Kit Setup

Everything in this kit works on Claude alone. This skill records which optional extra engines exist on this machine, so other kit skills can offer second opinions. Detect, then confirm. Never ask the user what subscriptions or plans they have, and never list, describe, or ask about tools that weren't detected.

Phase 1 — Engines

1. Detect

One Bash call:

bash
command -v codex; command -v gemini; date +%Y-%m-%d

For each CLI found, run a quick <tool> --version; if that fails, say it looks installed but isn't currently working, record it as unavailable, and move on — don't troubleshoot here. Version output only means "detected," not "usable."

2. Confirm each FOUND tool, in plain language

One short yes/no question per tool that was found — and nothing about tools that weren't:

• codex: "I found the Codex CLI installed — want me to use it for second opinions and research? (yes/no)"
• gemini: "I found the Gemini CLI installed — want me to use it for second opinions and very long documents? (yes/no)"

If nothing was found, skip the questions entirely. Either way, close Phase 1 with one line: "If you later install the Codex or Gemini CLI, run /kit-setup again."

3. Smoke-test each accepted tool

For each detected tool the user accepts, create one private temp directory with mktemp -d, then run one tiny smoke test through the shipped wrapper:

bash
smoke_dir="$(mktemp -d)"
printf 'Reply with exactly: starter-kit-ok\n' | bash "${CLAUDE_PLUGIN_ROOT}/scripts/codex-run.sh" "$smoke_dir/codex.txt" light
printf 'Reply with exactly: starter-kit-ok\n' | bash "${CLAUDE_PLUGIN_ROOT}/scripts/gemini-run.sh" "$smoke_dir/gemini.txt"

If the wrapper succeeds and the output is non-empty, mark the engine available: true. If it fails, mark it unavailable but preserve what happened, for example { "available": false, "detected": true, "auth_unverified": true }. Keep the user-facing message plain: "I found Codex/Gemini, but it did not answer a test prompt yet. Everything still works on Claude."

4. Write the config

Target: ~/.claude/starter-kit/engines.json. Run mkdir -p ~/.claude/starter-kit first. If the file already exists, Read it and preserve any fields you don't recognize — newer kit versions may have added some. Schema (full contract: ${CLAUDE_PLUGIN_ROOT}/references/engines-contract.md):

json
{
  "schema_version": 1,
  "codex":  { "available": true },
  "gemini": { "available": false },
  "oracle": { "available": false },
  "paste_loop": ["chatgpt", "gemini"],
  "updated": "<today, YYYY-MM-DD>"
}

• available: true only if the CLI was detected, the user said yes, and the smoke test returned a non-empty answer.
• If the user says no to an installed tool, write { "available": false, "user_declined": true } so a re-run doesn't quietly switch it back on.
• paste_loop: write the default shown above on first write; keep any existing value on re-runs.
• oracle: always write { "available": false } — the field is reserved for a future release; this version doesn't probe or enable it.

Phase 2 — Personalize (optional; offer once)

Ask once: "Want to spend two minutes personalizing how I write for you? Five quick questions — totally skippable."

If yes, ask one at a time, conversationally:

1. What kind of work will you mostly use this for — and what's your role or field?
2. How do you like answers — short and direct, or thorough and explained?
3. What annoys you in AI writing? (buzzwords, bullet overload, hedging…)
4. How formal should drafts be by default?
5. Anything I should always do — or never do?

Then write ~/.claude/starter-kit/my-preferences.md: start from the template at ${CLAUDE_PLUGIN_ROOT}/references/prompt-preferences-TEMPLATE.md and fill it in from their answers. If the template is missing, write a simple file with one short section per question. If they skip, say they can run /kit-setup again anytime — and don't offer a second time this run.

Finish — report in plain language

Two or three sentences, no jargon, no JSON. Pattern:

> "You're set up. Everything runs on Claude — that's the normal setup and it all works."

Add one sentence per enabled engine (e.g. "Codex is also connected, so you can ask it for second opinions with /codex.") and one if preferences were saved. If nothing extra was found, the first sentence is the whole report.

Re-running

Safe anytime. Re-detect and update engines.json: a newly installed tool gets confirmed (Phase 1, step 2) and added; a removed one gets marked unavailable; unknown fields and paste_loop are preserved. If my-preferences.md already exists, offer to update it rather than starting over.