---
title: "LLM Content Audit: Review and Refresh Pages at Scale"
description: "Run an LLM content audit with a validated classification rubric, source checks, human review, and clear rules for refreshing, merging, or pruning pages."
canonical: "https://nikoalho.fi/writing/llm-content-auditing/"
language: "en"
---

> Canonical source: [https://nikoalho.fi/writing/llm-content-auditing/](https://nikoalho.fi/writing/llm-content-auditing/)

[← writing](https://nikoalho.fi/writing/)

Agentic SEO Published 2026 · 05 · 20 Updated 2026 · 07 · 18

# LLM content auditing: kill low-value pages at scale

An LLM content audit reviews 10,000 legacy posts in hours, not weeks. The classification rubric, prompts, and prune-vs-refresh logic that frees real budget.

![Niko Alho](https://nikoalho.fi/assets/niko-alho-avatar-96.webp)

**Niko Alho**Operator in Turku · firsthand systems

ON THIS PAGE

[01 The Cost of Digital Hoarding: Why You Must Prune](#the-cost-of-digital-hoarding-why-you-must-prune) [02 The prune decision tree](#the-prune-decision-tree) [03 Framework for LLM-Based Content Quality Scoring](#framework-for-llm-based-content-quality-scoring) [04 Automating the ‘Kill, Update, or Keep’ Decision](#automating-the-kill-update-or-keep-decision) [05 The content scoring rubric](#the-content-scoring-rubric) [06 The Architect’s Blueprint: Prompt Engineering for Audits](#the-architects-blueprint-prompt-engineering-for-audits) [07 Execution: Moving from CSV to CMS](#execution-moving-from-csv-to-cms) [08 Estimate your index bloat](#estimate-your-index-bloat) [09 Criteria for Content Auditing Systems](#criteria-for-content-auditing-systems) [10 Conclusion: Clean the Engine](#conclusion-clean-the-engine)

PROGRESS

0%

ON THIS PAGE 10 sections

[01 The Cost of Digital Hoarding: Why You Must Prune](#the-cost-of-digital-hoarding-why-you-must-prune) [02 The prune decision tree](#the-prune-decision-tree) [03 Framework for LLM-Based Content Quality Scoring](#framework-for-llm-based-content-quality-scoring) [04 Automating the ‘Kill, Update, or Keep’ Decision](#automating-the-kill-update-or-keep-decision) [05 The content scoring rubric](#the-content-scoring-rubric) [06 The Architect’s Blueprint: Prompt Engineering for Audits](#the-architects-blueprint-prompt-engineering-for-audits) [07 Execution: Moving from CSV to CMS](#execution-moving-from-csv-to-cms) [08 Estimate your index bloat](#estimate-your-index-bloat) [09 Criteria for Content Auditing Systems](#criteria-for-content-auditing-systems) [10 Conclusion: Clean the Engine](#conclusion-clean-the-engine)

**TL;DR** The useful bits

-   8-min read
-   4 takeaways

1.  01 Manual review of 10,000 legacy posts burns cash and produces inconsistent verdicts — humans don't scale to enterprise index sizes.
2.  02 Index bloat is a technical liability: every millisecond Googlebot spends on a 2019 thin post is a millisecond not spent on your revenue pages.
3.  03 Don't ask an LLM 'is this good?' — architect a multi-step scoring agent that evaluates information gain, temporal relevance, and intent alignment.
4.  04 Pruning is gardening: cutting dead branches lifts site-wide quality signals and unblocks the money pages being suppressed by the Helpful Content System.

A/01 Direct answer

How do you audit content at scale with LLMs?

Architect a multi-step scoring agent — not a single 'is this good?' prompt. The agent evaluates each page on information gain, temporal relevance, intent alignment, internal-link value, and conversion contribution, then categorizes the page as keep, refresh, merge, or delete with reasoning attached.

Manual content audits are a relic of the past. If you are paying humans to manually review 10,000 legacy blog posts, you are burning cash. Most large-scale SaaS sites suffer from severe **Index Bloat** —carrying thousands of zombie pages that dilute topical authority and waste crawl budget. This guide details how to architect an **LLM content auditing** system that autonomously scores, categorizes, and flags pages for deletion with higher consistency than a human editor.

