setup-multi-agent

name	setup-multi-agent
description	Bootstrap a new project with multi-CLI agent config — drops AGENTS.md / CLAUDE.md / GEMINI.md from the cc-templates templates/ directory into a target project, patches its .gitignore, and prints verification commands. The "multi-agent" here means multi-CLI (Claude Code + Codex CLI + Gemini CLI), not multiple internal sub-agents. Use when this capability is needed.
metadata	{"author":"FuzzyKala"}

/setup-multi-agent — Bootstrap a new project

Drop the AGENTS.md / CLAUDE.md / GEMINI.md trio into a target project so all three coding CLIs (Claude Code, Codex CLI, Gemini CLI) load the same canonical project context.

Prerequisites

The cc-templates repo must be available locally. The skill reads four files from its templates/ directory:

• templates/AGENTS.md.template
• templates/CLAUDE.md.template
• templates/GEMINI.md.template
• templates/gitignore-additions.txt

Step 0 below locates that directory. If any of the four files is missing after resolution, fail loudly and stop.

Steps

0. Locate the templates directory

This skill is usually invoked from a symlinked install at ~/.claude/skills/setup-multi-agent/ that points back into the cct repo, so templates/ is NOT a sibling of SKILL.md. Resolve the real cct repo root by following the skill's own symlink and walking up three levels.

The harness reports the skill's base directory at invocation (e.g., Base directory for this skill: /home/<user>/.claude/skills/setup-multi-agent). Pass that path through readlink -f and ascend three directories — the skill lives at <cct-repo>/.claude/skills/setup-multi-agent/, so up-three lands at <cct-repo>:

# Replace <skill-base-dir> with the path the harness reported at invocation.
SKILL_DIR_RESOLVED="$(readlink -f "<skill-base-dir>")"
CCT_ROOT="$(readlink -f "${SKILL_DIR_RESOLVED}/../../..")"
TEMPLATES_DIR="${CCT_ROOT}/templates"

# Verify the four template files exist; fail loudly if any are missing.
missing=0
for f in AGENTS.md.template CLAUDE.md.template GEMINI.md.template gitignore-additions.txt; do
  if [ ! -r "${TEMPLATES_DIR}/${f}" ]; then
    echo "ERROR: Missing template at ${TEMPLATES_DIR}/${f}" >&2
    missing=1
  fi
done
[ "${missing}" -eq 0 ] || exit 1

echo "Templates dir resolved: ${TEMPLATES_DIR}"

Use the resolved ${TEMPLATES_DIR} (or the absolute path the snippet prints) in all subsequent steps. This works whether the skill is invoked from a symlinked install or directly from inside the cct repo; the readlink -f step is what lets symlinked installs resolve to the real cct repo root instead of ~/.claude/skills/.

1. Collect inputs

Ask the user (one batched question with AskUserQuestion if available, else sequentially):

• Project name (e.g., my-landing-page). Used to replace {{PROJECT_NAME}} in the template.
• One-line project description (replaces {{PROJECT_DESCRIPTION}}).
• Target market or audience (replaces {{TARGET_MARKET}}). If unsure, accept "TBD" and let the user fill in later.
• Tech stack summary (replaces {{TECH_STACK_SUMMARY}}). Single line. E.g., "Next.js 15 + TypeScript + Tailwind + Vercel".
• Key constraints (replaces {{KEY_CONSTRAINTS}}). E.g., "Solo dev, ship by Q3, no backend".
• Target directory (absolute path). Must be an existing directory or one the user wants the skill to mkdir -p.

2. Sanity-check the target

• If target does not exist, ask before creating it.
• If target is not a git repo (git -C <target> rev-parse --is-inside-work-tree), warn and ask whether to git init it.
• If <target>/AGENTS.md, <target>/CLAUDE.md, or <target>/GEMINI.md already exists, list which ones and ask: overwrite, skip, or abort.

