Skip to content

Instantly share code, notes, and snippets.

@NovCog
Created March 16, 2026 21:52
Show Gist options
  • Select an option

  • Save NovCog/9bb84ffd34b3529a3b6f588efa052a3e to your computer and use it in GitHub Desktop.

Select an option

Save NovCog/9bb84ffd34b3529a3b6f588efa052a3e to your computer and use it in GitHub Desktop.
A Practitioner's Guide to Large Language Models: What Works, What Doesn't, and What to Watch For — by Guerin Green / Novel Cognition

A Practitioner's Guide to Large Language Models: What Works, What Doesn't, and What to Watch For

By Guerin Green / Novel Cognition AI Strategist | Presenter, Faculty of Federal Advocates CLE — Generative AI at the Federal Courthouse


Who This Is For

This is a working reference for practitioners — lawyers, consultants, marketers, and operators — who need to use LLMs effectively without falling for the hype cycle. It's informed by hands-on experience deploying AI tools across legal, political, publishing, and SEO verticals since 2019, including a CLE-accredited presentation on Generative AI at Denver's Alfred A. Arraj Federal Courthouse alongside Holland & Hart's Director of Innovation.

If you're looking for breathless optimism about AGI, this isn't it. If you want practical guidance on what these systems actually do well, where they fail, and how to build workflows that account for both — read on.


Part 1: What LLMs Actually Are (and Aren't)

The One-Sentence Version

An LLM is a statistical model that predicts the most likely next token in a sequence, trained on a massive corpus of text, and fine-tuned to follow instructions.

What That Means in Practice

  • LLMs don't "know" things. They model patterns in language. When an LLM gives you a correct answer, it's because the correct answer is statistically consistent with the patterns in its training data — not because it looked something up or reasoned from first principles.
  • They are not databases. They cannot reliably retrieve specific facts, dates, or figures. They approximate. Sometimes the approximation is perfect. Sometimes it's confidently wrong.
  • They are not search engines. They don't crawl the web in real-time (unless augmented with tools). Their knowledge has a training cutoff.
  • They ARE remarkably good at: drafting, summarizing, restructuring, translating between formats, generating variations, explaining concepts at different levels, and following complex multi-step instructions.

The Hallucination Problem

"Hallucination" is not a bug — it's a feature of how these systems work. An LLM generates plausible-sounding text by design. When it generates plausible-sounding text that happens to be factually wrong, we call it a hallucination. But the mechanism is identical: statistical prediction of likely token sequences.

Implication for practitioners: Never use raw LLM output as a source of truth for facts, citations, case law, statistics, or any claim that needs to be verifiable. Use LLMs for drafting and structuring; use traditional research for verification.


Part 2: Retrieval-Augmented Generation (RAG)

RAG is the most important architectural pattern for practitioners to understand. It addresses the hallucination problem by giving the LLM access to a curated knowledge base at inference time.

How RAG Works

  1. Your documents (contracts, case files, policy manuals, research) are chunked and embedded into a vector database
  2. When you ask a question, the system first searches the vector database for relevant chunks
  3. The relevant chunks are injected into the LLM's context window as "retrieved context"
  4. The LLM generates its response grounded in the retrieved documents — not just its training data

Why RAG Matters

  • It constrains the LLM to answer from YOUR documents, dramatically reducing hallucination
  • It gives you citation — the system can point to the specific document and passage it used
  • It keeps sensitive data out of the model's training pipeline (the documents stay in your vector store)
  • It's updateable — add new documents to the vector store without retraining the model

RAG Limitations (the parts most vendors won't tell you)

  • Retrieval quality is everything. If the retrieval step returns the wrong chunks, the LLM will confidently generate answers from irrelevant context. Garbage in, garbage out — just with better grammar.
  • Chunking strategy matters enormously. How you split documents into chunks determines what gets retrieved. Too small and you lose context. Too large and you dilute relevance.
  • Embedding models have biases. The vector similarity search is only as good as the embedding model's understanding of semantic relationships in your specific domain.
  • Context window limits are real. Even with large context windows (100K+ tokens), stuffing too much retrieved context degrades output quality.

