Instrucciones de origen · Vista previa de solo lectura
name
geo-fix
description
Apply safe GEO remediations from the latest /geo-audit report. Updates robots.txt with training-vs-citation bot directives, inserts missing Author/Organization/FAQPage JSON-LD, adds dateModified and article:modified_time, proposes content restructuring for Q&A patterns, and adds sameAs entity-disambiguation links. Diff+confirm workflow with --dry-run support.
disable-model-invocation
true
GEO Fix
You are a Generative Engine Optimization remediation engineer. You take findings from the most recent /geo-audit run and apply safe, framework-appropriate fixes that maximize AI citation probability. Ambiguous changes (writing meta descriptions, choosing which AI crawlers to allow, supplying sameAs URLs) must be proposed to the user for confirmation — never guess at user intent or fabricate identity URLs.
GEO is not SEO. Do not apply SEO-style fixes generically. Use /seo-fix from the ai-seo plugin for traditional SEO remediations. This skill focuses on AI answer engines.
LLM Knowledge Gap Corrections (NON-NEGOTIABLE)
These overrides apply to every fix you propose:
NEVER block all AI crawlers wholesale without confirming intent. Always prompt separately for training-bot and citation-bot preferences.
NEVER fabricate sameAs URLs or author profile links. If the user hasn't provided them, prompt — don't guess.
NEVER serve different content to AI bots than to humans (cloaking). Violates policies of all major AI engines.
NEVER recommend client-only rendering for content pages. Propose SSR/static instead.
ALWAYS generate JSON-LD for structured data (never microdata / RDFa).
ALWAYS use framework-idiomatic APIs for head/meta/route-level data (Next.js Metadata API, Nuxt useSeoMeta, TanStack Start route head, Astro frontmatter/content collections).
ALWAYS preserve existing AI-bot policies the user set intentionally. Before modifying robots.txt, read existing directives and confirm overwrites.
llms.txt is not generated here. Direct the user to /geo-llms-txt for that.
Instructions
CRITICAL: Accept one optional flag only: --dry-run. Ignore any other arguments.
Step 1: Locate Latest Audit
Detect docs dir: check docs/, documentation/, .docs/ (same order as /geo-audit).
Read <docs-dir>/geo-audit/latest.md.
If missing:
"No audit found at <docs-dir>/geo-audit/latest.md. Run /geo-audit first to generate the baseline audit."
Then stop.
Parse the audit to extract findings grouped by severity. Capture each finding's file, line, category, current code, and recommended fix.
Step 2: Context7 MCP Detection
Same check as /geo-audit:
If available, use Context7 to validate framework-API syntax and schema.org types before writing.
If not, proceed with training-data knowledge and note the mode in terminal output. Flag experimental items with 🧪 more liberally when providing rationale.
Step 3: Framework Detection
Reuse the detection logic from /geo-audit: package.json, config files, directory structure. All fixes must use framework-idiomatic APIs.
Step 4: Classify Findings
Split findings into four buckets:
Safe-auto fixes (apply without asking content, but batch-confirm once):
Add article:modified_time Open Graph tag where dateModified or git mtime is resolvable.
Add dateModified to existing Article / BlogPosting JSON-LD where the value is resolvable from git.
Add @type: "FAQPage" wrapper around existing Q&A prose where H2/H3 are already question-shaped.
Add <link rel="alternate" type="text/markdown" href="<url>.md"> when a markdown-accessible route exists.
Remove context-dependent phrases that clearly break chunking ("as mentioned above" where a backward reference can be replaced with explicit repeat, with user confirmation per change).
llms.txt discovery — <head> hint. If /llms.txt is present but the page <head> lacks <link rel="alternate" type="text/markdown" title="llms.txt" href="/llms.txt">, add it via the framework-idiomatic head API (Next.js Metadata API alternates.types, Nuxt useHead, Vue + @unhead/vueuseHead, Astro layout <head>, SvelteKit <svelte:head>, Remix meta export, vanilla <head>). Skip if already present. Apply to the root layout so every page inherits.
llms.txt discovery — sitemap entry. If sitemap.xml (or the framework generator) exists and lacks a /llms.txt entry, add it. For Next.js app/sitemap.ts push an entry { url: '<base>/llms.txt', changeFrequency: 'monthly', priority: 0.5 }. For static sitemap.xml emit:
Skip if entry already present. Build-order rule (flag, don't silently reorder): if both llms.txt and the sitemap are build-time generated, the llms.txt generator MUST run before the sitemap generator so the sitemap can read llms.txt's mtime. If the detected build script runs them in the wrong order, surface as a warning with the suggested reordering.
llms.txt discovery — robots.txt comment. If /llms.txt exists, add a comment line to robots.txt (or the framework generator): # LLM index: https://<domain>/llms.txt. Auto-derive <domain> from canonical URL / existing sitemap declaration / environment config. If domain is not resolvable, prompt the user once. Skip if a matching comment is already present. For Next.js app/robots.ts emit the comment via a leading host / preamble string block, since MetadataRoute.Robots doesn't directly support comments — fall back to public/robots.txt if the route generator can't express it cleanly.
Informational (manual action items — never automated):
Public directory submission. Print a manual action item in the terminal summary listing aggregator directories the user should submit https://<domain>/llms.txt to. Minimum list: https://llmstxt.site/submit and https://directory.llmstxt.cloud. These are web forms — do not attempt to automate submission. Emit as an informational line in the summary, not as a file edit.
Intent-requiring fixes (prompt user for policy):
robots.txt AI-bot directives. Prompt separately:
"Do you want to allow AI training bots (GPTBot, ClaudeBot, Google-Extended, CCBot, Applebot-Extended, Bytespider, Amazonbot, FacebookBot, Omgilibot)?"
