AutoFix

Status: built, and we need your test

The skill is written and its code ran against two test apps with planted bugs (details below). One thing is not verified: installing the skill on a clean machine and asking Claude for it in plain words. On the machine where it was built there was already a personal skill that answers to the same word, autofix, and the two collided, so Claude could not be trusted to pick this one. Install it, try it on your app, and post in the community what happened: which stack (Node, Next.js, Express, Python, Flask, FastAPI, Django), what worked, and what looked wrong. The next version comes from those reports.

Where this comes from

Most apps students build with Claude Code fail quietly. A page breaks, a payment call returns HTML instead of JSON, a background job dies at 3 a.m., and nobody knows until a customer says so. Professional teams use Sentry or Datadog for this. AutoFix is the small version for your own apps: no account, no dashboard, one file in your project, and the alert lands on your phone with an explanation a beginner can act on. It was built by Hector Diaz for the AI Auto Base community, with Claude Code.

What you build

You install the skill once. Inside any project you say "ponle autofix a mi app" (or "add autofix"). Claude reads the project first, copies one file next to your .env, adds one or two lines where the app starts, asks three questions one at a time (which AI, where the alerts go, Spanish or English), walks you through each key, and sends one real test alert that you confirm arrived.

From then on, every error is written to a daily log kept 30 days, with anything that looks like a key or password masked. Each new error gets a plain-language explanation and a report file. The same error alerts once a day at most, and never more than 10 alerts an hour, so a broken app cannot flood your phone or burn your AI credit.

When an alert comes in, open the project in Claude Code and type autofix. Claude lists the open reports, confirms the real cause in your code (the AI note is a hint, not the truth), makes the smallest fix, proves it, and marks the report fixed with what changed.

After each fix Claude also writes a short lesson in .claude/autofix-lessons.md: what failed and what to do instead. A block in your project's CLAUDE.md makes Claude read it at the start of every session, so it does not repeat the mistake. It also notes mistakes Claude makes while coding in that project, or when you correct it. The list stays short (at most 10 per section) and never holds log text, keys or passwords, because it is committed to git with your project.

What it deliberately does not do: it never changes your code on its own. Alerts and reports happen automatically; fixes happen only when you type autofix.

Recommended setup for an app the public will use

• AI: OpenRouter. One key with a spending limit set on that key and prepaid credit, so when the credit runs out it stops instead of billing more. You can switch models without touching code. The default model is Claude Haiku 4.5 through OpenRouter, a fraction of a cent per new error. For an app that only you use, a local Ollama is free.
• Server: a Hostinger VPS. AutoFix keeps the 30-day log on the server's disk. On Vercel or Netlify functions that disk is wiped on every restart, so the alerts work but the history is lost. A KVM 2 plan runs the app and AutoFix with room to spare: hostinger.com. This is an affiliate link; if you buy through it, AI Auto Base may earn a commission at no extra cost to you.

Do you need this? Be honest first

This is for you if you built a Node or Python app with Claude Code, it runs somewhere (your laptop or a VPS), and you want to hear about its errors before your users do.

Close the tab if your site is plain HTML with no server (GitHub Pages, a static Netlify site): there is nowhere safe to keep the AI key, and AutoFix will tell you so. PHP, Go and Ruby are not supported yet. A company with an on-call team should use a real monitoring service.

What you need

• The machine. Any Windows, Mac or Linux computer with Claude Code, plus Node 18 or newer or Python 3.8 or newer, whichever your app uses.
• The accounts. Claude Code. One AI: OpenRouter, Anthropic or OpenAI with a few dollars of credit, or Ollama for free. One alert channel: Telegram is free and takes about three minutes. WhatsApp alerts need your own Evolution API server.
• The time. Two minutes to install the skill. Ten to twenty minutes per app, most of it creating the key and the Telegram bot.

What is in the zip

autofix/
  SKILL.md           the instructions Claude follows: install, fix mode, remove mode
  GUIA.md            the full Spanish guide: what it catches, what it does with each error, every setting, troubleshooting
  SETUP.md           how to get each key and token, step by step, in Spanish
  README.md          install in two minutes, how to use it, costs, privacy
  CREDITS.md         who built it, what was verified
  DISCLAIMER.md      as is, no warranty, use at your own risk
  templates/
    autofix.py       Python 3.8+, standard library only
    autofix.cjs      Node 18+, no dependencies (email uses nodemailer if you add it)
    autofix-browser.js  the web page snippet; holds no keys
    env.example      every setting, with comments
    autofix-lessons.md  the lessons file template
    claude-md-block.md  the CLAUDE.md block that makes Claude read the lessons
    n8n-autofix-hub.json  the ready n8n workflow

The same files are on GitHub: AIAutoBase/autofix-skill.

What was verified and what was not

Verified with real sends on Windows 11, Node 26 and Python 3.14: all four AI providers (Claude Haiku 5.5 on Anthropic, gpt-5-mini on OpenAI, Claude Haiku 4.5 on OpenRouter, a local Ollama), and three channels (Telegram, WhatsApp through Evolution API, email over SMTP). Two test apps were built with planted bugs, a Next.js store and a Flask booking system, and two separate Claude sessions that knew only the skill played the student and installed it. They found seven problems, all fixed. A real browser then clicked through both apps: all five planted bugs, server and browser side, arrived explained in Spanish.

New in 0.3.0: an n8n workflow (templates/n8n-autofix-hub.json). Import it, add one OpenRouter credential, enable the channels you want (email with the .md report attached, Telegram, WhatsApp, Slack, Discord, SMS, a phone call that reads you the summary, ntfy push to your phone, outgoing webhook) and point the app at its webhook URL. n8n does the frontier-model analysis and the sending; the app only reports. Verified with real sends: email, Telegram, WhatsApp and ntfy, with and without AI.

New in 0.4.0: lessons per project, adapted from napkin (MIT). Verified with Claude Sonnet 5.5 on a Python project with a planted bug: the file and block are created on install, a fix writes one lesson, the same cause in another file updates that lesson instead of adding a second, a password hidden in the error never reaches the file, a fresh session applied the lesson without being asked, reinstalling keeps saved lessons, and removal asks before deleting them.

Not verified: installing the skill on a clean machine and triggering it by voice (see the top of this page), Slack, Discord, Make/Zapier, email from a Node app, Django and FastAPI; inside the n8n workflow: Slack, Discord, SMS and the phone call. Those are where your reports help most.

Read the skill before you install it

Open SKILL.md and the template for your language before you copy the folder into your skills. Hand them to Claude and ask what they will change in your project and what they send out. The answer should be: one file and one or two lines in your app, an AUTOFIX_ block in your .env, and outbound calls only to the AI and the channels you chose.