svelte-mcp-feedback — Claude Code skill: use and setup
svelte-mcp-feedback is a Claude Code skill that turns corrected Svelte and SvelteKit mistakes into autofixer feedback. It runs after a correction, not during an initial implementation.
MIT 102 installs ★ 344 updated 2026-10-23
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a claude-codeWhat svelte-mcp-feedback does
svelte-mcp-feedback converts an agent mistake corrected by the user into a report for the Svelte MCP autofixer. It applies the correction, identifies the relevant rule, reduces the code to a minimal reproduction, and describes the expected transformation.
When the MCP is available, the skill checks its documentation and runs svelte-autofixer on the reduced example if that file type is supported. It searches sveltejs/ai-tools for matching issues, then can create one with GitHub CLI and --repo sveltejs/ai-tools. If gh is unavailable or unauthenticated, it falls back to a prefilled GitHub issue form. Personal style choices and unrelated application bugs do not start this flow.
What it helps with
- extract a reusable Svelte mistake from a user correction
- reduce a project bug to a minimal reproduction
- check whether
svelte-autofixermissed the case - search
sveltejs/ai-toolsfor a duplicate issue - prepare a focused MCP autofixer report
Example prompts for the agent
- I corrected your mistake in a
.sveltecomponent. Turn the pattern into an autofixer issue. - Review this fix in
+page.tsand check whether Svelte MCP already detects the problem. - Reduce the error in
hooks.server.ts, runsvelte-autofixer, and file a non-duplicate report.
How to install svelte-mcp-feedback
1Skills CLI (Vercel)
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a claude-code2Manually: put the files in place
~/.claude/skills/svelte-mcp-feedback/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/svelte-mcp-feedback/
1Skills CLI (Vercel)
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a codex2Manually: put the files in place
.agents/skills/svelte-mcp-feedback/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a cursor2Manually: put the files in place
.cursor/skills/svelte-mcp-feedback/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install sveltejs/ai-tools tools/skills/svelte-mcp-feedback2Manually: put the files in place
.github/skills/svelte-mcp-feedback/SKILL.md1Skills CLI (Vercel)
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a gemini-cli2Platform CLI
gemini skills install https://github.com/sveltejs/ai-tools --path tools/skills/svelte-mcp-feedbackManually: put the files in place
.devin/skills/svelte-mcp-feedback/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about svelte-mcp-feedback
How do I install svelte-mcp-feedback in Claude Code?
Run npx skills add sveltejs/ai-tools --skill svelte-mcp-feedback -a claude-code. The skill file will be placed at ~/.claude/skills/svelte-mcp-feedback/SKILL.md.
What does svelte-mcp-feedback do?
It reacts to a corrected Svelte or SvelteKit mistake, builds a minimal reproduction, checks the autofixer result, and helps create an issue in sveltejs/ai-tools.
What does svelte-mcp-feedback need to work?
An available svelte-autofixer is used when it supports the file type. Creating an issue requires an installed and authenticated GitHub CLI gh; a GitHub browser form is the fallback.
