Teach your AI coding agent proven dev workflows
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.
12 skills · install once, use in every repo
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. If you can use a terminal, you can use everything here. The Start-here guide assumes zero AI experience →
Set up in three steps#
Get an agent
Claude Code or Codex — both run in your terminal, and everything here works with both.
npm install -g @anthropic-ai/claude-code npm install -g @openai/codex
Install ai-devkit
Inside a Claude Code session, type:
/plugin marketplace add alexandremorgado/ai-devkit /plugin install ai-devkit@ai-devkit
In Codex it’s the same two commands, prefixed codex plugin in your shell.
Use it — in any repo
The workflows now work everywhere you code. Try one:
/create-issue the favorites list flashes when filtering
In Codex the same skill is $create-issue. No agent at all? The skills are just markdown — read or paste the playbooks directly.
What you get: the daily loop#
These workflows chain into one loop — from “someone found a bug” to “PR opened”. Each is a skill: a written playbook the agent follows step by step, using your repo’s own labels, branches, and test commands. Click one for what it does, a copy-paste example, and the full playbook.
/create-issue
Describe a bug or idea in one sentence — get a well-formed GitHub issue with labels, repro steps, and duplicate detection.
/issue-to-branch
Turn an issue into a branch (or isolated worktree) with a development plan built from your repo.
/smart-commit
Group a messy working tree into 2–5 clean, atomic commits. Shows you the plan first; never pushes.
/ensure-tests
Decide what needs tests, run the suite, and fix failures until everything passes.
/finish-branch
Check the branch is really done, finalize its plan, and open (or update) the PR.
Plus, at any moment
/create-branch
Start work without an issue — turn your changes or a one-line description into a named branch.
/cleanup
Sweep out debug prints, leftover comments, and commented-out code before they ship.
/update-branch-plan
Keep the branch plan honest — sync its checkboxes with what you actually committed.
/deepthink
Structured extended reasoning for hard problems — decompose, weigh options, produce a strategy.
/ultrafix
Hunt a stubborn bug with isolated worktrees and structured debug logging — evidence before fix.
/codex-buddy
Bring in a second agent (Codex) for an independent review, a second opinion, or a deeper debugging pass — then cross-check what it finds.
Repetitive work is exactly what a skill is for#
A skill is a written playbook your agent follows step by step. The payoff is biggest on the rituals you repeat all day — the multi-step dances you do the same way every time, where forgetting step 3 quietly costs you twenty minutes. Write the dance down once, and run it with a single command — in any repo, forever.
- Consistent every time. The same steps run the same way — nothing skipped because it’s late and you’re tired.
- Your conventions, baked in. Branch names, labels, commit style, test commands — written once, applied everywhere.
- Repetition becomes one command. The boring multi-step rituals collapse into a single invocation you can trigger anywhere.
- Shareable and auditable. It’s just markdown — hand a teammate your workflow in one PR, and read exactly what the agent will do before it does it.
- 1Open the issue, re-read the scope
- 2Think up a branch name that fits convention
- 3Create the branch (worktree if it’s risky)
- 4Write a plan doc — goals, steps, test plan
- 5Link the plan back to the issue
For months I started every task the same way: open the issue, re-read the scope, invent a branch name that matched our convention, create the branch, spin up a plan doc, list the steps, link it back. The same ten-minute warm-up before any real work — every single time.
One afternoon I wrote the whole ritual down as a skill. Now I type /issue-to-branch #482 and the agent runs the entire warm-up in one shot, the same way every time. That was the moment skills clicked for me: the parts of my day that were pure repetition turned into a single command.
Not on the stack a skill was written for? Adapt it#
A skill is a written playbook, not a compiled binary — so it doesn’t have to be ported by hand. Every skill page ends with an Adapt to your platform prompt: paste it into your agent, tell it your stack, and the agent rewrites the skill for your project. Each skill is tagged with how well it travels:
Built something other teams could use?#
A skill is just a markdown file — if your team has a workflow worth sharing, contributing it takes one PR. How to contribute →