Start here

This page assumes zero experience with AI tools. If you can use a terminal, you'll be productive in about ten minutes — it walks you from "what even is an AI coding agent?" to running these workflows in your own repo. Already set up? Jump to step 3 or browse the Catalog tab above.

1. What is an AI coding agent?#

An AI coding agent is a program that runs in your terminal, inside any repo. You type what you want in plain English; the agent reads your code, edits files, and runs commands to get it done — showing you everything it does and asking before anything risky. Think of it as a fast pair programmer that never gets tired of the boring parts.

A session looks like this:

$ cd your-repo$ claude> explain what src/sync/ does⏺ Reading src/sync/ (6 files)…  It syncs local edits with the backend: SyncQueue batches changes, retries on failure…> rename the SyncManager class to SyncCoordinator everywhere⏺ Found 23 references across 9 files✓ Renamed — review the diff with git diff

ai-devkit works with two popular agents. Both work the same way, and everything on this site works with both:

Both need an account or license to sign in on first run.

2. What is ai-devkit, and why install it?#

Out of the box, an agent is a capable generalist — but it doesn't know your conventions: how you like to write issues, name branches, group commits, or decide when a branch is ready for a PR.

ai-devkit is a plugin: a bundle of skills that teaches the agent a proven daily loop. A skill is a written playbook (a markdown file) the agent follows step by step. After one install, in every repo you work on, you can type things like:

• /create-issue the favorites list flashes when filtering by date — and get a well-formed GitHub issue with the right labels, repro steps, and a duplicate check.
• /smart-commit — and your messy working tree becomes 2–5 clean, atomic commits (you approve the plan first; it never pushes).
• /finish-branch — and the agent verifies the branch is actually done, runs the tests, and opens the PR.

(That's the Claude Code form — in Codex the same skills are invoked with a $, e.g. $create-issue, or by simply naming them in a sentence.)

The skills are versioned in the alexandremorgado/ai-devkit repo — this site is generated from that repo, so what you read here is exactly what installs.

3. Install ai-devkit (two commands)#

Claude Code#

Run claude anywhere, then type these two commands at its prompt:

/plugin marketplace add alexandremorgado/ai-devkit
/plugin install ai-devkit@ai-devkit

That's it — the skills are available immediately, in every repo. (The first command points Claude Code at the repo, the second installs the bundle.)

Codex#

In your shell (not inside a session):

codex plugin marketplace add alexandremorgado/ai-devkit
codex plugin add ai-devkit@ai-devkit

Then start a new Codex session so the freshly installed skills load.

Check it worked#

In Claude Code, type / — create-issue, smart-commit, and the rest should appear in the command list. In Codex, just ask: "what skills do you have from ai-devkit?"

Keeping it updated#

Plugins don't auto-pull new commits, so refresh when the repo ships changes:

Claude Code

/plugin marketplace update ai-devkit
/plugin install ai-devkit@ai-devkit
/reload-plugins                       # or just start a new conversation

Codex — in your shell, then start a new Codex session:

codex plugin marketplace upgrade
codex plugin add ai-devkit@ai-devkit

4. Your first five minutes#

Try these in a repo you actually work on. A skill is invoked by typing /its-name in Claude Code, or $its-name in Codex, followed by your input in plain words — there is no other syntax to learn. (In Codex you can also just name the skill in a sentence.) The examples below use the Claude Code form.

Turn one sentence into a GitHub issue:

/create-issue the favorites list flashes when you filter by date

The agent checks for duplicates, infers the type and labels from your repo's own label set, writes repro steps and acceptance criteria, and shows you the draft before creating anything.

Turn an issue into a planned branch:

/issue-to-branch #482

You get a branch named by your conventions plus a branch plan — a markdown checklist of the work, which the other skills keep updated as you go.

