ck-setup

name: ck-setup description: First-time setup for new cruijff_kit users — walks through claude.local.md configuration. Re-run for a health check on an existing config.

ck-setup

Naming note: the ck- prefix is deliberate. A bare setup would collide with a built-in skill of that name, so the prefix is load-bearing — a future naming audit shouldn't "fix" it away.

You help users get a working cruijff_kit environment. There are two modes — pick based on whether claude.local.md already exists:

• Greenfield — no claude.local.md. Walk the user through claude.local.md.template interactively, explaining what each section is for, then write a populated claude.local.md.
• Validate — claude.local.md exists. Read it, scan for unresolved placeholders, check that required fields are filled, run a few lightweight environment probes, and report a green/yellow/red summary with concrete next actions.

Default to never overwriting an existing claude.local.md. If one is present, validate-mode is the only option unless the user explicitly asks to start over (in which case follow the "Start over" procedure in walkthrough.md, which renames the existing file to a dated .bak and then enters greenfield mode).

Audience

The user may be one of two profiles. Use what they say (and what their environment looks like) to figure out which:

• New user — has just cloned the repo, hasn't configured anything, may not know what cruijff_kit does. They benefit from rationale ("this matters because…") on each setting, not just a question.
• Returning user — has configured their environment before, may have just pulled an update that added a field, or is debugging a workflow that's failing. They want a fast "is anything broken?" check, not a tutorial.

Mode detection above maps roughly: greenfield ↔ new user, validate ↔ returning user. But not always — a returning user on a fresh clone is also greenfield. Don't assume.

Your task

1. Detect mode (greenfield vs validate) by checking for claude.local.md.
2. Greet the user and tell them what's about to happen — mode, estimated time, the kind of questions or checks they'll see.
3. Greenfield: walk through the template section-by-section, prompt the user for each value, write claude.local.md at the repo root.
4. Validate: read the existing file, scan for placeholders and missing fields, run environment probes, report findings.
5. Point the user at next steps (docs/PREREQUISITES.md for accounts and Software Carpentry tutorials; /design-experiment for their first experiment).
6. Log the run to logs/ck-setup.log.

See the modular sub-files for details:

• walkthrough.md — greenfield Q&A flow, section-by-section
• validation.md — placeholder scan, required-field check, env probes
• logging.md — action types and log file location

Prerequisites

This skill is itself the prerequisite-check for other skills, so its own prereqs are minimal:

• The user has cloned the cruijff_kit repo and is in its root directory (where claude.local.md.template lives).
• A shell with bash, grep, and standard Unix tools.

If claude.local.md.template is missing, the user is either not in the repo root or has an unusual checkout — stop and ask before proceeding.

Workflow

1. Detect mode

ls -la claude.local.md claude.local.md.template

• Both present → validate mode.
• Template only → greenfield mode.
• Template missing → stop. Ask the user where they are; the skill assumes the repo root.

2. Greet the user and set expectations

After detecting mode and before diving into work, greet the user and tell them what's about to happen. Tone: friendly, like welcoming someone on day one. Don't assume technical fluency — name what you're going to do in plain language.

Greenfield mode (no claude.local.md):

Hey! Welcome to cruijff_kit. I'll walk you through creating your claude.local.md — a personal config file the toolkit uses to learn about your HPC cluster, your SLURM account, and your conda environment.

This takes about 5 minutes. I'll ask ~10–15 short questions, one at a time, and explain why each one matters. If you're not sure of an answer, you can say [skip] and we'll come back to it later. Before I write the file, I'll show you a summary and ask you to confirm.

Ready when you are.

Validate mode (claude.local.md already exists):

Hey! I see you already have a claude.local.md. I'll do a quick health check — about 30 seconds — to make sure your environment is set up right.

I'll scan for any placeholders that haven't been filled in, check that the required fields are present, and run a few lightweight probes (does your conda env exist, is your scratch directory reachable, is gh on PATH). I won't change anything in your config. At the end you'll get a green / yellow / red summary with concrete next steps for anything that needs attention.

Then proceed to the chosen mode below.

3a. Greenfield → walkthrough.md

The template is 158 lines and covers HPC, SLURM, conda, GitHub, wandb, paths, and quick commands. Walk it section-by-section. For each section:

1. Show the section name and a one-line "this matters because…" rationale.
2. Read the template's placeholder values. Auto-detect what you can (whoami, id -gn, pwd). Treat auto-detected values as defaults to confirm, not facts to write. In particular, id -gn returns the user's primary group token, but on some clusters the directory case differs from the group token (e.g. MSALGANIK directory vs msalganik group); ask the user to confirm the casing that appears in their actual paths.
3. Ask the user for non-detectable values one at a time. Don't ask for everything at once — that's a wall of questions, not a walkthrough.
4. Allow [skip] answers — the user may want to come back to that section later. Skipped sections keep their <placeholder> values; validate-mode will catch them on the next run.

When all sections are done, write the populated content to claude.local.md at the repo root.

3b. Validate → validation.md

Three classes of check:

1. Placeholder scan: grep -n '<[^>]*>' claude.local.md in the settings sections only (HPC Environment, SLURM Defaults, Common Paths, Weights & Biases, GitHub Configuration). Lines under "Quick Commands" legitimately contain <placeholder> tokens — don't flag those.
2. Required-field check: confirm a small set of must-have values are non-empty and non-placeholder: Cluster, Username, Group, Scratch directory, Default conda environment, Account (SLURM).
3. Env probes (lightweight only — don't actually run jobs or hit external APIs):
   • Conda env exists: conda env list | grep -q <env_name>
   • Scratch directory exists and is writable: test -w <scratch_dir>
   • gh is on PATH after activating the documented conda env: which gh
   • Module system available if the file mentions module load: command -v module

Report a per-check pass/fail with the concrete remediation step beside each fail. Don't return a binary "all good / all bad" — partial-pass is normal and useful.

4. Point at next steps

After either mode, finish with two pointers:

• docs/PREREQUISITES.md — accounts and skills the user should have (HuggingFace token, Git/shell/conda comfort). The skill does not validate accounts — don't try to curl HF with the user's token, that's a credential-leak risk and brittle.
• /design-experiment — the natural next skill once the environment is configured.

Acceptance / validation contract

A successful run satisfies all of:

• ✓ Either claude.local.md was created (greenfield) or a structured validation report was produced (validate).
• ✓ The user knows what the next step is (/design-experiment or "fix these N findings first").
• ✓ logs/ck-setup.log was appended with the action types from logging.md.
• ✓ No existing claude.local.md was overwritten without explicit user consent.

If any of these fail, the skill should error loud, not silently produce something half-correct.

Out of scope (explicitly)

• Validating HuggingFace tokens, Princeton SSO, or any other authenticated external service. Point at docs/PREREQUISITES.md instead.
• Running fine-tuning, eval, or any GPU work as a "smoke test." That's the job of the workflow test specs in .claude/workflow_tests/ (see CLAUDE.md), not first-time setup.
• Splitting claude.local.md into Claude-specific and tool-agnostic config files. That's tracked separately as #457.
• Auto-installing the conda environment. The user runs make install per claude.local.md.template — the skill verifies the env exists, doesn't create it.
• Migrating an old claude.local.md to a new schema. If the template gains fields, validate-mode will flag the missing ones; the user fills them in by hand.