snowman

Take working-tree changes from branch to merged PR in one pass — branch off main, commit smart-git-commit style, open a PR, then squash-merge and delete the local and remote branch. Command-only: /ship asks before merging, /ship clean goes straight through. Detects unrelated changes and offers to split them into separate branches/PRs.

Version-aware Prefect 3 guidance: a live docs-lookup protocol, CLI-first instance queries, and house standards. Prefect 2.x out of scope.

Forces the main thread to act as an orchestrator and delegate ALL work to subagents instead of doing it itself — plan, dispatch, verify with a separate agent, synthesize. Agent-neutral across Claude Code, Cursor, and Cortex Code.

Groups changed files by affected area, creates one conventional commit per group, then pushes to remote — with safety rules against force-pushes, skipped hooks, and committed secrets.

Invocation discipline and failure triage for running dbt — a static preflight script kills environment failures before the first command, invocation rules prevent the self-inflicted ones (sandboxed shells, piped output, silent empty selections), and a signature-indexed catalogue maps error strings to ranked causes and fixes, including dbt-fusion quirks. Bootstraps a committed per-project context file.

Prefect 3 agent skill: greenfield scaffolding + project audit with a live docs-lookup protocol (llms.txt). Part of relentless-data-skills.

Bootstrap a dlt ingestion project: install the dltHub AI Workbench project-scoped (only the toolkits the project needs), then layer Relentless Data house conventions (Snowflake, Prefect, DuckDB dev loop) as a committed always-on rule. Idempotent re-runs add new source types.

Produce a single self-contained HTML visual report — an explainer, writeup, or diagram-heavy document built with Tailwind and Mermaid via CDN plus hand-crafted CSS/SVG.

Read-only Snowflake exploration via the snow CLI — schema discovery, profiling, hypothesis testing, and data-quality investigation. Bootstraps a committed per-project context; a guardrail wrapper hard-enforces read-only execution and stages user-requested DML/DDL as scripts for manual execution.

Implement a GitHub issue end-to-end: claim it, branch, code in small commits, run tests + a runtime smoke check, and open a draft PR — with explicit stop conditions instead of improvising.