* * *

## The Cost of Digital Hoarding: Why You Must Prune

In March 2026, the concept of “more is better” is the fastest way to kill your organic visibility. For years, agencies sold the lie that publishing volume equals growth. The result? Enterprise sites are now drowning in technical debt. **Index Bloat** is not a content problem; it is a technical liability. [Google’s crawl budget for your domain is finite](https://nikoalho.fi/writing/programmatic-seo-architecture/). Every millisecond a bot spends crawling a low-value, 400-word blog post from 2019 is a millisecond *not* spent indexing your new, revenue-generating product pages. This is a zero-sum game. The algorithms have evolved to penalize site-wide quality signals. The integration of the Helpful Content System into the core algorithm means that a high ratio of unhelpful URLs acts as an anchor on your entire domain. If 40% of your site is thin, the other 60%—your money pages—are being suppressed. We are not talking about “optimizing” these pages. We are talking about a **content pruning strategy** that treats your sitemap like a garden: if a branch is dead, you cut it off to save the tree. You must approach this with surgical precision, viewing pruning as **automating technical debt removal**. If a page does not drive revenue or assist in a conversion, it has no business existing on your domain.

* * *

## The prune decision tree

01

Visual model

Agentic LLM Content Audit Pipeline

LLM Auditor Reads XML  
Parses 10,000 URLs

Evaluates Content Quality (1-10)  
& Checks GSC Traffic Data

Score > 8 | Traffic > 0

KEEP

Do nothing. Monitor.

Score < 5 | Traffic> 0

UPDATE

Send to Writer Agent  
to rewrite.

Score < 4 | Traffic=0

KILL (410)

Delete via API.  
Prune from XML.

## Framework for LLM-Based Content Quality Scoring

Traditional content audits fail because they rely on vanity metrics. While modern crawlers like Screaming Frog now offer native AI integration for qualitative analysis, many teams still rely on basic word counts or metadata reviews. A 2,000-word article can be completely useless, while a 300-word glossary definition can be highly valuable. To solve this at scale, we don’t ask ChatGPT, “Is this good?” That prompt is too vague and results in hallucinations. Instead, we architect a multi-step scoring agent using models like **GPT-5** or **Claude 4** , which possess the reasoning capabilities required for nuance. We evaluate content based on specific vectors:

### 1\. Information Gain

Does this URL add unique value to the internet, or is it derivative? If the LLM determines that the content simply regurgitates what is already in the top 10 SERP results, the Information Gain score is 0.

### 2\. Temporal Relevance

Is the information factually obsolete? We don’t just look for old dates; we check for validity. An article from 2021 might be evergreen, while a “2025 Industry Trends” post is now dead weight. The LLM checks for deprecated technology references and expired actionable advice.

### 3\. Intent Alignment

Does the content actually answer the target query? We often see “guides” that are actually thin sales pitches. This mismatch kills engagement metrics. This moves us beyond basic **thin content identification**. We calculate a composite quality score ($QS$) using a weighted formula: $$ QS = (0.4 times InfoGain) + (0.3 times IntentMatch) + (0.3 times Freshness) $$ If the $QS$ falls below a threshold (e.g., 6.0/10), the page is flagged for immediate review. This is **quality content verification** at scale—impossible for humans, trivial for an architected system.

* * *

## Automating the ‘Kill, Update, or Keep’ Decision

Data without action is noise. The purpose of this audit is not to create a spreadsheet; it is to execute a cleanup. We automate the decision-making process by cross-referencing the LLM’s qualitative analysis with quantitative performance data (GSC clicks, impressions, and conversion data). Here is the operational workflow for **automating technical debt removal** :

1.  **Extraction:** A Python script fetches the URL list, extracts the Main Content (stripping navigation and footers), and pulls the last 12 months of Search Console data.
2.  **Analysis:** The LLM agent analyzes the text against our rubric.
3.  **Synthesis:** The system merges the Quality Score with the Traffic Data to output a decision.

