Docs: Issue-to-PR Skills

3 ways in: pick one

Any agent · easiest

npx skills add AI-Degen-69/issue-to-pr-skills

Works with Claude, Cursor, Copilot, OpenCode and 70 more. Try --list to browse first.

Claude Code · plugin

/plugin marketplace add AI-Degen-69/issue-to-pr-skills/plugin install issue-to-pr-skills@ai-degen-marketplace

Auto-updates.

Manual

copy skills/ → .agents/skills/

Copy folders yourself. See getting-started guide.

What you need

git + gh (GitHub CLI). CodeRabbit is optional: without it, the AI does the review.

First run: 1-2-3

1. Open a repo with issues → 2. Type /i-pick-issue → 3. Pick an issue and follow the stations.

How to know it worked

Each station ends with a short report: what was done, proof it works, and the single next step. Run npm run validate to check locally.

The words you'll see, in plain English

No jargon needed to use this. Here is what the recurring words mean.

Station

One step of the process. There are seven, in order: pick the task, plan it, build it, fix what you flag, review it, merge it, tidy up.

Skill

A set of instructions the AI follows for one kind of job. Installing the pack gives your AI all 47 of them.

Agent (reviewer)

A specialist the AI calls in to double-check the change. Seventeen exist; it only summons the ones that fit.

Issue

A reported problem or requested change, written as a task. It is the unit the whole pipeline works on.

Proof gate

A checkpoint where the AI must show the change actually works — a passing test or a live look — before moving on.

Repo

Short for repository: the folder where your project and its whole change history live.

Merge

Folding a finished change into the main project once it has passed review and the automated checks.

Tests

Automatic checks that run the software and confirm it still behaves. They are the "proof" the gates rely on.

Closeout

The final tidy-up after a merge: mark the task done, delete leftovers, leave the project clean.

Why it works: 3 mistakes it stops

A normal AI tends to take the shortcut. Each of these is a checkpoint that makes it stop and prove itself first.

Building the wrong thing

Before any code, the AI writes down the plan and the rules it must not break — so it builds what you actually asked for, not what it guessed.

Shipping something broken

The AI has to show it works — a passing check or a real look in the browser — before anything can be merged. No "looks fine to me".

A review that approves everything

Instead of one rubber stamp, several specialists each check their own angle, and the "did we build the right thing" check is kept separate from the "is it built well" check.

Your first 10 minutes

Three steps. You do not need to read any code.

Step 1
Install
Run one command (see "3 ways in" above). It gives your AI all 47 skills at once.
Step 2
Open a project and start
Open any project folder, then type /i-pick-issue. The AI lists the open tasks and suggests which to do first.
Step 3
Let it run, step by step
Follow the stations. Each one finishes with a short report — what it did, proof it works, and the one next step. Nothing is guessed.

Changes land in CHANGELOG.md, one line per fix.

Tutorials: pick your path

What to type, depending on what you are trying to do.

Start something new

You have an idea, not a task yet. Type /create-issue and describe it in a sentence. The AI researches it and writes it up as a proper task, then plans, builds, and shows you the result.

Add a feature

There is already a task for it. Type /i-pick-issue, pick it, and the AI plans it, builds it in small steps, gets it reviewed, and merges it once everything passes.

Fix something broken

Describe what is going wrong. The AI first reproduces the problem, fixes the actual cause (not just the symptom), adds a check so it cannot come back, then reviews and closes.

How it compares

What you get here versus dropping a single generic skill into your AI.

	A single generic skill	This full pipeline
How much it handles	One small task at a time	The whole job, from idea to finished and merged
Proof it works	Optional, often skipped	Required before every review — it must show you
Who checks it	One quick pass	Several specialists, and a separate "right thing?" check
After it finishes	You clean up the leftovers	It tidies up and marks everything done for you

Common worries, answered plainly

The questions first-timers ask before trusting an AI with their project.

What if the AI gets it wrong?

It is not allowed to guess past a gate. If it cannot prove a step works, it stops and tells you what is missing instead of carrying on. Every change is reversible, and it never pushes anything without showing you first.

What if it changes something I did not ask for?

The plan is written and agreed before any change is made, so it works to that plan. If a change turns out to be outside it, it flags it rather than quietly doing it.

Can I stop partway and come back?

Yes. Nothing is lost if you stop. When you return, type /pipeline-triage and it checks exactly where things left off and points to the next step — no re-explaining needed.

Does it need anything installed?

Just git and gh (GitHub's tool). Everything else is optional — if the extra reviewer is not set up, the AI does the review itself instead.

Is it going to cost me anything?

No. It is free and open-source forever. You only pay for the AI you already use to run it.

What if I just have a tiny fix?

You do not have to run the full process. Type /quick-fix for a one-line change like a typo. It double-checks the change is safe to do quickly, then does it.

Layer	Paths	Purpose
Shared core	skills/ (47)	Workflows
Reviewers	agents/ (17)	Specialists
Docs	docs/	Guides
Site	site/	This website

Get your AI to follow a real process.