agent-skills

Map of the boettiger-lab data-and-agent stack — which GitHub repo owns which piece of the data → STAC → MCP → LLM-agent chain. Use when unsure which repo to edit for a given problem, when routing cross-repo issues, or when the root cause could plausibly be in more than one layer.

Manage S3 object storage on the NRP (National Research Platform) Nautilus cluster, which uses Ceph S3 (not AWS). Covers public vs internal endpoints, rclone usage, bucket creation, public-read policies, CORS configuration, credential handling, and syncing to source.coop. Use when working with S3 buckets on NRP Nautilus, uploading or downloading data, configuring bucket access, or syncing datasets to source.coop.

Use when storing, versioning, or sharing large files alongside a GitHub repo — model checkpoints, datasets, build artifacts, binaries, fixtures — and you want collaborators to use their GitHub access rather than separate cloud credentials. Covers ghcr.io, oras CLI, content-addressing, public/private access, and the tradeoffs vs git-lfs, S3 buckets, and GitHub Releases.

Use when browsing, querying, or reasoning about what datasets exist in a STAC catalog — especially before concluding a dataset is missing or absent. Triggers include STAC browser URLs, s3 stac-collection.json links, or any question about what data is available in a catalog.

Analyze geo-agent app behavior from LLM proxy and MCP server logs, identify inefficiencies, and trace root causes to the correct layer: STAC metadata, MCP tool descriptions, geo-agent framework, or app system prompt. TRIGGER when: reviewing agent behavior, analyzing LLM logs, improving system prompts based on observed failures, debugging tool-use loops, or optimizing query patterns in geo-agent apps.

Deploy and manage workloads on the NRP Nautilus Kubernetes cluster. Covers batch jobs (opportunistic priority class, resource requests, GPU node avoidance), ingress with HAProxy CORS and timeout annotations, NRP usage policies, and credential wiring. TRIGGER when the user mentions: kubectl, k8s, kubernetes, NRP, Nautilus, rollout, restart deployment, apply yaml, pod, job, namespace, ingress, or any cluster operation. Namespace is 'biodiversity'. Always load this skill BEFORE running any kubectl command.

Apply preferred branch-protection rulesets and squash-merge-only settings to GitHub repos. Two profiles: 'standard' (no approval required) for smaller projects and 'strict' (1 approval, dismiss stale reviews) for larger ones. TRIGGER when: setting up a new repo, applying branch protection, configuring merge settings, or the user mentions rulesets.

Authenticate to GitHub using the boettiger-lab-llm-agent GitHub App. Covers minting installation tokens from an age-encrypted private key, git push/pull via credential helper, gh CLI usage, and the full session workflow. TRIGGER when an agent needs to: push to GitHub, create/comment on PRs or issues, use gh CLI, or perform any GitHub API operation in the boettiger-lab org.

Non-obvious MapLibre GL JS expression gotchas: legacy vs modern filter syntax, when to use 'match' vs 'in' for categorical membership, and other traps LLMs commonly fall into. Use when writing filters, paint expressions, or default_filter/default_style config for MapLibre layers.

Process geospatial datasets into cloud-native formats (GeoParquet, PMTiles, H3 hex Parquet) using the cng-datasets CLI and NRP Kubernetes. Covers the full workflow: URL verification, raw S3 upload, YAML generation, cluster deployment, monitoring, and documentation. Use when processing any geospatial dataset in the data-workflows repo, or when working with the cng-datasets CLI.

Read and convert remote geospatial files with GDAL/OGR. Covers the VSI virtual filesystem layer (/vsicurl/, /vsis3/) that enables remote access without downloading, when to use each prefix, ogrinfo and ogr2ogr patterns, geometry edge cases, and when to use DuckDB instead. Use when inspecting or converting GDB, GPKG, Shapefile, GeoTIFF, GeoParquet, or other formats locally or from remote URLs.

Critical gotchas for DuckDB spatial extension: the ST_Read() vs read_parquet() distinction, vendored GDAL limitations, S3 secret configuration for non-AWS endpoints, and URL style rules. Use when working with geospatial data in DuckDB, especially with remote files on S3 or HTTPS, or when hitting errors with ST_Read() and Parquet files.