### The Logic Tree

The output determines the fate of the URL:

-   **High Quality + High Traffic = KEEP/MONITOR.** The asset is performing. Do not touch it.
-   **High Quality + Low Traffic = UPDATE/RE-PROMOTE.** The content is good, but distribution failed or keyword targeting is off. This enters the optimization queue.
-   **Low Quality + High Traffic = REWRITE IMMEDIATELY.** This is a high-risk category. The page ranks, but the content is bad. It is a ticking time bomb. Rewrite it to match the ranking intent before you lose the position.
-   **Low Quality + Low Traffic = KILL (410 GONE).** This is the sweet spot for pruning. These pages have no traffic, no backlinks, and low quality. Delete them. Serve a 410 status code to signal permanent removal to Google.

* * *

## The content scoring rubric

02

Reference table

| Evaluation Vector | Weight | LLM Prompt Context (The "Judge") |
| --- | --- | --- |
| **Information Gain** | 40% | "Does this page present unique data, original viewpoints, or deep technical architecture not found in the SERP average?" |
| **Intent Alignment** | 30% | "Does the primary entity match the core problem the reader is trying to solve, without verbose fluff?" |
| **Temporal Freshness** | 20% | "Are the technical specs, pricing models, and software integrations still valid for the current year?" |
| **Content Formatting** | 10% | "Is the payload scannable? Does it use JSON-LD structures, clear H2s, and lists rather than walls of text?" |

## The Architect’s Blueprint: Prompt Engineering for Audits

The success of **LLM content auditing** relies entirely on the precision of your prompt engineering. You cannot use “zero-shot” prompting here. You must use “Chain of Thought” reasoning to force the model to justify its score before assigning a number. This reduces variance and hallucination rates significantly. Below is the blueprint for a “Judge-LLM” prompt designed for **quality content verification**.

### The “Judge” Prompt

> **Role:** You are a Senior Technical SEO Editor for a B2B SaaS company. You are critical, harsh, and objective. **Task:** Evaluate the following content text for “Information Gain” and “Helpfulness.” **Input Text:** \[INSERT CONTENT HERE\] **Instructions:**
> 
> 1.  Analyze the text for depth, actionable advice, and unique data.
> 2.  Check for “fluff” sentences that add no value.
> 3.  Check for outdated references (e.g., deprecated software versions or expired trends).
> 4.  Assign a score from 1-10.
>     -   **1-3:** Generic, thin, AI-generated fluff.
>     -   **4-6:** Acceptable but derivative.
>     -   **7-10:** High expert value, unique data, highly actionable.

> **Output Format:** JSON { “score”: \[Integer\], “reasoning”: “\[One sentence justification\]”, “action\_recommendation”: “\[DELETE / UPDATE / KEEP\]” } By standardizing the output into JSON, we can pipe the results directly into a database or a visualization tool. This is how you audit 50,000 pages in an afternoon rather than a year.

* * *

## Execution: Moving from CSV to CMS

The audit is useless until the changes go live. In a manual workflow, a content manager would take the CSV file and manually delete pages in the CMS. This is slow and prone to error. To scale this across 100,000 pages, you cannot rely on manual inputs; you need a **programmatic architecture** that supports bulk operations. This is where we deploy **[Agentic AI](https://nikoalho.fi/writing/agentic-ai-seo/)**. While this guide covers the analysis, the execution phase is where **AI agents** take over. We script agents that can:

1.  Connect to the CMS API.
2.  Read the “Action Recommendation” from our audit database.
3.  If the action is “DELETE,” the agent unpublishes the page and triggers the 410 status protocol.
4.  If the action is “UPDATE,” the agent creates a ticket in the project management system assigned to a writer.

This closes the loop. We move from “insight” to “infrastructure change” without human intervention for the deletion tasks.

### Competitive Differentiation

Most agencies suggest using tools to simply “optimize” everything. They are scared to tell clients to delete 40% of their site because it sounds destructive. We frame deletion as a growth lever. By removing the dead weight, you free up crawl budget and consolidate authority into the pages that actually drive revenue. other strategies rely heavily on *traffic* as the only metric for pruning. Traffic is a lagging indicator. We use **LLM reasoning** to judge quality *before* traffic drops occur, allowing us to be proactive rather than reactive.

