Find contributable issues

Find Contributable Issues

Investigate open issues in a GitHub repository and rank them by how contribute-able they are, so a contributor can quickly see what's worth picking up.

What This Skill Does

One thing: read GitHub issue data via gh CLI, score it, present a ranked table. Read-only investigation — no comments, assignments, labels, PRs, or issue state changes. The skill ends after presenting the ranked table; Steps 5–6 are optional follow-ups the user may request.

Prerequisites

Run this first. If it fails, stop and tell the user to run gh auth login:

gh auth status

This skill never handles tokens or credentials — gh uses the user's existing auth.

Workflow

Step 1: Resolve the target repository

• If the user gave an explicit owner/repo, use it.

• Otherwise detect the current repo:

  gh repo view --json nameWithOwner -q .nameWithOwner
  

• If neither yields a repo, ask the user for owner/repo and stop.

Report the resolved repo before continuing.

Step 2: Fetch open issues (one pass, capped at 30)

gh issue list \
  --repo <owner/repo> \
  --state open \
  --limit 30 \
  --json number,title,body,labels,assignees,reactionGroups,closedByPullRequestsReferences,comments,updatedAt

Notes:

• --limit 30 caps the number of issues analyzed. If the user wants a different or larger set, they say so and you re-run Step 2 with a refined query (e.g. --search "label:bug" or a higher limit) — do not silently raise the cap.
• comments returns full comment bodies, which is the largest field. It is fetched deliberately: Step 3a needs the comment count and Step 3e needs comments[].authorAssociation. Do not treat the 30-cap as a token-cost bound on its own — the comments payload dominates cost. If the repo is very chatty, consider --search "comments:<5" to pre-filter.
• closedByPullRequestsReferences is the authoritative linked-PR signal — see Step 3c. Do NOT use gh pr list --search "fixes #N"; that search is a full-text match and returns unrelated PRs in active repos.
• createdAt is not needed (staleness uses updatedAt); omit it.
• If zero open issues, report that and stop.

Step 3: Score each issue

Compute six indicators per issue, in memory. Walk comments once per issue to get both the count (3a) and maintainer-engagement flag (3e).

3a. Difficulty — label-first, heuristic fallback (SOURCE-TAGGED)

GitHub has no built-in difficulty field. Derive it, and always tag the source so the contributor knows how much to trust the rating.

Label detection (authoritative when present): scan labels[].name case-insensitively. More specific labels (difficulty:*, effort:*) win over generic ones (help wanted).

Label pattern	Maps to
good first issue, beginner, starter, up-for-grabs, easy	easy
difficulty:easy, difficulty:hard, etc.	the named band
effort:S / effort:M / effort:L	S=easy, M=medium, L=hard
effort:<n>d (days)	≤3d=easy, ≤7d=medium, >7d=hard
help wanted (alone, no other difficulty signal)	medium

→ Source tag: label.

Heuristic fallback (when no difficulty label is found): estimate easy/medium/hard from body specificity, comment volume, and code/file references in the body. Directionally: a long body with many comments and references to specific files/architecture is harder; a short body with reproduction steps or code blocks is easier. Map your estimate to a band.

→ Source tag: estimated.

Display format: easy (label) or medium (estimated) — always both band and source. Never present an estimated difficulty as authoritative.

3b. Claimed

• assignees array non-empty → claimed, show ✓ claimed (@login).
• Empty → open, show — open.

3c. Linked PRs (from the Step 2 payload — no extra calls)

Read closedByPullRequestsReferences (already fetched in Step 2). This is GitHub's authoritative closing-reference list — the PRs that would close this issue. Display: count + states if present, e.g. 2 (1 merged, 1 closed), or none.

Do NOT issue per-issue gh pr list calls. --search "fixes #N" is a full-text search that returns many unrelated PRs in busy repos and is not a reliable closing-reference resolver.

3d. Staleness

Days since updatedAt. Show as a compact relative time, e.g. 3d, 2w, 6mo.

3e. Maintainer engagement

During the single pass over comments, check each comments[].authorAssociation. If any is OWNER, MEMBER, or COLLABORATOR → yes, else no. This signals whether a PR is likely to get reviewed.

