get-started

Prompt file imported from jfranecki/JustinsClaude (.claude/commands/get-started.md). Fill in {{arguments}}, {{PLACEHOLDER}}, {{HOME}}, {{GITHUB_USERNAME}}, {{GITHUB_REPO}}, {{MAIN_CHECKOUT}}, {{JIRA_PROJECT_KEY}}, {{REPO_NOTES}}, {{CLONE_ROOTS}}, {{SLACK_USER_ID}}, {{SLACK_CHANNELS_TABLE}}, {{KOKORO_DIR}}, {{HISTORY_SCRIPT}}, {{CLAUDISH_SCRIPT}}, {{ORCHESTRATOR_FRAMEWORK_PATH}}, {{MEMORIES_REPO}} before use. Copyright stays with the author.

/get-started — command installer & wiring assistant

You are setting up this repository's Claude Code commands for the current user. The files under commands/ are templates: they contain {{PLACEHOLDER}} tokens for everything specific to a person, team, or codebase. Your job is to gather the values, verify the user actually has the required credentials and connections, fill the templates, and install working copies into ~/.claude/commands/ (and ~/.claude/workflows/). A command whose required prerequisites fail verification must NOT be installed — report what is missing and how to fix it instead.

Optional pre-selection from the user (may be empty): {{arguments}} — if non-empty, treat it as the list of commands to install and skip the selection question.

Work conversationally but efficiently: batch your detection commands, use AskUserQuestion for choices, and never invent a value you could not detect or confirm with the user.

Step 0 — Locate the repo and sanity-check

1. Resolve the repo root: git rev-parse --show-toplevel (fall back to cwd). Verify commands/ exists there and contains the template .md files. If not, stop: tell the user to cd into their clone of this repo and re-run /get-started.
2. mkdir -p ~/.claude/commands ~/.claude/workflows so installs can't fail on missing dirs.

Step 1 — What's on offer

Present this menu (with AskUserQuestion, multiSelect, unless {{arguments}} already chose). Briefly describe each:

Step 2 — Detect the environment (batch these)

Run in parallel where possible; record every result:

• echo $HOME → {{HOME}}
• gh auth status and gh api user -q .login → proposed {{GITHUB_USERNAME}} (confirm with the user — some people have separate work/personal GitHub accounts; the one that matters is the one with access to the repo they review in)
• command -v acli, command -v jira, command -v python3, command -v jq, command -v afplay
• test -n "$ELEVENLABS_API_KEY" && echo set || echo missing — on Windows also check the persistent user scope, which the current shell may not have inherited: [Environment]::GetEnvironmentVariable('ELEVENLABS_API_KEY','User')
• audio player for speak-api: command -v afplay ffplay mpv mpg123 cvlc — any one is enough
• Slack MCP: use ToolSearch with query select:mcp__claude_ai_Slack__slack_read_channel. If it resolves, Slack is connected. If not, the fix is: run /mcp and connect "claude.ai Slack".
• Kokoro: check the conventional spots (~/ToolboxRepos/kokoro, ~/Tools/kokoro, ~/kokoro) for a dir containing both .venv/bin/python and speak.py.
• history.py for coldstart: it ships in this repo at <repo>/coldstart-setup/history.py. Confirm it is present and that python3 -c 'import ast,sys; ast.parse(open(sys.argv[1]).read())' <path> parses it.
• ollama: command -v ollama, then probe the daemon with curl -sS --max-time 5 http://localhost:11434/api/tags (a running daemon returns JSON containing a models array). List what is pulled with ollama list.
• claudish.sh: check the conventional spots (~/ToolboxRepos/claudish/claudish.sh, ~/.claude/bin/claudish.sh) for the wrapper script.
• Platform: speak (Kokoro) plays audio via afplay/open, so it is macOS-first — on Linux or Windows, warn that speak.py needs its playback command swapped. speak-api auto-detects a player and needs no changes.

Step 3 — Interview for the values

Only ask for what the selected commands actually need. The full placeholder map:

Use absolute paths everywhere (expand ~ before substitution).

Step 4 — Verify before allowing use (the gate)

For each selected command, evaluate its gate. Required failures block installation of that command (skip it, explain the fix); optional failures install with a clear warning.

