Dlt Bootstrap

dlt-bootstrap

Set a project up for dlt pipeline development the house way: install dltHub's official AI Workbench project-scoped, then layer Relentless Data conventions on top as an always-on rule. The workbench's toolkits own pipeline building. After bootstrap, step aside: day-to-day work runs through the workbench's own entry points (/find-source, /explore-data, ...), with the house rule applying automatically because it is a rule, not skill-mediated.

First: check for the house rule

Search the project for a committed dlt-house-conventions.md.

• Present → incremental mode: read its frontmatter (source_types, toolkits_installed), then do only what's missing. This typically means installing an additional toolkit for a new source type and updating the frontmatter.
• Absent → full bootstrap (below).

If invoked proactively (the user was working on something else and the rule is merely missing), don't start the bootstrap. Note that the project lacks dlt-house-conventions.md, offer to run the setup, and return to the user's actual task unless they accept.

Detect, never ask

Establish by inspection, creating what's missing in the install step:

• pyproject.toml / venv: greenfield vs existing project (same path either way).
• dlt[hub] already a dependency? uv run dlthub --version succeeds?
• Which agent is in use (Claude Code / Cursor / Codex), usually obvious from the session. Pass it explicitly to --agent rather than relying on detection.
• Existing .dlt/ directory, existing pipelines, git repo state.
• Where dlthub ai init placed its rules (see "Write the house rule").

Interview (three questions max, as plain prose)

1. Source type(s) this project ingests: REST API, SQL database, or files? Drives toolkit selection. Multiple is fine.
2. Pipeline / dataset name. Only ask if not derivable from the repo name.
3. Destination. Snowflake is the house default; confirm, allowing a per-client override (e.g. BigQuery) without editing this skill.

Install (verified fast path)

Say in one sentence what you're about to install, then run the sequence straight through. Speak again only for a failure, a decision the user must make, or the hand-off, not per command.

uv init                          # only if no pyproject.toml
uv add "dlt[hub]"
uv add "dlthub[mcp]"             # MCP server deps; without this the workspace MCP never starts
uv run dlthub init               # workspace init; follow its instructions (uv sync)
uv run dlthub ai init --agent claude
uv run dlthub ai toolkit install <toolkit> --agent claude   # per source type
uv run dlthub ai status          # verify: agent detected, toolkits + entry skills listed

If any command fails or a flag is rejected, suspect upstream drift before debugging: consult references/docs-map.md (start at the workbench README) and re-derive the command. Never invent flags.

Known upstream trap (verified 2026-06): if dlthub ai status warns to pip install "dlt[workspace]", ignore it. That extra does not exist; the correct fix is uv add "dlthub[mcp]" (already in the fast path above).

Toolkit policy

Install init (automatic dependency) plus only the pipeline toolkits matching the interview answer:

• Never install quick-start (this skill is the entry point) or dlthub-platform (we deploy via Prefect, not the dltHub platform).
• data-exploration, data-quality, transformations only on explicit request. The last two require a dltHub sign-up; flag that before installing and skip gracefully if the user has no account.
• Record every installed toolkit in the house rule's frontmatter.

Write the house rule

1. Copy references/rule-template.md into the project and fill every <placeholder> from the interview + detection. Then re-read the frontmatter: every key from the template present, one per line, no < left (easy to mangle destination / dev_destination). Fill the template, don't grow it: no extra sections, no restating what the workbench's own rules already cover.
2. Name it dlt-house-conventions.md and place it in the same location where dlthub ai init installed its own rules (find its dlthub-workspace.md) so it is always-on for the same agent. If that agent merges rules into a memory file (CLAUDE.md / AGENTS.md), append the filled template as a clearly delimited managed section instead.
3. Ensure the agent memory file carries dltHub's credential-safety line (it is in the template); their installer does not add it for Claude Code.
4. Commit the rule and the workbench-installed files. Never commit .dlt/secrets.toml. Neither uv init nor dlthub ai init gitignores it (the latter only writes .claudeignore). Add secrets.toml to .gitignore if absent, then verify: git check-ignore .dlt/secrets.toml.

Verify, then hand off

• uv run dlthub ai status shows the agent and pipeline toolkits, but it omits init; use uv run dlthub ai toolkit list to confirm the full set.
• MCP: confirm dlt-workspace-mcp is registered for the agent (for Claude Code, check the project .mcp.json).
• Tell the user to restart their agent session now. The workbench skills (/find-source, ...) and the MCP server are not active until they do.
• Tell the user the working loop: /find-source → scaffold → secrets via the MCP secrets tools → debug → validate on DuckDB → harden (incremental loading, remove dev limits) → wrap in a Prefect flow (prefect) → ship via /ship → inspect what landed in Snowflake with snowman.

Guardrails

• This skill sets up; it does not build pipelines, and it does not fork or re-teach workbench skill content, since upstream owns that. If asked to build a pipeline before bootstrap, bootstrap first, then route to the workbench's skills. When the bootstrap surfaces adjacent problems, such as missing tests, an untidy pyproject.toml, or a stale dependency, report them and finish the bootstrap; don't fix them on the way through.
• Bootstrap is a linear install sequence. Run every step inline, never fanned out to subagents.
• Credential safety is enforced at runtime by the house rule's Secrets section; see references/rule-template.md.
• Incremental re-runs must be idempotent: re-installing an existing toolkit or re-writing an unchanged rule is a no-op, not an error.