Source text
SKILL.md · MIT
--- name: svelte-mcp-feedback description: Report Svelte MCP autofixer opportunities when the user corrects a Svelte or SvelteKit mistake the agent made. Use after corrections to .svelte components, .svelte.ts/.svelte.js modules, or SvelteKit features where a reusable check or fix could prevent the mistake. Do not trigger for initial implementation requests, unrelated bugs, or personal style preferences. --- # Svelte MCP feedback Turn a user-corrected mistake into actionable feedback for the Svelte MCP autofixer in `sveltejs/ai-tools`. ## 1. Identify the autofixer opportunity Use the conversation and the code you just worked on to identify: - The incorrect code you produced and the user's correction. - The Svelte or SvelteKit rule that explains the correction. - A reusable code pattern the MCP could detect, and the fix or suggestion it could provide. Include SvelteKit JavaScript/TypeScript files such as `+page.ts`, `+layout.server.ts`, `+server.ts`, `hooks.server.ts`, and `svelte.config.js`; feedback is not limited to `.svelte` files. Only proceed if there is a concrete, generalizable Svelte/SvelteKit mistake. A changed product requirement, a personal naming preference, or an unrelated application bug does not justify an autofixer issue. If the correction is ambiguous, clarify it while continuing the user's task. Apply the user's correction and complete the requested work. Reporting feedback should not leave the original mistake unfixed. ## 2. Capture a minimal reproduction Reduce the before/after code to the smallest example that preserves the mistake. Use generic names and sample data instead of copying unrelated project code, credentials, or conversation history. If the Svelte MCP is available, check its relevant documentation and run `svelte-autofixer` on the incorrect example when it supports that file type. Record whether it missed the issue, suggested an incorrect fix, or already reported the issue and you overlooked it. Do not claim the tool missed something without checking its output. If the tool is unavailable or does not support the file type, say so; a proposed new rule can still be useful. If an existing diagnostic already clearly covers the correction, follow it instead of opening a new-rule request. Report only a concrete gap, such as an unclear diagnostic, a wrong suggested fix, or a missed case. Prepare a title like `Autofixer: detect <incorrect pattern>` and a Markdown body with: 1. **Summary** — the mistake and why it matters. 2. **Incorrect code** — a fenced minimal example with the relevant filename/file type. 3. **User correction / expected code** — the corrected example and a brief explanation. 4. **Proposed detection and fix** — the recognizable pattern, suggested transformation or guidance, and cases where a fix would be ambiguous. 5. **Current MCP behavior** — the actual relevant output, or explicitly "not checked" with the reason. 6. **Environment** — Svelte, SvelteKit, MCP, and model versions when known; do not invent missing versions. The report is ready when a maintainer can understand the pattern and expected outcome without access to the original project. ## 3. Open the issue Use GitHub CLI when it is installed and authenticated. Check with `gh --version` and `gh auth status`. Always specify `--repo sveltejs/ai-tools` so the issue does not go to the user's project repository. Search for an existing report of the same pattern: ```sh gh issue list --repo sveltejs/ai-tools --state all --search '<pattern keywords>' --limit 20 ``` Read likely matches with `gh issue view <number> --repo sveltejs/ai-tools`. If an issue already describes the same reproduction and expected fix, share its link instead of filing a duplicate. If a closed issue was supposedly fixed but your example still fails, describe that regression and link the old issue in the new report. Write the issue body to a temporary Markdown file using a file-writing tool or a quoted heredoc. Pass the title as a safely quoted argument and use `--body-file` so code snippets and backticks are not interpreted by the shell: ```sh gh issue create --repo sveltejs/ai-tools --title 'Autofixer: detect <incorrect pattern>' --body-file '<temporary-body-file>' ``` Capture and share the returned issue URL. Do not guess labels or assignees. If creation fails, report the failure and use the browser fallback. If the result is uncertain (for example, a network timeout after submission), check for the issue before retrying. ### Browser fallback If `gh` is unavailable, unauthenticated, or unable to create the issue, build a URL for: ```text https://github.com/sveltejs/ai-tools/issues/new?title=<encoded-title>&body=<encoded-body> ``` Encode each query value with `encodeURIComponent`, `URLSearchParams`, or an equivalent URL encoder; do not interpolate raw Markdown into the URL. Open it using an available browser tool or the platform URL opener (`open` on macOS, `xdg-open` on Linux, or `Start-Process` in PowerShell). This opens the new-issue form with the draft filled in for the user to submit. If the full report makes the URL too long, prefill the title and a short summary, and provide the full Markdown body for the user to paste. If no browser can be opened, provide the clickable prefilled URL and the body in the response. Distinguish the outcome: either give the URL of the created/existing issue, or say that the draft form is ready and the user still needs to submit it. Then return to the user's original task.
© sveltejs · MIT · sveltejs/ai-tools