with X-API-Key. It checks job status, pulls epoch metrics (train/loss, val/loss, val/accuracy, val/auroc, val/f1_macro), tails trainer logs, lists filtered jobs, stops active runs, and patches hyperparameters such as learning_rate, epochs, and early_stop patience when a job is STOPPED or FAILED. The skill encodes how agents should present metrics: decreasing val/loss is healthy; rising val/loss with falling train/loss signals overfitting; plateaued val/auroc suggests stopping. Solo builders training pathology or custom models on CellDX use it to avoid babysitting dashboards manually and to close the loop back into ai_model_trainer-style re-runs after adjustments. It complements slide-analyzer (inference on My Cases) and does not replace launching the initial training job.0installs