* * *

## Estimate your index bloat

03

Working tool

Index Bloat Extractor

Low-quality indexed pages drag down your entire domain's authority. Pruning leads to immediate crawl lift.

Total Valid Indexed Pages (GSC)

Pages with ZERO clicks in last 90 Days

System Pruning Impact

Current Bloat Ratio 45%

URLs to 410 / Delete 4,500

**New Domain Quality Profile** Hyper-Concentrated

## Criteria for Content Auditing Systems

To engineer a reliable LLM auditing system, the model must evaluate five specific data points. If you miss one, the system is flawed.

1.  **Information Gain:** Does the content offer unique data or perspective?
2.  **Topical Relevance:** Is the semantic distance between the content and the core business entity too wide?
3.  **Decay Velocity:** Is the content referencing outdated years or deprecated features? Early [content decay identification](https://nikoalho.fi/writing/content-decay/) prevents high-value assets from silently losing rankings.
4.  **SERP Alignment:** Does the format (e.g., guide vs. listicle) match current search intent?
5.  **Conversion Logic:** Is there a clear path to revenue, or is it a dead end?

* * *

Audit verdicts — what to do with each bucket

+ WORKS WELL

-   **Keep + monitor.** Pages still ranking with current intent and decent depth. Re-evaluate quarterly.
-   **Refresh.** Pages with traffic decay or stale entities. Update stats, inject missing entities, re-align intent.
-   **Merge.** Near-duplicate pages cannibalizing each other. Consolidate into one canonical and 301 redirect.

− WATCH OUT

-   **Delete.** Zero traffic, zero backlinks, irrelevant to revenue topics. 410 it. Don't keep dead weight.
-   **Don't trust 'is this good?' single-shot prompts.** LLMs hallucinate verdicts on single-shot prompts. Build a structured rubric agent.
-   **Don't auto-delete without human approval.** Edge cases (pages with backlinks but no traffic) need review. Gate deletion.

## Conclusion: Clean the Engine

You cannot build a high-performance vehicle on a rusted chassis. If your site is bloated with years of low-quality content, no amount of new backlinks or technical tweaks will save you. **LLM content auditing** is the only viable path to solving this at an enterprise scale. It allows you to process massive datasets with the nuance of a human editor but the speed of a machine. Stop hoarding URLs. Start engineering a lean, efficient growth engine. **Audit your system.** If you don’t know which 40% of your site is burning money, your competitors already have the advantage. Written by ![Niko Alho](https://nikoalho.fi/assets/niko-alho-avatar-192.webp) Niko Alho Technical SEO specialist and AI automation architect. Building systems that drive organic performance through data-driven strategies and agentic AI. [Connect on LinkedIn →](https://www.linkedin.com/in/nikoalho/) Related Articles

When the audit surfaces content that needs to be grounded in proprietary data — internal benchmarks, customer interviews, product docs — the next step is a [RAG system for B2B](https://nikoalho.fi/writing/rag-systems-b2b/), not another generic AI rewrite.

For a hands-off audit run as a service rather than as a DIY pipeline, see [agentic SEO services](https://nikoalho.fi/services/agentic-seo/).

WANT THIS AUDIT PIPELINE BUILT?

I run LLM-driven audits on enterprise content libraries that ship a triaged Jira board, not a verdict spreadsheet.

[Book a 20-min intro →](https://nikoalho.fi/book/)

Questions people actually ask

FAQ · 4

Q01 Won't deleting content hurt my SEO? +

The opposite, usually. The Helpful Content System penalizes site-wide quality signals when thin posts drag down average. Pruning often lifts rankings on remaining pages.

Q02 What LLM works best for content auditing? +

Claude or GPT-5 for reasoning depth, with structured output (JSON schemas) for downstream automation. Avoid temperature > 0.3 for audit verdicts.

Q03 How do I prevent the LLM from being too aggressive? +

Set conservative thresholds: only delete if the page has zero traffic in 12 months, zero backlinks, and doesn't match any cluster. Everything ambiguous goes to refresh, not delete.

