AI Agent Skills — Custom Slash Commands for Claude Code, Codex & Gemini CLI

Define reusable SOPs once, invoke them across your entire agent fleet. Updated July 2026.

Why skills matter

Without skills, every new AI agent session starts blank. You have to re-explain your commit conventions, your PR review checklist, your deploy sequence — every single session, every single time. As your fleet grows from 2 sessions to 20, this becomes the bottleneck.

amux Skills solve this with a library of reusable SOPs. Each skill is a Markdown file describing exactly what the agent should do when you invoke it. Skills are:

• On-demand — they do nothing until you type /skill-name. Unlike CLAUDE.md, they don't add to every prompt's context.
• Fleet-wide — define once in the Skills tab, auto-synced to every running session.
• Model-agnostic — the same skill works on Claude Code, Codex, and Gemini CLI. amux handles provider-specific delivery.

The result: a consistent, auditable, version-controlled playbook for your AI engineering team that scales as you add more sessions.

Creating your first skill

Step 1: Open the Skills tab

In the amux dashboard (localhost:8822), click the Skills tab in the top navigation. It shows all defined skills and a New Skill button. The Skills tab is hideable — pin it to the taskbar or keep it in the overflow menu.

Step 2: Click New Skill and give it a name

The name becomes the slash command. Use lowercase with hyphens: review-pr creates /review-pr, commit creates /commit, run-tests creates /run-tests. Keep names short and action-oriented.

Step 3: Write the skill Markdown

The skill body describes the SOP. Be specific: list the exact steps, the expected output format, and any constraints. The more precise the skill, the more consistent the agent behavior across all sessions.

---
name: review-pr
description: Review staged changes for correctness, edge cases, and breaking changes
---

## When to use
Run this before committing a PR branch. Works on any staged diff.

## Steps
1. Run `git diff --cached` to see staged changes.
2. For each changed file: identify the intent of the change, list edge cases not covered by the diff, flag any breaking changes.
3. Check for: missing error handling, hardcoded values, test coverage gaps, security issues (SQL injection, XSS, auth bypass).
4. Output a structured review: LGTM / NEEDS WORK, then bullet points per finding with severity (critical/major/minor).
5. If NEEDS WORK: list the specific changes required before this should be committed.

## Output format
```
**Status:** LGTM / NEEDS WORK
**Summary:** one sentence

**Findings:**
- [critical/major/minor] file.ts:42 — description
```

Step 4: Save — skills sync automatically

Click Save. amux writes the skill to the skills/ directory and syncs it to all running sessions. For Claude Code sessions, the skill appears in the session's skills list immediately. For Codex and Gemini CLI sessions, amux injects it as a system block on the next invocation.

Step 5: Invoke the skill in any session

In the steering bar of any session, type /review-pr and hit Enter. The skill SOP is sent as the prompt. Works across all providers in your fleet.

Example skills

/standup — summarize yesterday's work

---
name: standup
description: Summarize what you accomplished since the last commit
---

Run `git log --since="24 hours ago" --oneline` and `git diff HEAD~5 --stat`.

Summarize in standup format:
- **Done:** bullet list of completed tasks (from commit messages and diff)
- **In progress:** what's staged or in your current working context
- **Blockers:** anything waiting on me or an external dependency

Keep it to 5 bullets max. Post to the amux board as an issue titled "Standup [date]".

/commit — stage and commit with a conventional message

---
name: commit
description: Stage all changes and commit with a conventional commit message
---

1. Run `git status` and `git diff` to understand what changed.
2. Group changes by type: feat, fix, refactor, docs, test, chore.
3. Stage only the relevant files (not lockfiles, .env, or generated files unless they're the point of the commit).
4. Write a commit message following Conventional Commits:
   - Format: `type(scope): short description`
   - Keep the subject under 72 chars
   - Add a body if the change is non-obvious
5. Run `git commit -m "..."`. Do not amend previous commits.
6. Report: what was staged, the commit hash, and the message.

/deploy — test, build, push

---
name: deploy
description: Run tests, build, and push to origin
---

1. Run the test suite. If any tests fail, stop and report which tests failed and why.
2. Run the build command. If the build fails, stop and report the error.
3. If both pass: `git push origin HEAD`. Report the push result.
4. After push: check if there's a CI pipeline. If so, note the pipeline URL.

Do NOT push if tests fail. Do NOT force push. Do NOT skip the build step.

Skills vs alternatives

Model-agnostic delivery (July 2026)

The July 2026 Skills update (v0.9.x) made the Skills tab model-agnostic and cross-provider. Previously, skills only worked in Claude Code sessions (via the native /slash-command system). Now amux handles provider-specific delivery for all runtimes:

• Claude Code: skill Markdown is written to the session's skills/ directory. Invocation uses Claude Code's native /skill-name syntax.
• OpenAI Codex CLI: skill body is injected as a system context block prepended to the steering prompt on invocation.
• Gemini CLI: same injection approach as Codex — the SOP becomes the leading context of the delivered prompt.

From the user's perspective, /review-pr works the same in every session, regardless of which AI is running behind it. You can have a Claude Code session and a Codex session both using the same /commit skill with no changes.

Keeping skills in sync across a large fleet

When you update a skill in the dashboard, amux marks it dirty and syncs it on the next prompt delivery to each session. For active sessions this is nearly instant. For idle or detached sessions, the sync happens the next time they receive a prompt (whether from the steering bar, the scheduler, or another session via inter-session messaging).

To force-sync immediately: click Sync All in the Skills tab, or use the API:

curl -sk -X POST https://localhost:8822/api/skills/sync

Frequently Asked Questions

git clone https://github.com/mixpeek/amux && cd amux && ./install.sh
amux register myproject --dir ~/Dev/myproject --yolo
amux start myproject
# Open https://localhost:8822 → Skills tab → New Skill

Related guides

• Install amux and run your first parallel agents
• Schedule AI agents to run automatically on a cron
• Agent-to-agent coordination with amux
• Claude Code, Codex, and Gemini CLI together in one fleet
• AI agent daily standup — /standup in action
• Running 10+ AI agents in parallel