AI Agent Workflow Collection

Writes Gherkin .feature files from a user story so non-technical stakeholders can sign off on behavior. Use before implementing any user-visible behavior change.

Writes Gherkin .feature files from a user story so non-technical stakeholders can sign off on behavior. Use before implementing any user-visible behavior change.

Coordinates heterogeneous agents through a shared file-system blackboard at docs/blackboard/ instead of message passing. Use when the next action depends on accumulated shared state, agents join and leave at different times, and you want git-diffable coordination.

Google BigQuery remote MCP conventions (preview) — GCP Workload Identity Federation auth never a static service-account key, a read-only bigquery.jobUser + dataViewer role, mandatory partition pruning, and in-place query streaming. Use when querying BigQuery, wiring the BigQuery MCP, or setting up read-only warehouse access on GCP.

Databricks Managed MCP conventions (Public Preview) — Unity Catalog ACLs as the upstream gate, OAuth only never a personal access token, Genie spaces treated as downstream lineage consumers, and Mosaic AI model serving published by humans/CI not the agent. Use when querying Databricks, wiring the Databricks MCP, or working with Unity Catalog.

dbt-core conventions — contract-first models with contract.enforced true, mandatory unit tests, a semantic layer as the single metric source of truth, the dbt remote MCP as the governed agent surface, and auto-activating dbt-agent-skills. Use when building dbt models, contracts, unit tests, or semantic-layer metrics, or wiring the dbt MCP.

DuckDB + MotherDuck conventions — local-first SQL over Parquet/CSV/JSON with no warehouse round-trip, MotherDuck OAuth (never a static token), sample-then-scale even locally, and DuckDB as a join engine for Polars via pl.SQLContext. Use when querying local data files or using DuckDB/MotherDuck. The duckdb MCP server ships with this plugin.

UK AISI Inspect AI eval conventions — the dataset/solver/scorer three-piece shape, Docker-sandboxed task isolation, the 200+ pre-built inspectevals starter tasks, and wrapping Claude Code/Codex/Gemini CLI as the agent under test. Use when authoring or running Inspect AI evals, sandbox-isolated agentic evals, or adding a custom eval task.

Langfuse OSS LLM-observability conventions — production traces graduate to the next eval dataset, cross-family LLM judges, versioned reproducible datasets, the MCP at /api/public/mcp with public+secret keys, and PII masking in trace inputs. Use when instrumenting LLM tracing, wiring the Langfuse MCP, or triaging traces and regressions.

marimo reactive .py notebook conventions — notebooks are git-diffable Python files edited directly, the reactive dependency graph re-runs downstream cells, Restart-and-Run-All via marimo export script, and the marimo pair agent surface. Use when building or editing marimo notebooks, or when choosing a reactive notebook runtime over Jupyter.

MLflow tracking conventions — MLflow 3.5.1+ as the minimum, every training run logged, GenAI tracing for LLM apps, the run-comparator agent flagging suspicious improvements, and registry promotion to production gated to humans/CI. Use when wiring MLflow tracking, the MLflow MCP, comparing runs, or registering model artifacts.

Polars idioms — lazy frames via scanparquet with a single final collect, withcolumns over column assignment, DuckDB via pl.SQLContext for heavy joins, Polars v1.40 as the minimum, and pandas only as ecosystem glue. Use when doing DataFrame work, feature engineering, or choosing a DataFrame engine for analysis or ML.

Snowflake Cortex Managed MCP conventions — OAuth only never a static password, Cortex Agent Evaluations as the agent-run eval surface, a SELECT-only read-only role, and mutation routed through migration PRs. Use when querying Snowflake, wiring the Snowflake Cortex MCP, or setting up read-only warehouse access.

Astral uv toolchain conventions — lockfile-frozen by default with uv lock --frozen and uv sync --frozen, uv add for atomic pyproject + lock updates, frozen-only CI/production installs, and uv v0.7+ as the minimum. Use when managing Python dependencies, freezing environments, or setting up a reproducible Python project.

Weights & Biases conventions — the official wandb-mcp-server with WANDBAPIKEY auth, Weave as the GenAI trace surface, Reports as the human-review surface the agent can author, refuse-to-delete-runs discipline, and pairing with MLflow. Use when wiring the W&B MCP, tracing LLM runs in Weave, or authoring W&B Reports.

Exploratory notebook discipline — marimo first and never raw NotebookEdit on .ipynb, Restart-and-Run-All as the only acceptance test, sample-then-scale on every warehouse query, Polars + DuckDB over pandas, and chart-critic review. Use when .claude/HARNESS.toml selects data/data-analyst-notebook, or when doing ad-hoc analysis, exploratory reporting, or building reactive reproducible notebooks.

Shared rules for any data engineering work — warehouse is read-only via Managed-MCP/OAuth, sample-then-scale on every query, every reported metric traces to a logged query and data hash, evals live in a package separate from model/prompt code, cross-family LLM judges, and a dataset card for every dataset. Auto-loads for any analytics, ML, LLM, or notebook task.

LLM-app discipline — the three-tier assertion/judge/human eval ladder with no higher tier before the lower, cross-family judges, a single pinned model-version env var, prompt-regression suites, and production tracing. Use when .claude/HARNESS.toml selects data/llm-app, or when building RAG, agentic pipelines, prompt-driven products, evals, or LLM observability.

ML-pipeline discipline — every training run logged, lockfile-frozen environments, every seed pinned, the eval suite as a separate package the model never imports, a data hash per dataset, and leakage guards. Use when .claude/HARNESS.toml selects data/ml-pipeline, or when training models, writing train.py, running evaluation suites, packaging or registering model artifacts.

Scaffold a new dbt model with contract, unit test, and model body in the right order. Seeded from dbt-labs/dbt-agent-skills (Feb 9 2026).

Write the model contract (column types + constraints) before the model body. Every staging+ model declares contract.enforced true.

Define semantic-layer entities, measures, dimensions, and metrics in the MetricFlow shape. Seeded from dbt-labs/dbt-agent-skills (Feb 9 2026).

Write dbt unit-test specs with given/when/expect shape. Seeded from dbt-labs/dbt-agent-skills (Feb 9 2026).

AWS conventions — 15-minute STS session ceiling, AFT-on-Control-Tower account bootstrap, EKS Pod Identity over IRSA, OIDC trust federated to STS with tight sub-claim scope, and env/blast-radius account tagging. Use when targeting AWS for IaC, CI/CD trust policies, or EKS.