Clean up a messy working tree:

/smart-commit

The agent groups your uncommitted changes into 2–5 atomic commits with semantic prefixes (Fix:, Feat:, …), shows you the plan, and only commits after you approve. It never pushes.

5. The daily loop#

The skills chain into one loop — from "someone found a bug" to "PR opened". Use any step on its own; they also hand off to each other automatically.

The "You type" column shows the Claude Code form — in Codex, type $the-name instead (e.g. $smart-commit), or just name the skill in plain words.

6. Everything the plugin installs#

Exactly these skills — nothing else, no hooks, agents, or background processes:

• cleanup — Detect and optionally remove code noise — debug prints, transitional/obvious comments, and commented-out code — across languages. Pairs with smart-commit. Try: /cleanup --branch
• codex-buddy — Use a second AI agent (Codex) for review, a second opinion, debugging, an audit, or delegated implementation — context-enriched, run non-interactively, and cross-checked rather than trusted blindly. Any repo. Try: /codex-buddy review my changes on this branch
• create-branch — Create a git branch from current changes or a short description — infer a type/kebab-name, confirm, then git switch -c (or git worktree add for isolation). Any repo. Try: /create-branch add rate limiting to the api client
• create-issue — Create a well-structured GitHub issue with AI-inferred type, labels, and duplicate detection via gh — on any repo. Try: /create-issue the favorites list flashes when you filter by date
• deepthink — A structured extended-reasoning framework — decompose, explore options, analyze deeply, apply patterns, assess risk, and produce an implementation strategy — for any stack. Try: /deepthink how should we cache images app-wide without unbounded memory growth?
• ensure-tests — Analyze a branch's scope, run the test suite, fix failures to 100% pass, then conservatively annotate the branch plan's test section — on any stack. Try: /ensure-tests
• finish-branch — Validate a feature branch's readiness (plan %, tests, hygiene), finalize and archive its plan, then create or update a PR against the base branch — on any GitHub repo. Try: /finish-branch
• issue-to-branch — Turn a GitHub issue (or a text description) into a branch or isolated worktree with an AI-generated development plan — on any repo. Try: /issue-to-branch #482
• smart-commit — Group uncommitted changes into 2–5 atomic, semantically-prefixed commits by module, present a plan, and commit on approval — never auto-pushes. Any language. Try: /smart-commit
• ultrafix — Debug a stubborn bug with isolated git worktrees for parallel hypothesis testing plus structured logging — reproduce, isolate, find root cause, verify a minimal fix, clean up. Any repo. Try: /ultrafix tests flake on ci but pass locally
• update-branch-plan — Analyze recent commits and conservatively update a branch plan's checkboxes and phase status — marking tasks complete only at ≥80% confidence. Any repo. Try: /update-branch-plan
• update-from-branch — Sync the current branch from another (usually main) via merge or rebase, auto-stashing dirty work and surfacing conflicts clearly — never discards changes. Any repo. Try: /update-from-branch main

7. Words you'll see around here#

8. When a skill doesn't fit your stack: adapt it#

A skill is a playbook, not a binary — so you never port one by hand. Open any skill in the Catalog, copy its Adapt to your platform prompt, fill in your stack and conventions, and paste it into Claude or Codex. The agent reads the skill and produces an adapted version (markdown plus any scripts) for your project. If you ship that adaptation, consider contributing it back as a case so others see the worked example.

9. No agent? Read the skills directly#

A skill is just a markdown file — you don't need an agent to benefit. Clone the repo and read any skills/<name>/SKILL.md: it's a step-by-step playbook you can follow by hand, or paste into any LLM.

gh repo clone alexandremorgado/ai-devkit
cd ai-devkit

This site is generated from those same files (npm run build writes dist/), so the catalog never drifts from the source.

10. Questions, ideas, contributions#

Found a rough edge in a skill, or built a workflow worth sharing? A skill is just a markdown file — contributing one is a single PR. See the Contribute tab above, or open an issue on alexandremorgado/ai-devkit.