wjs-looping-feedback
Use when the user wants to add an in-site feedback loop to a website repo — a floating "提个建议" button where allowlisted visitors submit suggestions that become a GitHub Issue, which GitHub Actions turns into an automatic code change via Claude Code, auto-merges and deploys, and records on a /_feedback dashboard with one-click revert. Triggers — "给网站加个反馈对话框", "提一句话就自动改网站", "装上反馈闭环", "feedback loop", "/wjs-looping-feedback".
How do I install this agent skill?
npx skills add https://github.com/jianshuo/claude-skills --skill wjs-looping-feedbackIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill implements an automated feedback loop that turns website suggestions from authorized GitHub users into direct code changes via Claude Code. While useful, it creates a surface for indirect prompt injection: an authorized user could submit a malicious issue to manipulate the AI agent, which is granted high-privilege shell access (Bash tool) and write permissions to the repository. The security of the system depends entirely on the accuracy of the manual allowlist and the AI's internal guardrails.
- Socketwarn
4 alerts: gptAnomaly, gptSecurity
- Snykfail
Risk: CRITICAL · 3 issues
What does this agent skill do?
wjs-looping-feedback
Installs a self-driving feedback loop into any website repo. Runs entirely on the repo
owner's own GitHub Actions, authenticated with their Pro/Max OAuth token or their own
ANTHROPIC_API_KEY — no extra service or backend.
What it installs
- A floating feedback button (prefills a GitHub Issue — no backend, no keys client-side).
.github/workflows/feedback.yml— allowlist gate → Claude Code Action → auto-commit → push tomain(auto-deploy) → ledger + dashboard update → close issue./_feedbackdashboard — every suggestion, what Claude did, commit, status, one-click revert..feedback/runtime:feedback-lib.mjs,feedback-finalize.mjs,INSTRUCTIONS.md,feedback-ledger.json,config.json.
How it works
visitor clicks button → fills suggestion → prefilled GitHub Issue (label feedback) →
Actions checks the author is in FEEDBACK_ALLOWLIST (else closes the issue) →
Claude Code edits the site per .feedback/INSTRUCTIONS.md → workflow commits to main,
updates the ledger + dashboard → the deploy ships it →
the issue gets a comment with the commit and a link to /_feedback.
Revert is the same loop driven by a revert: #N issue.
Deploy triggering — read before you install (common gotcha)
The workflow pushes to main with the built-in GITHUB_TOKEN. Two cases:
- External push-deploy host (Cloudflare Pages git-connect, Vercel, Netlify): their webhook fires on every push, so the bot's commit deploys automatically. Nothing to do.
- Deploy is itself a GitHub Actions workflow in the SAME repo (e.g. GitHub Pages via
Actions): GitHub deliberately does NOT let a
GITHUB_TOKENpush trigger another workflow (recursion prevention), so the deploy workflow'son: pushwill NOT fire — the change lands onmainbut never goes live. Bridge it by adding aworkflow_runtrigger to the deploy workflow (seereferences/install.mdstep 4.5).
To install
Read references/install.md and follow it in the target repo. It detects the site type
(Hugo / Next.js / Astro / static), copies assets, injects the widget, and asks the user for
exactly two things: the allowlist of GitHub usernames and the auth (Pro/Max OAuth
token via claude setup-token, or an ANTHROPIC_API_KEY).
Scope (route 1)
Runs in the user's own repo with their own auth. Submitters need a GitHub account; only allowlisted authors trigger work. Not a hosted service — no anonymous submit, no billing, no multi-tenant. (A future GitHub-App packaging would reuse every asset here.)
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/jianshuo/claude-skills/wjs-looping-feedback">View wjs-looping-feedback on skillZs</a>