3. Materialise the files

For each template, read from ${TEMPLATES_DIR}/ (resolved in Step 0), substitute placeholders, and write to <target>/:

• ${TEMPLATES_DIR}/AGENTS.md.template → <target>/AGENTS.md
  • Replace: {{PROJECT_NAME}}, {{PROJECT_DESCRIPTION}}, {{TARGET_MARKET}}, {{TECH_STACK_SUMMARY}}, {{KEY_CONSTRAINTS}}.
• ${TEMPLATES_DIR}/CLAUDE.md.template → <target>/CLAUDE.md
  • Replace: {{TODAY}} with date +%Y-%m-%d output.
• ${TEMPLATES_DIR}/GEMINI.md.template → <target>/GEMINI.md
  • No placeholders typically (single @./AGENTS.md line).

4. Patch .gitignore

• If <target>/.gitignore does not exist, create it with the contents of ${TEMPLATES_DIR}/gitignore-additions.txt.
• If it exists, append ${TEMPLATES_DIR}/gitignore-additions.txt to it — but first grep for .gemini/ to avoid double-appending if the user runs the skill twice.

5. Print verification commands

Print these so the user can run them in a fresh terminal:

cd <target>

# Claude Code loads AGENTS.md via the @AGENTS.md import in CLAUDE.md:
claude --print 'Without using any tools, just from your loaded context: does it contain a section called "Tool Preference: CLI over MCP"? If yes quote the first sentence.'

# Gemini CLI loads AGENTS.md via the @./AGENTS.md import in GEMINI.md:
# GEMINI_CLI_TRUST_WORKSPACE=true is required for Gemini CLI headless mode (-p flag):
GEMINI_CLI_TRUST_WORKSPACE=true gemini -p 'Without using any tools, just from your loaded context: does the project context exist? Answer yes or no.'

# Codex CLI reads AGENTS.md natively (no wrapper needed). NOTE: Codex is an
# agent loop and will read files unless explicitly told not to. The prompt
# below asks for a verbatim quote from loaded context; if Codex answers
# without running shell commands, native auto-load is confirmed. If it ran
# `sed`/`cat`/`grep` on AGENTS.md first, that only confirms AGENTS.md exists
# and Codex can find it — NOT that it was auto-loaded into context:
codex exec 'Without running any shell commands or reading any files, answer purely from your loaded AGENTS.md context: does it contain a section titled "Tool Preference: CLI over MCP"? If yes quote the first sentence.'

If any of the three CLIs is not installed, mention which and link the user to:

• Claude Code: https://docs.claude.com/en/docs/claude-code
• Codex CLI: https://developers.openai.com/codex
• Gemini CLI: https://github.com/google-gemini/gemini-cli

6. Final summary

Print a concise summary:

✓ Bootstrapped <project-name> at <target>
  - AGENTS.md (canonical project context)
  - CLAUDE.md (thin @AGENTS.md wrapper + Current Sprint Status, Session 1)
  - GEMINI.md (thin @./AGENTS.md wrapper)
  - .gitignore patched (+ .gemini/, +.agents/, +.claude/*)

Next steps:
  1. Edit AGENTS.md to fill in any details the placeholders missed.
  2. Run the 3 verification commands above (fresh terminal each, in <target>).
  3. Commit: cd <target>; git add AGENTS.md CLAUDE.md GEMINI.md .gitignore; git commit -m 'feat: bootstrap multi-CLI agent config via cc-templates'

What NOT to do

• Do not install v2's 5 specialist agents. That pattern is gone in v3.
• Do not copy the .claude/skills/ directory into the new project. Skills are repo-local to cc-templates; if the new project wants /wrap etc, the user clones or symlinks separately. (Future work: a skills-bundle template.)
• Do not run the verification commands yourself in this same session — the user runs them in a fresh terminal so context-load is honest.

---

Source: FuzzyKala/cc-templates — distributed by TomeVault.