AI improvement prompts

How it works

Two pieces of data combine on click:

• The finding — title, description, evidence, remediation, CWE — already loaded with the report.
• Your codebase's framework — detected from the scan's discovery.tech-fingerprint findings (Next.js, React, Vue, Django, Express, Rails, Laravel, Flask). Falls back to a generic recipe when no framework is detected (the agent prompt then asks the LLM to detect it from the repo state).

Templates live in lib/scanner/fix-prompts.ts. The registry has check-specific guidance for crawlability, search presentation, semantic content, structured data, media, performance, accessibility, forms, mobile/i18n, runtime, owner journeys, and repo-quality issues. For everything else, the existing remediation field on the finding becomes the generic recipe.

What the prompt looks like

Fix the "Hero image is lazy-loaded and missing dimensions" finding on /pricing.

Issue: The largest above-the-fold image is marked loading="lazy" and has
no explicit width/height. That can delay LCP and create layout shift.

Codebase context: Next.js.

Recommended fix:
Use next/image or the existing image component with explicit width/height,
responsive sizes, meaningful alt text, and priority/fetchPriority for the
first major visual on the page.

Constraints:
- Don't break existing tests; run the test suite after the change.
- Match the codebase's existing style and lint config.
- Add a brief comment explaining the performance reasoning only where the
  fix would otherwise look arbitrary.
- If the fix needs a new dependency, install it via the project's
  package manager (npm / pnpm / pip / bundle / composer).

Reference: Core Web Vitals / Largest Contentful Paint guidance.

Supported frameworks

We surface framework-specific snippets for:

• Next.js, React, Vue, Nuxt, Svelte (frontend)
• Express, Fastify (Node.js backend)
• Django, Flask (Python)
• Ruby on Rails
• Laravel (PHP)
• ASP.NET Core (planned, fallback to generic today)

Framework detection is best-effort. We sniff __NEXT_DATA__ tags, __NUXT__, hash cookies (laravel_session), X-Powered-By headers, and a few other signals. If you're running a custom framework, the prompt falls back to the generic recipe and the agent figures it out from your package.json.

Use it from your AI agent

If you've wired up the MCP server, the same prompt is exposed as a slash command. From Claude Desktop:

/fixweb-fix finding_id=550e8400-e29b-41d4-a716-446655440000

The renderer looks up the finding, detects the framework from the parent scan when available, renders the templated prompt, and injects it into your conversation as the user message. No round-trip to our Claude API; templates are pure and free.

Why we don't hit Claude per click

On launch we considered calling the Anthropic API for each click to refine the prompt with codebase context. We didn't, because:

• The agent the user pastes into already has codebase context — they're using Cursor / Claude Desktop with their repo open.
• Templating per-(check × framework) covers ~80% of the value without any per-click cost.
• A “Refine with AI for my codebase” opt-in could fire the API later if users want it. Today, no.