init

/talksmith:init

One-shot bootstrap, run once per working directory. It writes the Talksmith working-directory stub as CLAUDE.md in the current directory, writes a thin AGENTS.md pointer so Codex (and other agents that read AGENTS.md) start the same workflow, and adds a .gitignore covering Talksmith's regenerable build artifacts and caches (so they never get committed to the shared subject repo). On the next session, that stub loads the full operating spec and Talksmith introduces itself and starts; this command does not replicate that, it just puts the stubs in place.

CLAUDE.md is the single source of truth; AGENTS.md only points at it. The real boot instructions live in CLAUDE.md; AGENTS.md is a short stub that tells a non-Claude-Code agent to open CLAUDE.md and follow it (with the two Claude-Code-specific fallbacks — the @-import and ${CLAUDE_PLUGIN_ROOT} — spelled out). Nothing is duplicated between them, so they can't drift.

The stub is intentionally minimal and stable — it only knows how to load the real spec from ${CLAUDE_PLUGIN_ROOT}/orchestrator.md. The spec itself is not copied; it lives in the plugin install and refreshes automatically on /plugin update, so /talksmith:init rarely needs re-running.

This command does not scaffold config/ files, the talks/ directory, or the profile — those are the orchestrator's job (Editor subagent, Step 0.5 and Step 1) once the workflow is running.

What to do

1. Resolve the destination as ./CLAUDE.md (cwd-relative — the user's project root).
2. Refuse to run inside the plugin source repo. If ./.claude-plugin/plugin.json exists in the cwd, stop immediately and emit:

   [init] Refusing to run: this directory is the Talksmith plugin source repo (./.claude-plugin/plugin.json exists). /talksmith:init is for user subject working directories, not the plugin source. The dev CLAUDE.md here is the plugin development notes — overwriting it would clobber documentation. cd into a separate working directory and re-run.
   

   Do not write anything. This guard exists because /talksmith:init always overwrites — without it, running this command in the plugin repo would destroy the dev CLAUDE.md.
3. Write both stubs, byte-for-byte, always overwriting any existing ./CLAUDE.md and ./AGENTS.md:

   cp -f "${CLAUDE_PLUGIN_ROOT}/talksmith-orch.md"   ./CLAUDE.md
   cp -f "${CLAUDE_PLUGIN_ROOT}/talksmith-agents.md" ./AGENTS.md
   

   If ${CLAUDE_PLUGIN_ROOT} is unset or those paths fail, locate the plugin install (find talksmith-orch.md / talksmith-agents.md under the Claude Code plugins directory) and copy both files. Re-running /talksmith:init is the supported way to pick up stub changes; user content belongs in config/, talks/, never in CLAUDE.md or AGENTS.md. AGENTS.md is a pointer to CLAUDE.md — it carries no independent content, so keeping it in sync is automatic. Emit:

   [init] CLAUDE.md written from the Talksmith stub (talksmith-orch.md), and AGENTS.md written as the Codex pointer to it. Any prior copies were overwritten.
   

4. Add a .gitignore for build artifacts. Talksmith regenerates decks, icon caches, and critique previews on every render — those must never be committed to the shared subject repo. Append the Talksmith ignore block to ./.gitignore (creating the file if it doesn't exist), guarded by a marker so it's idempotent and any pre-existing .gitignore is preserved — never overwrite the user's own entries:

   # Idempotent: append only if the marker isn't already present; existing entries untouched.
   # printf (quoted args) so the written lines never inherit stray leading whitespace.
   grep -qF "# --- Talksmith ---" ./.gitignore 2>/dev/null || printf '%s\n' \
     '' \
     '# --- Talksmith --- (regenerable build artifacts & caches — keep out of Git)' \
     '.DS_Store' '.venv/' '__pycache__/' '*.pyc' \
     'talks/*/output/' 'talks/*/images/.critique/' '.icons/' \
     '/index.html' '/talksmith-index.html' \
     '.claude/settings.local.json' \
     >> ./.gitignore
   

   The root index.html is the generated landing page every HTML render rewrites (it links into the ignored output/ decks, so it is regenerable too). The source a Talk is built from — draft.md, final.md, memory.md, research/corpus/, images/ (the committed SVG/PNG diagrams), and config/ — stays tracked; only the regenerable output/, caches, and local settings are ignored. Emit:

   [init] .gitignore updated with Talksmith build-artifact ignores (any existing entries preserved).
   

5. Finish by telling the user to restart:

   Done. Close this session and open a new one in this directory to start creating —
   in Claude Code or Codex, the stub loads Talksmith, which introduces itself and
   walks you through Step 0 → Step 0.5 (profile) → Step 1 (Frame your first Talk).
   
   Everything else — config/profile.md, config/learnings.md, the feedback logs,
   and the talks/ tree — is created by the orchestrator on demand. Nothing to scaffold by hand.