Skills and subagents for Claude Code that carry a change from an issue to a merged PR.
Engineers who work through GitHub issues and PRs with Claude Code and want a repeatable flow: pin down the problem, build with subagents, review, ship, and keep context across sessions. The skills report what they find and leave the decisions to you.
| Name | What it does |
|---|---|
ground |
Understand the problem behind a change before any work starts, then the approach. Read-only. |
grill |
A Q&A session that refines a problem and its solution until nothing is left assumed, then records the decisions. |
orchestrate |
Build a change end to end as the lead: subagents implement, you plan and review, then review, open the PR and post findings. |
deep-review |
Pre-merge review of the branch diff: the built-in /code-review plus one Opus agent per axis (conventions, fit, tests, spec match), merged into one verdict. |
gate |
Pre-push gate: runs the repo's own complexity or lint ratchet if it has one, then an Opus review of new code for KISS, YAGNI, DRY and surgical change. |
codex-triage |
Validate Codex review findings on a PR, fix only the real in-scope ones, reply on every thread, sync main and push. |
ship-pr |
Push, open or find the PR, and drive it to merge-ready: watch CI, fix Sonar findings, triage Codex, re-watch. |
cleanup |
After PRs merge, remove their worktrees and branches and bring the main checkout to latest main. |
debug |
Find why something is red, failing or stuck, fix the root cause and pin it with a test. |
perf |
Measure a performance problem, change one thing, re-measure, and report before and after numbers. |
research |
Answer "which is better" questions with the repo's current practice, credible sources and one recommendation. |
handover |
Write a handover note and a paste-ready prompt so a fresh session can resume cold. |
recap |
Say where the work stands from real state, not from memory of the conversation. |
retro |
Look back over a session and suggest changes to checks, steering files, memory, skills and tooling. Manual only. |
day-update |
Build the day's work update from git, GitHub and session transcripts across repos, and append it to ~/updates.txt. |
minutes |
Turn meeting notes or a transcript into decisions, action items and open questions, checked against the real work. |
issue |
Raise a GitHub issue, or add a report comment to an issue or PR, with facts and evidence only. |
ui |
Design and build a UI change, mockup first: show it, wait for the OK, build in the repo's style, check in a browser. |
builder (agent) |
Implements one bounded slice of a written plan in a worktree, runs the repo's checks, commits, and reports. |
merge-main (agent) |
Commits pending work, merges the latest base into the feature branch, resolves conflicts base-first, never pushes. |
test-runner (agent) |
Runs the requested tests, returns exact counts and each failure's class, method and cause. Never fixes code. |
/ground → /grill → /orchestrate → /deep-review → /ship-pr → /codex-triage → /cleanup
Run /gate before a push if the repo has a gate. /handover and /recap bridge sessions; /debug and /perf are used when something breaks or is slow.
git clone <this repo> && cd <this repo>
mkdir -p ~/.claude/skills ~/.claude/agents
ln -s "$PWD"/skills/* ~/.claude/skills/ # or cp -r
ln -s "$PWD"/agents/* ~/.claude/agents/ # or cpOptionally adapt CLAUDE.example.md into ~/.claude/CLAUDE.md. It holds the personal rules the skills are written to pair with: short answers, ask on decisions, minimal changes, subagent conventions.
Some skills use placeholders instead of fixed values. Set them by naming them in your CLAUDE.md:
<notes-dir>— where ledgers, design records and handovers live, outside your repos.<repos-dir>— where your local checkouts live.<org>— your GitHub org or user, for cross-repo issue and PR searches.<project-key>— your Sonar project key.
Project-specific facts, such as the test command or gate, come from each repo's own CLAUDE.md and docs.
- Claude Code
ghCLI, authenticated- Optional: Codex CLI for
codex-triage - Optional: Sonar CLI for the Sonar steps in
ship-pranddebug
- Ground claims in code. Trace whether a real caller reaches a concern before raising or fixing it.
- Subagents are leaves with small tasks and a to-do list. The lead keeps context for decisions.
- Run targeted tests, not full suites, and send module-wide runs to
test-runnerso output stays out of the lead's context. - Check before outward actions:
issuedrafts before posting,codex-triagesupports--askand--no-push,ship-prasks first where CI is live. - Reply with plain, short output — the verdict first.
MIT. See LICENSE.