ai-devkit
Teach your AI coding agent a proven dev loop. ai-devkit is a toolbox you install into an AI coding agent — Claude Code or Codex. One install teaches the agent a battle-tested daily loop: turn a sentence into a GitHub issue, an issue into a planned branch, messy changes into clean commits, and a finished branch into a PR — in any repo you work on, in any language.
It is two things at once:
- An asset repo — versioned, reusable agent skills, discoverable and adaptable to any stack.
- A knowledge portal — alexandremorgado.github.io/ai-devkit, generated from the assets in this repo so it never drifts from source.
New to AI coding agents? An agent is a program that runs in your terminal: you type what you want in plain English, and it reads your code, edits files, and runs commands — showing its work and asking before anything risky. The portal's Start here guide assumes zero AI experience.
Quickstart
1. Get an agent
| Agent | Install | Run |
|---|
| Claude Code (Anthropic) | npm install -g @anthropic-ai/claude-code | claude inside a repo — docs |
| Codex (OpenAI) | npm install -g @openai/codex | codex inside a repo — docs |
2. Install ai-devkit (two commands)
Claude Code — inside a session, type:
/plugin marketplace add alexandremorgado/ai-devkit
/plugin install ai-devkit@ai-devkit
The skills are available immediately, in every repo. Prefer a single skill instead of the whole plugin? Drop its SKILL.md (with any references/) into ~/.claude/skills/<name>/.
Codex — in your shell:
codex plugin marketplace add alexandremorgado/ai-devkit
codex plugin add ai-devkit@ai-devkit
Start a new Codex thread afterwards so the skills load.
Updating — plugins don't auto-pull new commits, so refresh when the repo ships changes:
- Claude Code:
/plugin marketplace update ai-devkit, then re-run /plugin install ai-devkit@ai-devkit, then /reload-plugins (or start a new conversation).
- Codex:
codex plugin marketplace upgrade, then codex plugin add ai-devkit@ai-devkit (then start a new Codex thread).
3. Use it — the daily loop
A skill is invoked by typing /its-name in Claude Code, or $its-name in Codex, plus plain words — no other syntax to learn. The skills chain into one loop, from "someone found a bug" to "PR opened" (the table shows the Claude Code form):
| Moment in your day | You type | What happens |
|---|
| Someone reports a bug, or you have an idea | /create-issue …one sentence… | Well-formed issue: type, labels, repro steps, duplicate check |
| You pick up an issue | /issue-to-branch #482 | Branch (or worktree) + a development plan built from your repo |
| Starting without an issue | /create-branch …a sentence… | A well-named branch from your changes or a short description |
| Your branch fell behind main | /update-from-branch main | Merge/rebase from main, auto-stashing dirty work, conflicts surfaced |
| The working tree is messy | /smart-commit | 2–5 atomic commits, plan shown first, never pushes |
| Before you push | /ensure-tests | Decides what needs tests, runs the suite, fixes failures to 100% |
| The work feels done | /finish-branch | Readiness checks, plan archived, PR opened or updated |
| Anytime, before review | /cleanup --branch | Finds debug prints, leftover comments, commented-out code |
| Stuck on something hard | /deepthink …the problem… | Structured extended reasoning → an implementation strategy |
| A bug won't reproduce or won't die | /ultrafix …the symptom… | Isolated worktrees + structured logging → root cause + verified fix |
| Want a second agent's take | /codex-buddy review this branch | Codex reviews or debugs independently; Claude cross-checks the findings |
| The plan drifted from reality | /update-branch-plan | Conservatively syncs plan checkboxes with your commits |
The core idea: adapt, don't install
A skill is a workflow + conventions, not a binary. A skill written for one stack rarely runs as-is on another — but its intent and structure transfer. So instead of copying, you read it and have your agent regenerate an adapted version for your stack.
Every skill in the catalog is tagged for portability and ships an "Adapt to your platform" prompt you paste into Claude or Codex:
- Portable — transfers directly with config changes (e.g.
gh/git workflows like create-issue).
- Adaptable — concept transfers, implementation differs (e.g. testing/review/debugging workflows).
- Platform-specific — instructive, but tied to one stack (read it for the pattern).
What the plugin installs
Exactly the contents of skills/ — twelve SKILL.md playbooks, no hooks, agents, or background processes:
