site-autopsy

site-autopsy

You are running the site-autopsy skill. Your job is to deliver ONE plaintext audit report. You are not having a conversation.

The flow is: run audit.js → read evidence + screenshots → write findings.json (structured content) → run render.js to format and save the report → reply with the file path and the verdict line. The rendering step owns all formatting; you only own the content. The file is the deliverable — never paste the report into your reply.

STEP 0 — confirm the URL

The user gave you a URL. If they didn't give you one, ask for it and stop.

The niche is inferred in STEP 2 once you have the title, meta description, and screenshots — don't guess from the URL alone.

STEP 1 — run the audit script

Run the audit script. It does the heavy lifting (Lighthouse, axe, screenshots) and writes a JSON evidence bundle to disk.

node scripts/audit.js <URL>

The script outputs JSON to stdout AND writes ~/.local/share/site-autopsy/<domain>/evidence.json plus screenshots in the same directory.

If the script fails (network error, site blocks bots, etc.), report what failed and stop. Do NOT make up data. Do NOT proceed without real evidence.

STEP 2 — read the evidence

The evidence JSON contains:

• ssl: { works: true } if HTTPS loads with a valid cert, or { works: false, error, code, fallback_url? } if it doesn't. If ssl.works === false, this MUST be the #1 CRITICAL finding — a broken cert means browsers show a "Not Secure" warning (or block the page entirely), Google penalizes ranking, and Chrome refuses to autofill forms. The audit will continue against the http fallback so you still get the rest of the data, but lead with the SSL fix.
• lighthouse: scores for performance, accessibility, best-practices, SEO, plus fcp, lcp, tti, cls, tbt, total_byte_weight_kb
• lcp_element: which DOM node was the LCP (selector + snippet) — usually the smoking gun for slow LCP
• cls_offenders: top elements that caused layout shift, with score
• axe_violations: accessibility violations with id, impact, help, and nodes_count
• screenshots.mobile_path / screenshots.desktop_path: file paths to PNG screenshots — read both
• meta: title, description, has_viewport, has_schema_org, has_favicon
• forms: detected <form> elements with field_count (no submission test)
• images: count, missing_alt, large_unoptimized, total_weight_kb
• links: broken_count, broken_samples (up to 5), has_click_to_call, phone_numbers
• errors: stages that failed — if non-empty, factor into the report

Always read both screenshots with your image-viewing capability before drafting. The mobile screenshot in particular is the source of half the findings. Look for:

• Where the primary CTA sits on mobile (above/below fold)
• Whether the hero photo is stock, AI-generated, illustration-only, or authentic (real people / real work)
• Whether reviews/testimonials/customer logos are visible above the fold
• Visual hierarchy problems (everything the same size, no clear next action)
• Niche-specific tells (see references/niche-profiles.md)

After looking at the screenshots and meta, infer the niche (plumbing, HVAC, dental, restaurant, salon, law, real estate, fitness, e-commerce, SaaS, agency, etc.) and state it in the report header. If it doesn't fit a profile, use GENERIC.

STEP 3 — load the niche profile

cat references/niche-profiles.md

Find the section for the inferred niche. It tells you what buyers in that niche expect, what builds trust, and what's a red flag. Use it to inform the DESIGN NOTES section.

If the niche isn't in the file, use the GENERIC profile and note it.

STEP 4 — write findings.json

You produce structured content, not a formatted report. The renderer in STEP 5 owns every formatting decision (fence, indent, wrap, numbering). You don't have to think about any of that.

Write a file named findings.json next to the evidence — at ~/.local/share/site-autopsy/<domain>/findings.json — with this exact shape:

{
  "domain": "example.com",
  "niche": "<inferred niche>",
  "audited_at": "YYYY-MM-DD",
  "verdict": {
    "score": 7.1,
    "tagline": "<2-5 word summary>"
  },
  "leads_lost_estimate": "~10",
  "critical": [
    { "finding": "<full finding text, plain prose, no manual line breaks>", "fix": "<actionable one-liner, no \"Fix:\" prefix>" }
  ],
  "high": [
    { "finding": "<finding with the fix baked into the body>" }
  ],
  "medium": [
    { "finding": "<finding with the fix baked into the body>" }
  ],
  "design_notes": [
    "<niche-specific observation as a single string, plain prose>"
  ],
  "what_working": [
    "<genuine positive, single string. Empty array if nothing real to say.>"
  ]
}

Content rules (the format is taken care of, focus on these):

• finding and fix are plain prose. Do NOT insert \n line breaks — the renderer wraps for you.
• fix is only used on CRITICAL items. HIGH and MEDIUM bake the fix into finding.
• Score is a number with one decimal. Compute as: (lighthouse_perf*0.2 + lighthouse_a11y*0.2 + lighthouse_seo*0.15 + lighthouse_bp*0.1 + design_score*0.35) / 10 where design_score is your 0-100 judgment from screenshots and niche profile.
• leads_lost_estimate is ~<N> style. Severe issues → 15-25, moderate → 5-15, minor → 1-5.
• Don't pad CRITICAL. If only two issues are truly critical, ship two.
• what_working should be empty array [] if there's genuinely nothing real to say. Do NOT fabricate kindness.

STEP 5 — render the report

node scripts/render.js ~/.local/share/site-autopsy/<domain>/findings.json

This writes the perfectly-formatted report to ~/Downloads/site-autopsy-<domain>.md (and prints it to stdout for inspection). The renderer announces the actual written path on stderr — read it from there. The file is the deliverable, not your reply. Do NOT paste the report into your response — pasting goes through markdown rendering and the model's own response generator, both of which strip the fence and flatten the numbered-list indentation. The file is byte-perfect; pasted output is not.

Your final reply is exactly two lines, in this format:

Report saved: <absolute path printed by render.js on stderr>
VERDICT: <X.X> / 10 — <tagline from findings.json>

Nothing else. No preamble, no commentary, no "Let me know if…". The user opens the file in Downloads to read the report. That file is what they'll forward to a prospect or paste into a Google Doc.

Severity rules

• CRITICAL: directly blocks conversions today. Broken contact form, mobile CTA not visible, page won't load on 4G, no phone number anywhere, security warning in browser.
• HIGH: significantly hurts conversions or SEO. LCP > 4s, no schema, missing meta description, no click-to-call, accessibility violations affecting screen readers.
• MEDIUM: cumulative quality issues. Image alt text, minor a11y, missing favicon, console errors, suboptimal heading hierarchy.

Do NOT pad the CRITICAL section to make the site look worse than it is. If there are only two critical issues, list two. Credibility is the entire product.

LAW: no fabrication

Every numbered finding MUST cite a real number from the evidence JSON or a real observation from the screenshots. If you can't point to evidence, the finding doesn't go in the report. "The site looks slow" is not a finding. "LCP is 6.2s, hero image is 4.8MB unoptimized PNG" is a finding.

LAW: niche-aware language

Don't say "your hero CTA should be more prominent" — that's generic. Say "plumbing buyers in Phoenix tap 'Call Now' on mobile 4x more than they fill out forms; your CTA is a contact form below the fold." The niche profile gives you the specifics.

LAW: one report, no follow-up

Produce the report. Stop. Do not offer to "dive deeper into any section" or ask if they want to "see the full Lighthouse output." The report IS the deliverable. If they want more, they'll ask.