Q04 Can I trust an LLM to grade my content without human review? +

Not for delete decisions. The agent's job is to surface candidates with reasoning; the human still approves destructive actions. Auto-apply only the safe verdicts (e.g. add internal links, refresh stale stats) and queue everything else for review.

Sources & further reading

1.  \[01\]
    
    [Helpful Content System](https://developers.google.com/search/docs/fundamentals/creating-helpful-content)
    
    Google Search Central
    
    DOC
2.  \[02\]
    
    [Content pruning research](https://ahrefs.com/blog/content-pruning/)
    
    Ahrefs
    
    STUDY

![Niko Alho](https://nikoalho.fi/assets/niko-alho-avatar-192.webp)

Niko Alho

I run agentic SEO and build custom AI for B2B companies. Based in Turku.

[About →](https://nikoalho.fi/about/)

KEEP READING

## More on agentic seo.

-   [![Editorial illustration for Agentic SEO: the operator playbook for autonomous search](https://nikoalho.fi/visuals/agentic-ai-seo.webp)
    
    Agentic SEO 2026 · 05 · 20
    
    Agentic SEO: the operator playbook for autonomous search
    
    Agentic SEO replaces passive tools with systems that execute search goals end-to-end. Here is how op…
    
    read →](https://nikoalho.fi/writing/agentic-ai-seo/)
-   [![Editorial illustration for The cost economics of agentic SEO at $40 per ranked post.](https://nikoalho.fi/visuals/agentic-seo-cost-economics.webp)
    
    Agentic SEO 2026 · 05 · 20
    
    The cost economics of agentic SEO at $40 per ranked post.
    
    Model the cost of agentic SEO per brief, draft, review, publish, and ranked page. Includes assumptio…
    
    read →](https://nikoalho.fi/writing/agentic-seo-cost-economics/)
-   [![Editorial illustration for AI content eval loops: pass-fail rubrics that scale](https://nikoalho.fi/visuals/ai-content-eval-loops.webp)
    
    Agentic SEO 2026 · 05 · 20
    
    AI content eval loops: pass-fail rubrics that scale
    
    Build AI content eval loops with pass-fail rubrics, source checks, style tests, and escalation rules…
    
    read →](https://nikoalho.fi/writing/ai-content-eval-loops/)

[More writing →](https://nikoalho.fi/writing/)

Direct with Niko · 20-min intro, no pitch [Book a slot →](https://nikoalho.fi/book/)