• onboard — no gate.
• orchestrator — no gate. Optional: <repo>/Orchestrator Framework/ present in the clone (it ships with the repo) — without it the command still works, carrying the tier rule inline, and only loses the full writeup and templates.
• coldstart — required: python3 present; <repo>/coldstart-setup/history.py present and parsing. Also installs the sidecar (Step 5). The command degrades gracefully with no transcript history — it reports that and delivers the filesystem pass — so an empty ~/.claude/projects/ is not a gate failure.
• bye — no gate. Optional: gh authenticated — without it, merge confirmation falls back to plain git checks.
• review-deep — required: gh auth status logged in. Optional: acli/jira present (without it, Jira spec-mapping degrades gracefully); each CLONE_ROOTS directory exists.
• pr-autoreview — required: review-deep being installed in this same run (or already present in ~/.claude/commands/); gh access to {{GITHUB_REPO}}; valid {{MAIN_CHECKOUT}}; python3 present. Also installs both workflow files (Step 5). Warn explicitly: this command POSTS REVIEWS to GitHub under their account when run — they should read commands/pr-autoreview.md before scheduling it.
• slack-updates — required: Slack MCP tools resolvable via ToolSearch; a plausible {{SLACK_USER_ID}}; at least one tracked channel resolved.
• speak — required: validated {{KOKORO_DIR}}; afplay present (macOS).
• speak-api — required: ELEVENLABS_API_KEY resolvable (key from elevenlabs.io → Profile → API Key; if missing, export ELEVENLABS_API_KEY="sk_..." in the shell profile, or on Windows [Environment]::SetEnvironmentVariable('ELEVENLABS_API_KEY','sk_...','User'), then start a new Claude Code session — a running session keeps its old environment); jq; at least one audio player from afplay/ffplay/mpv/mpg123/cvlc. If the key is only set for the persistent user scope and not yet visible to the running shell, install it but report status as ready after restart, not ready.
• memorize — required: <repo>/memories/ exists in the clone (it ships with the repo, so a failure means a broken or partial clone).
• claudish — required: a validated {{CLAUDISH_SCRIPT}}; jq (the script exits without it); the ollama daemon answering on /api/tags (if not: brew services start ollama, or ollama serve); at least one model pulled — ollama list must be non-empty. If the script's default gemma3:4b is not among the pulled models, still install, but tell the user which model they do have and that they must set CLAUDISH_MODEL (shell profile, or the env block of ~/.claude/settings.json) or the command will fail with model isn't available. Optional but worth reporting: if they have an NVIDIA GPU with ≥24 GB VRAM, mention they can run a much larger model and should set CLAUDISH_KEEP_ALIVE=5m.

Step 5 — Install

For each command that passed its gate:

1. Read the template from <repo>/commands/<name>.md.
2. Replace every {{PLACEHOLDER}} with the gathered value (do the substitution in memory — do not use sed, paths contain slashes).
3. If ~/.claude/commands/<name>.md already exists, tell the user and ask before overwriting (offer to show a diff).
4. Write the result to ~/.claude/commands/<name>.md.
5. After writing, grep the installed file for {{ — any leftover token means a substitution was missed; fix it before declaring success.

For coldstart additionally install the transcript-search sidecar before substituting {{HISTORY_SCRIPT}}:

• mkdir -p ~/.claude/bin, then copy <repo>/coldstart-setup/history.py → ~/.claude/bin/history.py (verbatim — it takes all config via runtime args).
• Verify with python3 ~/.claude/bin/history.py index run from any git repo. Both a transcript table and no transcript history found for this project are passes; a traceback is not.
• {{HISTORY_SCRIPT}} is that absolute path with ~ expanded.

For pr-autoreview additionally copy the two workflow files (they take all config via runtime args, so they are copied verbatim):

• <repo>/workflows/pr-review-fanout.js → ~/.claude/workflows/pr-review-fanout.js
• <repo>/workflows/pr-draft-fanout.js → ~/.claude/workflows/pr-draft-fanout.js

Step 6 — Final report

End with a single summary the user can act on:

1. A table: command → installed path → status (ready / skipped: <missing requirement + fix> / declined).
2. Any warnings (e.g. acli missing, Kokoro deferred, Linux playback caveats).
3. Remind them: newly installed commands are picked up in a new Claude Code session — restart or open a new session, then try /onboard in any repo as a smoke test.
4. If pr-autoreview was installed, repeat the one-line warning that it posts reviews as them, and that scheduling it is a separate, deliberate step — never arm it automatically.

Do not install anything the user did not select, and do not leave half-written files behind: if a substitution fails mid-way, delete the partial file and report it.