Fixing issues with fix prompts

Every check in the AI Site Scanner comes with a fix prompt: a short instruction written for an AI coding assistant that says what to add or change. To fix an issue, copy its prompt, paste it into Claude, ChatGPT, Cursor, Codex or a similar assistant that can see your site's code, review the change it proposes, deploy it, and scan again to confirm the check now passes.

Where do I find the fix prompts?

Where	What you get
Scanner page and stored report	A Copy fix prompt button on each of the up to six prioritized issues
Command line with --fix	One combined prompt covering every failed check, ready to paste or pipe
MCP server scan_site tool	Every failed check with what was found and its fix, returned to your assistant
HTTP endpoint	A fixPrompt field on every entry in checks, passed or failed

On the web the list stops at six, critical first. If a site fails more checks than that, fix the top ones and re-scan, or run the CLI with --fix to get all of them in one go.

What does a fix prompt look like?

It names the file or tag, where it goes and what it should contain. The prompt for the llms-txt check, for example:

Create a /llms.txt for my site per https://llmstxt.org: '# Site Name' H1, a '> one-line summary' blockquote, then ## sections of markdown links to my most important pages with one-line descriptions.

The prompts are the same for every site. They do not include your content, so the assistant still needs access to your code or CMS, and you still need to check that what it writes is true for your site. The issue card shows what the scanner found (for example "Blocked at root: GPTBot, ClaudeBot"), which is useful context to paste along with the prompt.

The CLI's --fix output adds that context for you: it starts with the URL, score and grade, then lists each failed check with what was found, the fix and a reference link, and ends with the command to re-check.

Which issues should I fix first?

The heaviest ones, because they move the score most. Issues are already sorted by severity, which follows weight:

1. Critical (weight 7 or 8): HTTPS, llms.txt, JSON-LD, server-rendered content, AI bot rules. Each one is worth 7 or 8 points of weight on its own.
2. Warning (weight 4 to 6): robots.txt, the sitemap, HTTP status, time to first byte, Markdown negotiation, semantic HTML, title and OpenGraph, noindex, API catalog.
3. Info (weight 2 or 3): the rest, mostly new protocols that few sites publish yet.

Some fixes are quick with a Toolz.dev tool: the robots.txt generator for crawler rules, the schema markup generator for JSON-LD, the sitemap generator for sitemap.xml, and the meta tag generator for the title, description and OpenGraph tags. The server-rendered content check usually needs a change to how the site is built, such as server-side rendering or static generation.

How do I confirm a fix worked?

Deploy the change, then scan the same URL again with the same mode and the same site-type filter. A different filter changes which checks run, and so the score. Each scan is a new report; the old one is kept.

If a check still fails after a deploy, look at the detail line first. A cache in front of the site (a CDN, a hosting cache, or a static file that was not redeployed) is the usual reason the scanner still sees the old version.

How do I see the improvement over time?

Open any stored report of the domain. Its Earlier scans list shows the domain's other public scans with their date, score and grade, each linked to its report, so you can compare before and after side by side. Reports, sharing and history has more on stored reports.

When a domain's first and latest public scans show a higher score, the domain can appear in Improved after a scan on the scanner page: a table of sites that "scanned publicly, fixed what the report flagged, and scanned again", with the two scores, the gain and links to the before and after reports. It lists up to eight domains, largest gain first. Only public scans count, and nothing in the table claims who made the fix.

Can I fail a build when the score drops?

Yes, with the command line: npx @toolz-dev/aiscan example.com --min-score 85 exits with code 1 when the score is below 85. Command line (npm package) has a GitHub Actions example.