setup-aftergrid — Add the optional Engine for evidence and reproducibility.

Optional Engine setup. Not required for the portable skills.

What it does

Collects inputs, runs aftergrid setup, and reports what the command reported — six separate facts, none rounded up. unknown stays unknown.

It is user-invoked. Setup writes files and installs a Claude Code hook, so a human asks for it.

When to reach for it

You start it: /setup-aftergrid. Use it for typed Findings, retained evidence, and publication controls. Do not run it to try diagnose-change on a CSV.

Inputs

Every input is a flag on the command; the skill's first job is to collect them in one round rather than guess.

Input	Notes
Instance directory	Defaults to `analytics/`.
Backend	Optional, and asked about only once. With no `--adapter`, the Instance is written for the recorded path: your harness runs the SQL and `aftergrid record` writes down what it ran. The skill asks one question — do you want aftergrid to be able to re-run these queries itself, for `check --mode rerun` and Revisit? — and only a yes turns into `duckdb` (a `.duckdb` file or a directory of `<table>.csv` files, inside the Instance root) or `postgres`.
Postgres connection	Only when the answer was Postgres. The name of the environment variable holding the connection string, never the string. A write-capable role is refused, and there is no flag that accepts one.
Owner	Name and contact — where a Reader's flag lands.
Publication	`owner/repo`, the automation login that opens Finding pull requests, and one or more trusted approver logins. The automation login may not be a trusted approver.
Hook settings	Path to the Claude Code settings file, if not `<repo>/.claude/settings.json`.

Setup runs without the publication identities. The Instance still works and still renders drafts; the report then says publication is not configured.

What it verifies

Each of these is a separate step in the report, reported completed, incomplete or skipped:

scaffold — which files were created, which were kept unchanged, and which were kept and differ from what setup would have written. Nothing is ever overwritten.
dependencies — Node's version, the mattpocock-skills skills (grilling and writing-for-agents), and the prebuilt DuckDB binding. A missing one carries the exact install line.
connection — the source was really opened and its capability matrix was read from the adapter. For DuckDB, "role probing unsupported" is the correct answer, not a gap. With no adapter the step is skipped and the report says what the recorded path gives (artifact_replay, verified hashes, aftergrid record) and what it costs (capture and execute refuse, check --mode rerun is rerun_unavailable, Revisit and unattended intake are unavailable). A skipped connection does not make the setup incomplete. A source whose tables are larger than the Instance's admission cap still connects and still passes setup: capture refuses those tables with admission before reading a byte, and the report says so and names the windowed-Instance pattern (docs/contracts/adapters.md, "Large sources") rather than a narrower capture.
hook — the guard is installed and its self-test just blocked a write and allowed a read. Installed without a passing self-test is reported as exactly that.
publication_preflight — whether the policy could ever produce a verified approval. Without a GitHub token this is unknown, and unknown is what it says.
smoke — a throwaway Finding ran new → check → draft render in a temporary copy. The Instance itself is left holding only the scaffold.

Setup is safe to rerun: it never overwrites a file, never writes a credential, and records completed steps in <instance>/.aftergrid-setup.json so a rerun resumes. --dry-run writes nothing at all.

What it does not verify

That the data is correct.
That aftergrid.yaml is protected by CODEOWNERS and branch protection — that is repository configuration, and nothing in setup can check it.
That any Finding has been approved, or that a trusted approver will ever read one.
That the guard covers a query path it says it does not cover. The hook is not a shell sandbox; source permissions and runtime isolation remain the boundary (docs/contracts/hook.md).

The full contract, including every limit: docs/contracts/setup.md.

It's working if

The skill asks for the inputs it does not have, in one round, and guesses no login, contact or warehouse path.
It does not ask for a warehouse path or a connection variable at all until you have said you want rerun and Revisit. No answer, or "not yet," sets the Instance up on the recorded path.
On an adapterless Instance, the summary names aftergrid record as the route and lists what is unavailable, rather than reporting the missing adapter as something still to fix.
The summary it gives back names each of the six steps with the word the command used for it.
unknown survives into the summary as unknown.
The hook is described as protecting something only when the self-test line is in the output.
The next command it hands you is /grill-question and then aftergrid new finding <slug>.

Where it fits

Run-once Engine setup, not part of ordinary analysis. The fourteen portable skills work without it. After a working Instance, start with ask-aftergrid, grill-question, or analyze.

Keep exploring← Back to the skill library Try itInstall your first skill ↗