3f. Area

Labels first (e.g. area:frontend, area:backend, scope:docs). Else keyword scan of title + body:

Keywords	Area
UI, button, render, css, component, layout	frontend
API, route, handler, endpoint, server, db	backend
test, spec, coverage	test
README, docs, documentation	docs
CI, deploy, docker, infra, build	infra

No match → unknown.

Step 4: Rank and present

Rank issues by these factors, in priority order (the score is for sort order only — DO NOT show a numeric score in the table):

1. Unclaimed issues rank above claimed ones.
2. No linked PR (closedByPullRequestsReferences empty) ranks above issues with linked PRs.
3. Lower difficulty ranks above higher (easy > medium > hard).
4. Fresher (updatedAt more recent) ranks above staler.
5. Maintainer engaged ranks above not, as a tiebreak.
6. Reaction count desc as the final tiebreak. Reaction count = sum(reactionGroups[].users.totalCount) over all reaction types (NOT reactionGroups.length, which is just the count of reaction types present).

Output a single Markdown table:

#	Title	Difficulty	Claimed	Linked PRs	Stale	Maintainer	Area
123	Fix login redirect loop	easy (label)	— open	none	3d	yes	backend
87	Add dark mode toggle	medium (estimated)	— open	none	2w	no	frontend

Above the table, show the resolved repo and the issue count analyzed (Analyzed 30 open issues in owner/repo). Keep titles short; truncate to ~50 chars with ….

The compact summary in the Report Format section is shown in addition to the table, not instead of it.

Step 5: Refine (interactive, no refetch unless needed)

The user can refine with natural language. Re-filter against the in-memory scores — do NOT re-run gh unless the user changes repo or asks for more than 30 issues (in which case go back to Step 2).

Recognized refinements:

• Difficulty: "only easy", "hide hard"
• Claimed: "unclaimed only", "show only claimed"
• Linked PRs: "no existing PR", "already has a PR"
• Staleness: "fresh only (<14d)", "hide stale (>90d)"
• Maintainer: "maintainer engaged", "no responses"
• Area: "backend only", "hide docs"
• Combined: "unclaimed easy backend issues"

Re-present the filtered table. State the active filter line above the table (Filter: unclaimed, easy, backend — 4 of 30 issues).

Step 6: Explain on request

If the user asks "why is #123 ranked high?" or "tell me about #87", look up the cached scores for that issue and explain each factor's contribution. Pull a short excerpt from the issue body (first ~200 chars) for context. Do NOT re-fetch unless the user asks for something not in the cache.

Boundaries

• Read-only. Never comment, assign, label, close, or open a PR. If the user asks to take action ("just claim #123 for me", "comment that I'll take it"), do NOT do it — give them the exact gh command to run themselves (e.g. gh issue edit 123 --add-assignee @me). The user asking is not authorization to write; this skill only investigates.
• No secrets. Never touch tokens, gh config, or credentials.
• Cap is intentional. 30 issues bounds the analysis set. The comments payload is the dominant cost, not the issue count. If the user wants broader coverage, refine the gh issue list query and re-run Step 2.
• Heuristics are estimates. Always tag difficulty source (label vs estimated). Never present an estimated difficulty as authoritative.

Report Format

In addition to the Step 4 table, end with a compact summary:

Repo: owner/repo
Analyzed: 30 open issues
Filter: <none | active filter line>
Top pick: #123 — Fix login redirect loop (easy, unclaimed, maintainer engaged)

Common Mistakes

• Skipping gh auth status. If gh isn't authenticated, every later step fails confusingly. Check first.
• Using gh pr list --search "fixes #N" for linked PRs. That is a full-text search returning unrelated PRs in busy repos. Use closedByPullRequestsReferences from the Step 2 payload instead.
• Taking write action on GitHub because the user asked. "Just claim it for me" is not authorization. Hand back the gh command; do not run it.
• Guessing the repo. If no explicit arg and repo detection fails, ask. Don't invent an owner/repo.
• Treating estimated difficulty as fact. Always show the source tag.
• Showing a numeric composite score. Rank order is the output; the score is internal.
• Using reactionGroups.length for reaction count. That is the number of reaction types present (max 8). Sum reactionGroups[].users.totalCount.