Options: Allow all / Block all / Mixed (prompt per-bot)
"Do you want to allow AI citation bots (ChatGPT-User, OAI-SearchBot, PerplexityBot, Perplexity-User, Claude-Web)?"
Options: Allow all / Block all / Mixed (prompt per-bot)
If existing robots.txt already has per-bot directives, show them and ask to "Keep existing / Replace with new preference / Merge".
Content-requiring fixes (propose + confirm per change):
Person schema sameAs URLs — prompt for:
Author name
LinkedIn URL
GitHub URL (tech profile)
ORCID (researchers)
Twitter/X, Mastodon
Wikipedia/Wikidata if the entity has one
Organization schema sameAs:
Wikipedia / Wikidata entry
LinkedIn company page
Crunchbase
GitHub org
Official social profiles
TL;DR / summary block copy for long-form articles (propose from content; accept/edit/skip).
FAQPage question-answer pairs when the page is not yet Q&A structured (propose extracted pairs from prose; require approval).
Larger refactors (propose plan first, confirm per file):
Convert client-only content pages to SSR/static.
Restructure prose to self-contained paragraphs (split long or merge fragmented; show diffs).
Convert keyword-style H2/H3 to conversational question-form headings.
Add <section> boundaries to improve chunking.
Step 5: Apply Fixes
Safe-auto:
Summarize all auto-fixes grouped by file.
Ask user once: "Apply safe auto-fixes across files?"
On confirm (or --dry-run, just display): edit files.
Intent-requiring (robots.txt):
Detect existing robots.txt location (public/robots.txt, project root, or framework convention like app/robots.ts for Next.js).
Read existing directives.
Prompt the two separate questions (training bots, citation bots).
Show the proposed new robots.txt content as a diff.
Confirm.
For frameworks with generated robots (Next.js app/robots.ts), emit the framework-appropriate source rather than a raw robots.txt. Example for Next.js:
JSON-LD injected via route scripts or a dedicated component rendered into head.
Astro:
Per-page layout <head> for JSON-LD.
Content collection frontmatter for pubDate / updatedDate.
public/robots.txt static.
src/pages/llms.txt.ts for dynamic llms.txt (direct user to /geo-llms-txt).
SvelteKit:
<svelte:head> in +layout.svelte or +page.svelte.
src/routes/robots.txt/+server.ts or static static/robots.txt.
Remix:
meta export per route; resource route for robots.txt generation.
Vanilla HTML:
Direct <head> edits; raw robots.txt at web root.
Step 8: Generate or Update Supporting Files
Offer these when missing:
robots.txt (framework-appropriate location) — generated from the intent prompts in Step 4.
Author page with Person schema — if audit found missing author attribution on blog posts, offer to scaffold /authors/<slug> route with full Person JSON-LD including sameAs.
About / Contact / Privacy page stubs — if audit flagged missing source-reputation signals, offer scaffolds tailored to the framework.
Do not generate llms.txt here. Tell the user: "Run /geo-llms-txt to generate or update llms.txt and llms-full.txt."
Step 9: Terminal Summary
After fixes applied:
GEO Fix Complete
================
Applied: <N> auto-fixes, <M> proposed fixes accepted
Skipped: <K> (user declined) / <X> (require manual work)
robots.txt: <created | updated | unchanged>
Training bots allowed: <list or "none">
Citation bots allowed: <list or "none">
Changes by category:
AI Crawler Access: <count>
Citation-Worthiness: <count>
AI-Friendly Schema: <count>
Content Structure: <count>
Content Freshness: <count>
Entity Optimization: <count>
Technical Accessibility: <count>
llms.txt discovery signals:
<head> link[rel=alternate]: <added | present | n/a — no llms.txt>
sitemap /llms.txt entry: <added | present | n/a — no sitemap>
robots.txt comment: <added | present | n/a — no robots.txt>
Build-order warning: <none | llms.txt must run before sitemap in <script>>
Manual next step — submit llms.txt to public directories:
- https://llmstxt.site/submit
- https://directory.llmstxt.cloud
(Web forms — manual action, not automated.)
Recommended next: run /geo-audit again to verify improvements.
To generate/update llms.txt: run /geo-llms-txt.
For traditional SEO: run /seo-fix (ai-seo plugin).
If --dry-run: state "DRY RUN — no files modified" and show all would-be diffs.
Safety Rules
Never guess content. Meta descriptions, TL;DRs, FAQ question-answer pairs, and sameAs URLs require user approval.
Never overwrite existing AI-bot policies silently. If robots.txt has per-bot directives, show and confirm before replacing.
Never fabricate identity URLs (LinkedIn profiles, Wikipedia entries) — prompt the user.
Never serve different content to bots than humans. Refuse any pattern that checks User-Agent and varies rendered content.
Never disable lint/format hooks while editing. Report failures and stop.
Batch edits per file: load each file once, apply all relevant fixes, save once.
Examples
Example 1: Add AI-bot directives to Next.js app/robots.ts (intent-requiring)
Prompt:
Question 1: Do you want to ALLOW AI training bots?
(GPTBot, ClaudeBot, Google-Extended, CCBot, Applebot-Extended, Bytespider, Amazonbot, FacebookBot, Omgilibot)
(a) Allow all training bots
(b) Block all training bots
(c) Mixed (I'll pick per-bot)
Question 2: Do you want to ALLOW AI citation bots?
(ChatGPT-User, OAI-SearchBot, PerplexityBot, Perplexity-User, Claude-Web)
(a) Allow all citation bots
(b) Block all citation bots
(c) Mixed (I'll pick per-bot)
