- Home
- Skills
- Agents & Orchestration
- ai-loop-skills: Issue-to-PR workflow for Claude Code
Works with the AI Agents you already use
ai-loop-skills: Issue-to-PR workflow for Claude Code
It separates the commander, which designs and checks, from the implementer, which writes the code.
$19
ai-loop-skills: Issue-to-PR workflow for Claude Code
Example session with this skill installed
Act as the commander using the commander skill. I want the README to have a short "Troubleshooting" section that tells users what to check when the verify command fails. Keep it to a few bullet points.
- Read your context and instructions
- Compiled the ai-loop-
Summary
One Issue (#12) and one pull request (#13) for a 9-line README change: drafted for approval, implemented by a subagent, checked by the commander, and cleaned up after the merge. Nothing was created before the draft was approved, and the merge was left to the user.
1. The draft, before anything is created
The commander showed the full change and the Issue text, then stopped and waited for "approve".
+## Troubleshooting
+
+If the verify command (`bash .claude/ai-loop/verify.sh`) fails, check the following.
+
+- Run it from the repository root. If it prints `verify: not a git repository`, you are outside the repository.
+- Check that Python 3 is available with `python3 --version`.
+- Run `python3 -m unittest -v` and read the name and message of each failing test.
+- If a test you added does not run, check that its file name starts with `test` and ends with `.py` (`python3 -m unittest` only looks for files named this way).
Completion conditions written into the Issue: verify exits 0 with VERIFY PASSED; README.md matches the approved sha256 and has 12 lines; exactly 1 ## Troubleshooting heading and 4 bullet lines; git diff --numstat lists only README.md, with 0 removed lines.
2. After approval: the commander's report
PR #13 is ready for you to merge: "Add a Troubleshooting section for verify failures to README.md" (closes #12). I recommend merging it.
| What the commander checked itself | Result |
|---|---|
| Commits | One commit, 0540abf, with no trailers (count 0) |
| Diff | 9 0 README.md. Only README.md changed, with no removed lines |
| README.md at the PR head | Byte-identical (cmp) to the approved version |
| Commit subject and PR title | Both match the finalized text (sha256) |
| PR body | The cloud session had added an attribution footer. Restored to the finalized body; the fetched body now matches (sha256) |
| Closing keyword | check_closing_keywords.py exits 0 with Line 1: #12 |
| CI | n/a (0 check-runs), so the commander ran verify itself in a worktree at the PR head: exit 0, 9 tests, VERIFY PASSED |
| Review and blockers | No automatic review. No blockers. Nothing was sent back |
The implementer's final response (two lines omitted):
Result: done
Issue: #12
Branch: claude/skill-pack-update-review-rwyer1
Commits: 0540abf
Changed files: M README.md
verify: exit=0 (9 tests OK, 0 failed, VERIFY PASSED)
pre-push-check: verify=0, personal-info scan=0
PR: #13
PR body: mismatch (commander to handle)
Stop reason: none
Self-check: personal-info scan exit 0 (0 warnings)
The commander also caught a mistake in that report. The implementer said the branch did not exist on origin before its push. The commander looked at the history of the remote-tracking ref, found that the push had fast-forwarded an existing branch, and reported the correction. The result was the same, and nothing needed fixing.
3. After the merge: cleanup
The user merged PR #13 on GitHub and asked for the cleanup.
The post-merge cleanup for PR #13 (Issue #12) is done, and every step passed.
| Step | Result |
|---|---|
| Merge | PR #13 is merged, with merge commit 739554e |
| Issue | #12 is closed with state_reason: completed, closed automatically by Closes #12 |
| verify on the updated main | verify: test command exit 0, 0 failed, VERIFY PASSED |
| Branch | The local branch was deleted after the ancestor check gave YES. GitHub had already deleted the remote branch |
Next steps for the user: none. The repository is on main at the merge commit, and the commander is waiting for the next request.
From a real session in a Claude Code cloud session. The replies are shortened: work-directory paths and repeated hashes are omitted, and the attribution footer is specific to cloud sessions.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
ai-loop-skills is a pack of 8 skills that takes a request from a GitHub Issue through implementation and a pull request to the post-merge cleanup, following a fixed procedure.
It separates the commander, which designs and checks, from the implementer, which writes the code. Verification, scanning and comparison are built into the procedure, so nothing unchecked ends up in a report. Merging and closing Issues stay with you.
How it works
- You describe what you want to the commander.
- The commander splits the request into Issues and shows you the draft bodies. It creates them only after you approve.
- The commander hands an Issue to the implementer (a subagent). The implementer creates the branch, implements, verifies, runs the pre-push checks, pushes and opens the PR.
- The commander reads the CI and review results and reports its recommendation.
- You merge the PR on GitHub.
- The cleanup skill updates the merge target, verifies, checks the Issue state and deletes the branch.
What is included
- setup: first-time setup and re-diagnosis. Shows the full text of every file before creating it.
- commander: creates Issues from a request, delegates to the implementer, reads CI and review, decides what to accept.
- publish-github: creates an Issue or PR from finalized text only, then fetches the body and confirms it matches.
- pre-push-check: verification, diff scan, added-line comparison and extra checks before every push.
- post-merge-cleanup: the housekeeping after you merge.
- diagnose-failure: identifies why CI, review or a run failed.
- report-self-check: scans a report for personal information before it is sent.
- replace-doc-range: replaces part of a document and confirms nothing else changed.
Safeguards
- Permission settings stop Claude Code from merging, closing Issues, pushing to the default branch, force pushing, and editing the verify command or the settings. You do these yourself.
- Pass or fail is decided by the exit code of your verify command and of CI, not by the agent's impression.
- Commits and PRs carry no attribution lines.
- Scripts are never saved in your repository. Each one is written to a temporary directory, checked against its sha256, run, and removed.
- When something does not match the procedure, the skills stop and report instead of working around it.
Good to know
- It follows the full Issue and PR procedure even for small changes, so it uses more time and tokens than asking Claude Code directly. It is meant for work where you want a checked, reviewable trail.
- CI and review workflows are not part of the pack. In a repository without CI, the result of your verify command is used.
- The skills are written in English and respond to you in your language. Issue and PR text is written in the language you use.
Claude Code only. Developed and tested with Claude Opus; other models are untested. GitHub only (no GitLab or Bitbucket). It runs the full Issue and PR procedure even for small changes, so it uses more time and tokens than a direct request. CI and review workflows are not included. Merging and closing Issues are left to you by design. macOS has not been tested on a real machine yet. The pack handles the known differences (shasum instead of sha256sum, /tmp shown as /private/tmp), but other differences may remain.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 1 day ago
Requires
Permissions
Allowed Hosts
File Scopes
Shell: runs git, gh and your verify command. Network: talks to GitHub through gh only. Write: setup creates files under .claude/ after you approve their full text, and the implementer edits the files your request needs. Scripts run from a temporary directory and are removed afterwards.