AI Automation Skills: 9 Skills I Run Daily

AI automation skills are reusable instructions that turn Claude Code or Codex from a coding assistant into a daily operating layer. Instead of writing a fresh prompt every morning, you install a small library of skills that know how to prioritize, scan, draft, publish, and reflect.

I put together a public repo with 9 practical automation skills you can install and adapt. The set is built around the loop I care about most: decide what matters, clear the noisy inputs, ship useful updates, and review where the day actually went.

The GitHub Repo

Install the AI automation skills from https://github.com/travisseh/ai-automation-skills.

Steps:
1. Clone the repo to a temporary folder.
2. Copy every folder under skills/ into my local skills directory:
   - Claude Code: ~/.claude/skills
   - Codex: ~/.codex/skills
3. Preserve each skill folder exactly, including SKILL.md, scripts/, agents/, and other supporting files.
4. Do not overwrite an existing skill without first showing me the conflicting folder names and asking for confirmation.
5. After copying, list the installed skill names and tell me to restart the agent app if needed.

Then help me pick one skill to run first based on my current workflow.

Paste that prompt into Claude Code or Codex. It tells the agent to clone the repo, copy the skills into the right local folder, avoid overwriting existing skills without permission, and report what it installed.

What You Get

The repo includes 9 skills. They are intentionally small. Each one teaches the agent a specific workflow instead of trying to become one giant personal operating system.

Why Skills Beat One Long Prompt

A long prompt works once. A skill becomes part of the agent's operating memory. When you ask for a daily prioritization pass, inbox cleanup, or release-note draft, Claude Code can pull the relevant workflow automatically and follow the same procedure every time.

That consistency matters more than clever wording. The difference between a toy automation and a useful one is usually guardrails: what sources to check, what to ignore, what to ask before writing, and what evidence to show before taking action.

If you're new to skills, start with my Claude Code skills guide. If you're still getting set up, use the Claude Code setup guide first.

The Daily Automation Loop

The skills are grouped around 4 jobs.

• Prioritize: decide what matters today and extract promises you already made.
• Automate: clear email, scan Reddit, summarize podcasts, and review spending.
• Publish: prepare LinkedIn work and release notes.
• Reflect: build a daily chronicle so you can see how your time actually moved.

The important part is that these are connected. The prioritizer should notice yesterday's promises. The chronicle should show which automations ran. The publishing skills should turn work you already did into useful updates.

Run Your First Skill

After installation, restart Claude Code or Codex. Then run one narrow request. Do not try to wire every automation on day one.

Use $daily-prioritizer.

Look at my available context for today and give me the 3 most important things to focus on.
For every source you check, mark it as checked, unavailable, auth failed, empty, or partial.
Do not invent context from memory if a source is unavailable.

That prompt tests the highest-value part of the system: can the agent gather real context, separate important from noisy, and show source coverage instead of guessing?

Turn It Into A Real Automation

Once a skill works manually, schedule it. For Claude Code, that usually means a cron, launchd job, or plugin command that runs the same prompt on a cadence. I cover the broader automation pattern in the first automation guide.

Use $what-i-promised.

Look at yesterday across my available communication sources.
Extract only first-person commitments I made.
Separate clear commitments from lower-confidence ones.
Ask me to route each item as done, schedule, codex, backlog, or drop before creating events or closing tasks.

This is the pattern I recommend: run manually, tighten the skill, then schedule it. If it cannot behave reliably when supervised, it will be worse when unattended.

Customize Before You Trust It

These skills are sanitized templates. They cannot know your actual Gmail account nicknames, Notion database names, Slack workspaces, timezone, local scripts, or calendar setup until you adapt them.

Open each SKILL.md and replace generic placeholders with your real tool paths and account names. Keep secrets out of the skill files. Store tokens in local config files, environment variables, or the auth system your CLI already uses.

My rule: the skill should describe the workflow, not contain private credentials. The agent can use credentials that are already configured locally, but the reusable instruction file should stay shareable.

What To Change First

Start with the source list. If you do not use Slack, remove Slack from the prioritizer and promise extractor. If you use Linear instead of Notion, change the task database instructions. If you live in a different timezone, fix the date-window language before you schedule anything.

Then run a dry pass and look for bad behavior. Does it skip sources silently? Does it confuse ideas with commitments? Does it draft emails when you only asked it to identify replies? Those are skill problems. Tighten the instructions until the agent behaves predictably.

The Course Version

In ClaudeFluent, we build this kind of system live. Students do not just watch a demo. They install Claude Code, write skills, connect tools, and leave with automations that fit their own work.

If you want the repo, start with the install prompt above. If you want to understand how to build your own version from scratch, join the next ClaudeFluent cohort.