Open Contribute

한국어 | English

The problem

good first issue labels lie. On a popular repo the labelled issue you find is usually already claimed, already fixed by a different PR, or quietly resolved with nobody closing the ticket. You read the code, get invested, and then find a linked PR from six months ago.

That is the visible half. The invisible half is worse: a repo can look thriving — daily commits, hundreds of merges a quarter — while merging almost nothing from anyone outside its core team. Stars and activity don't tell you whether your PR will land.

Same day, same method: grafana/grafana merges 70% of its PRs from outside its five most frequent authors. chroma-core/chroma merges 8% — and every one of the 17 PRs opened against its good-first-issues had failed to merge. Both look healthy from the outside.

Then there's the race. Anything recent, well-scoped and labelled tends to be claimed within hours.

This skill measures all of that before it recommends anything.

What's in the box

Each is a self-contained Agent Skill: a SKILL.md plus a references/ folder loaded on demand.

How it works

1. Narrows down what you want — domain, sub-area, stack, experience level, contribution type — before searching, so the recommendation rests on your preferences rather than a guess.
2. Filters on availability first. A repo with no currently-open, unassigned contribution issues can't produce a contribution today, and that check is one cheap API call. Only the survivors get the expensive evaluation.
3. Runs the PR merge reality gate — the share of open PRs sitting 90+ days, and the share of merges from outside the top-5 authors. You see the numbers and decide; the skill never silently drops a repo.
4. Triages every candidate issue against an 8-step checklist: internal-only notices, linked PRs, assignees, resolving comments, claim pile-up, age — and finally whether the defect still exists in current code, because a fix can land without anyone closing the issue.
5. Drafts the outreach — claim comments, community-channel messages, PR descriptions — in the project's own register.
6. Walks the contribution — gh-based fork and clone, scoped edits, the project's real test suite, a pre-PR quality checklist covering DCO, signatures and CLAs, then submission and follow-up etiquette.

Prerequisites

The skill drives forking, cloning and PR creation through the GitHub CLI (gh), so no manual personal access tokens.

• Mac: brew install gh
• Windows: winget install --id GitHub.cli
• Linux: sudo apt install gh

gh auth login --scopes workflow

If you hit SSH host key errors later, run gh config set git_protocol https.

Install

Claude Code — install as a plugin

Two commands, all three skills, no cloning:

/plugin marketplace add chjnett/open_contribute
/plugin install first-contribution@open-contribute

Or paste this prompt to your coding agent

For agents without plugin support. Swap the folder name for whichever variant you want.

Install the first-contribution agent skill for me, and nothing else.

1. Figure out my skills directory (Claude Code → ~/.claude/skills/,
   otherwise my tool's documented skills directory).
2. Clone https://github.com/chjnett/open_contribute into a temp folder.
3. Copy the first-contribution/ subfolder into <skills directory>/first-contribution/.
4. Confirm SKILL.md exists at the root of the copied folder.
5. Do not modify any other files, settings, or skills. Report the final
   install path and stop.

Restart your agent to pick up the skill.

Claude Code — copy the folder yourself

git clone https://github.com/chjnett/open_contribute /tmp/open_contribute
cp -r /tmp/open_contribute/first-contribution ~/.claude/skills/first-contribution

Claude.ai / Cowork — upload a packaged .skill

Download the variant you want from the latest release and upload it under Settings → Skills. To build the files yourself from a clone:

bash scripts/package_skill.sh

Use it

The skill triggers on intent — no commands:

"I want to start contributing to open source, help me find a good repo and issue"

"Is this repo beginner-friendly? https://github.com/langchain-ai/langchain"

"Find me a good-first-issue in a RAG/Python project."

Field notes

Every rule in the checklist was added after it cost a real run something.

What it has produced

Eighteen pull requests, each found and root-caused by the skill from a cold start — four already merged, the rest in review:

Two other outlines PRs (#2013/#2014) and older attempts on sqlfluff (#8350) and directus (#28092) were withdrawn or superseded during review; each taught a documented lesson (competing-PR consolidation, AI-disclosure, reviewer-widens-scope) rather than contributing to the merge count.

The second run is the better illustration of the funnel: 30 backend repos screened for availability → 13 with open contribution issues → 11 through the merge gate → 229 candidate issues checked for linked PRs → 26 still claimable → one with a defect verified live in main.

What each rule cost

• chroma-core/chroma — 29k stars, pushed daily, 11 open good first issues. Also 275 of 452 open PRs sitting 90+ days, 8% of merges from outside the top-5 authors, and 17 of 17 PRs against those labelled issues failed to merge. The gate exists because of this repo.
• deepset-ai/haystack — passes the gate comfortably, and its contribution labels are empty: 199 such issues total, zero open. An empty label on a healthy repo means "come back later," which is the opposite diagnosis from a rotten one.
• argoproj/argo-cd#29051 — 8 days old, unassigned, zero linked PRs, severity:critical. It passed every social check and had already been fixed by v3.3.14. That is why the checklist now verifies the defect against current code.
• Both PRs' existing tests asserted the buggy output. Grafana's expected orgId=org-5; Directus's would have broken on the obvious fix because its mock schemas define relations and no collections. Read the fixtures before choosing how to write a fix.
• CLAs are not all the same. Grafana uses an external signing service; Directus asks you to add your username to contributors.yml inside the PR — a one-line diff that is nonetheless the act of signing. The skill will never do that for you.
• A second-hand summary of your own PR is a claim, not the state. One arrived asserting 28 workflows awaited approval on the Grafana PR; the API showed zero. Acting on it would have burned the single follow-up nudge on a non-problem.

Reproducing the charts

The merge-gate and triage charts come only from measurements taken on 2026-08-12, with the method printed on each chart; the contribution charts track PRs opened through this skill via gh (as of 2026-08-28):

python3 scripts/generate_charts.py

Troubleshooting

Contributing

Built and refined in public while being used for real contributions. If it recommends a stale issue, misjudges a repo, or misses a project convention, open an issue — that feedback is what the checklist is made of.

License

MIT. Use it, fork it, build on it.