## Structured data

```json
{
  "@context": "https://schema.org",
  "@type": "WebSite",
  "@id": "https://nikoalho.fi/#website",
  "url": "https://nikoalho.fi/",
  "name": "Niko Alho",
  "description": "Agentic SEO and custom AI builds for B2B companies.",
  "inLanguage": "en",
  "publisher": {
    "@id": "https://nikoalho.fi/#person"
  },
  "potentialAction": {
    "@type": "SearchAction",
    "target": {
      "@type": "EntryPoint",
      "urlTemplate": "https://nikoalho.fi/search/?q={search_term_string}"
    },
    "query-input": "required name=search_term_string"
  }
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "Person",
  "@id": "https://nikoalho.fi/#person",
  "name": "Niko Alho",
  "givenName": "Niko",
  "familyName": "Alho",
  "url": "https://nikoalho.fi/about/",
  "image": "https://nikoalho.fi/og/default.png",
  "jobTitle": "Agentic SEO & Custom AI Consultant",
  "email": "mailto:contact@nikoalho.fi",
  "telephone": "+358401539426",
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Turku",
    "addressCountry": "FI"
  },
  "knowsAbout": [
    "Search Engine Optimization",
    "Agentic SEO",
    "Topical Authority",
    "Retrieval-Augmented Generation",
    "Large Language Models",
    "Custom AI Builds",
    "B2B SaaS Content Strategy",
    "Schema.org Structured Data",
    "Generative Engine Optimization"
  ],
  "knowsLanguage": [
    "en",
    "fi"
  ],
  "worksFor": {
    "@id": "https://nikoalho.fi/#organization"
  },
  "sameAs": [
    "https://www.linkedin.com/in/nikoalho/",
    "https://github.com/alhoniko"
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "ProfessionalService",
  "@id": "https://nikoalho.fi/#organization",
  "name": "Niko Alho — SEO & AI Automation",
  "alternateName": "Niko Alho",
  "description": "Agentic SEO and custom AI builds for B2B companies.",
  "url": "https://nikoalho.fi/",
  "image": "https://nikoalho.fi/og/default.png",
  "logo": "https://nikoalho.fi/assets/logo-mark.svg",
  "email": "mailto:contact@nikoalho.fi",
  "telephone": "+358401539426",
  "priceRange": "$$$",
  "founder": {
    "@id": "https://nikoalho.fi/#person"
  },
  "employee": {
    "@id": "https://nikoalho.fi/#person"
  },
  "knowsLanguage": [
    "en",
    "fi"
  ],
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Turku",
    "addressCountry": "FI"
  },
  "areaServed": [
    {
      "@type": "City",
      "name": "Turku"
    },
    {
      "@type": "City",
      "name": "Helsinki"
    },
    {
      "@type": "Country",
      "name": "Finland"
    },
    {
      "@type": "Place",
      "name": "European Union"
    },
    {
      "@type": "Place",
      "name": "Worldwide (remote)"
    }
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "Article",
  "@id": "https://nikoalho.fi/writing/llm-content-auditing/#article",
  "headline": "LLM content auditing: kill low-value pages at scale",
  "name": "LLM content auditing: kill low-value pages at scale",
  "description": "An LLM content audit reviews 10,000 legacy posts in hours, not weeks. The classification rubric, prompts, and prune-vs-refresh logic that frees real budget.",
  "image": "https://nikoalho.fi/og/llm-content-auditing.png",
  "url": "https://nikoalho.fi/writing/llm-content-auditing/",
  "datePublished": "2026-05-20T00:00:00.000Z",
  "dateModified": "2026-07-18T00:00:00.000Z",
  "inLanguage": "en",
  "isAccessibleForFree": true,
  "wordCount": 1636,
  "articleSection": "Agentic SEO",
  "keywords": "llm content audit, ai content audit, content audit automation, content pruning seo, ai seo audit",
  "author": {
    "@id": "https://nikoalho.fi/#person"
  },
  "publisher": {
    "@id": "https://nikoalho.fi/#person"
  },
  "mainEntityOfPage": {
    "@type": "WebPage",
    "@id": "https://nikoalho.fi/writing/llm-content-auditing/"
  },
  "about": {
    "@type": "Thing",
    "name": "Agentic SEO"
  },
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [
      "h1",
      ".tldr",
      ".article-body > .prose > p:first-of-type"
    ]
  },
  "citation": [
    {
      "@type": "CreativeWork",
      "url": "https://www.linkedin.com/in/nikoalho/"
    }
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "Home",
      "item": "https://nikoalho.fi/"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "Writing",
      "item": "https://nikoalho.fi/writing/"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "LLM content auditing: kill low-value pages at scale",
      "item": "https://nikoalho.fi/writing/llm-content-auditing/"
    }
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "Won't deleting content hurt my SEO?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "The opposite, usually. The Helpful Content System penalizes site-wide quality signals when thin posts drag down average. Pruning often lifts rankings on remaining pages."
      }
    },
    {
      "@type": "Question",
      "name": "What LLM works best for content auditing?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Claude or GPT-5 for reasoning depth, with structured output (JSON schemas) for downstream automation. Avoid temperature > 0.3 for audit verdicts."
      }
    },
    {
      "@type": "Question",
      "name": "How do I prevent the LLM from being too aggressive?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Set conservative thresholds: only delete if the page has zero traffic in 12 months, zero backlinks, and doesn't match any cluster. Everything ambiguous goes to refresh, not delete."
      }
    },
    {
      "@type": "Question",
      "name": "Can I trust an LLM to grade my content without human review?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Not for delete decisions. The agent's job is to surface candidates with reasoning; the human still approves destructive actions. Auto-apply only the safe verdicts (e.g. add internal links, refresh stale stats) and queue everything else for review."
      }
    }
  ]
}
```