commit 2fac6d2e74e75f5038824b595fd73326435b6209
Author: gm <nitro820@gmail.com>
Date:   Fri Jun 5 10:31:05 2026 +0900

    Initial agent skills setup
    
    Common skills shared across Claude Code, Codex, and Antigravity,
    plus per-agent extras. sync.sh deploys to each tool's config path.
    
    Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

diff --git a/agents_extra.md b/agents_extra.md
new file mode 100644
index 0000000..3b1587a
--- /dev/null
+++ b/agents_extra.md
@@ -0,0 +1,30 @@
+# Codex & Antigravity — Agent-Specific Skills
+
+## Skill Discovery
+
+- Skills in `~/.agents/skills/[name]/SKILL.md` are auto-discovered.
+- Each skill directory can include optional `scripts/`, `references/`, and `assets/` subdirectories.
+- Skills can be scoped: user (`~/.agents/skills`), repo (`.agents/skills`), or system (`/etc/codex/skills`).
+
+## SKILL.md Format
+
+- Always include YAML frontmatter with `name` and `description` fields.
+- `description` should clearly state when and how the skill applies.
+- Example:
+  ```yaml
+  ---
+  name: my-skills
+  description: Personal coding conventions and preferences
+  ---
+  ```
+
+## Task Execution
+
+- Prefer running tasks in isolated environments when available.
+- Use `agents/openai.yaml` to configure agent-specific behavior per skill directory.
+- Leverage built-in Codex skills before writing custom ones.
+
+## Project Awareness
+
+- Read `.agents/skills` at repo root for project-specific overrides.
+- Project-level skills take precedence over user-level skills.
diff --git a/claude_extra.md b/claude_extra.md
new file mode 100644
index 0000000..72f185e
--- /dev/null
+++ b/claude_extra.md
@@ -0,0 +1,30 @@
+# Claude Code — Agent-Specific Skills
+
+## CLAUDE.md & Project Context
+
+- Always check for `CLAUDE.md` at project root before starting tasks.
+- Instructions in `CLAUDE.md` override default behavior — follow them exactly.
+- Use `@path/to/file` syntax to import additional context files inline.
+
+## Tool Use
+
+- Prefer dedicated tools (Read, Edit, Write) over Bash when one fits.
+- Run independent tool calls in parallel to maximize efficiency.
+- Reserve Bash for shell-only operations.
+
+## Memory
+
+- Save non-obvious user preferences and feedback to `/root/.claude/projects/` memory files.
+- Do not save things derivable from code, git history, or documentation.
+- Verify memory against current file state before acting on it — memories can go stale.
+
+## MCP & Hooks
+
+- MCP servers extend tool capabilities — check available MCP tools before falling back to generic approaches.
+- Hooks in `settings.json` execute automatically on events; use them for automated behaviors.
+- Never skip hooks (`--no-verify`) unless the user explicitly requests it.
+
+## Permissions
+
+- For destructive or hard-to-reverse actions (force push, rm -rf, drop table), confirm with user first.
+- One-time approval does not mean blanket approval — re-confirm in new contexts.
diff --git a/common.md b/common.md
new file mode 100644
index 0000000..114be5b
--- /dev/null
+++ b/common.md
@@ -0,0 +1,33 @@
+# Common Agent Skills
+
+## Code Style
+
+- Write clean, minimal code. No unnecessary comments unless the *why* is non-obvious.
+- Prefer editing existing files over creating new ones.
+- Do not add error handling or validation for scenarios that cannot happen.
+- Three similar lines is better than a premature abstraction.
+- No half-finished implementations.
+
+## Task Approach
+
+- For exploratory questions, respond with a recommendation and the main tradeoff in 2-3 sentences before implementing.
+- Do not add features, refactor, or introduce abstractions beyond what the task requires.
+- Default to the simplest solution that satisfies the requirement.
+
+## Security
+
+- Never introduce command injection, XSS, SQL injection, or other OWASP top 10 vulnerabilities.
+- Only validate at system boundaries (user input, external APIs). Trust internal code and framework guarantees.
+- Do not hardcode secrets or credentials in code.
+
+## Git
+
+- Write concise commit messages focused on *why*, not *what*.
+- Prefer creating new commits over amending published ones.
+- Never force-push to main/master without explicit confirmation.
+
+## Communication
+
+- Be concise. Match response length to task complexity.
+- State results and decisions directly.
+- When referencing code, include file path and line number when possible.
diff --git a/sync.sh b/sync.sh
new file mode 100755
index 0000000..ed1a0c9
--- /dev/null
+++ b/sync.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+set -e
+
+SKILLS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# Claude Code
+mkdir -p ~/.claude
+cat "$SKILLS_DIR/common.md" "$SKILLS_DIR/claude_extra.md" > ~/.claude/CLAUDE.md
+echo "  [ok] ~/.claude/CLAUDE.md"
+
+# Codex + Antigravity (shared ~/.agents/skills)
+SKILL_DIR=~/.agents/skills/my-skills
+mkdir -p "$SKILL_DIR"
+{
+  printf -- "---\nname: my-skills\ndescription: Personal coding skills and conventions\n---\n\n"
+  cat "$SKILLS_DIR/common.md"
+  echo ""
+  cat "$SKILLS_DIR/agents_extra.md"
+} > "$SKILL_DIR/SKILL.md"
+echo "  [ok] $SKILL_DIR/SKILL.md"
+
+echo "synced."