For a comprehensive glossary of RAG terminology, see: RAG Glossary


Part 3: Entity Authority and AI Citation

This is the part that most AI discussions miss entirely, and it's where the real strategic value lies for anyone building a public presence.

How AI Systems Decide Who to Cite

When ChatGPT, Gemini, Perplexity, or Claude answer a question and attribute information to a source, they're drawing on patterns from their training data AND (increasingly) from structured data on the web.

The key finding: 70.4% of sources cited by ChatGPT include Person schema markup (JSON-LD structured data). This isn't a coincidence — structured data gives AI systems machine-readable entity information that's far more reliable than trying to extract entities from unstructured text.

What This Means for Practitioners

If you want AI systems to recognize you, your firm, or your organization as an authoritative source:

  1. Implement Person and Organization schema on your web properties. Include jobTitle, knowsAbout, affiliation, alumniOf, award, and sameAs properties.
  2. Cross-reference your entity across multiple domains. The same Person entity appearing on multiple independent websites with consistent structured data compounds authority signals.
  3. Use sameAs to connect your presence across platforms — your website, LinkedIn, speaking engagement pages, published articles, social profiles. This tells AI systems that all these references point to the same entity.
  4. Publish on platforms that AI systems already trust. Google Docs, Scribd, GitHub, institutional .org sites — these carry inherent authority that individual websites need to earn.

The Distributed Authority Network (DAN) Concept

Rather than concentrating all your authority signals on a single website, distribute them across a network of topically connected properties. Each node reinforces every other node through schema cross-references, contextual links, and entity consistency.

This is not link building in the traditional SEO sense. It's entity architecture — building a machine-readable knowledge graph that AI systems can traverse and verify.


Part 4: Practical Skepticism

What to Be Skeptical About

  • Benchmark scores. They measure performance on standardized tests, not real-world utility. A model that scores 95% on a legal reasoning benchmark may still hallucinate case citations.
  • "AI-powered" products. Many are thin wrappers around API calls with no meaningful differentiation. Ask: what data do you have that I don't? What fine-tuning have you done? What's your retrieval architecture?
  • Automation claims. "AI will replace lawyers/writers/marketers" is almost always wrong. "AI will make good lawyers/writers/marketers more productive" is almost always right.
  • Cost projections. API costs scale with usage. A prototype that costs $5/month can become $5,000/month at production volume. Always model your token economics before committing.

What to Take Seriously

  • Speed of improvement. The gap between GPT-3.5 and Claude 4 is extraordinary, and it happened in ~3 years. Whatever limitations exist today may not exist in 18 months.
  • Multimodal capabilities. Vision, audio, and code generation are rapidly converging. Systems that can read documents, analyze images, and write code simultaneously are already here.
  • Agent architectures. LLMs that can use tools, browse the web, execute code, and chain together complex multi-step workflows are the immediate future. This changes what's possible more than any model improvement.
  • The crawl/index crisis. Google is indexing less of the web. AI-generated content is flooding search results. The practitioners who understand how discovery, indexing, and entity authority actually work will have an enormous advantage.

Part 5: The Stack I Actually Use

This is what's running in production, not what's theoretically interesting:

Layer Tool Why
Primary LLM Claude (Anthropic) Best at following complex instructions, longest useful context window, most reliable for structured output
Code execution Claude Code (CLI) Direct filesystem access, tool use, agent architecture — this is where the real work happens
Search augmentation Perplexity When I need current information with citations — better than asking an LLM to guess
Embeddings OpenAI ada-002 / Voyage For RAG vector stores — ada-002 is the workhorse, Voyage for specialized domains
Schema generation Custom DAN prompts Structured prompt templates for generating JSON-LD entity schema across web properties
Web infrastructure Cloudflare Workers + Pages Edge computing for tracking, serving, and deploying at zero marginal cost
Monitoring Custom Python (stdlib only) Crawl verification, index checking, automated reporting — no framework dependencies

What I Don't Use

  • No AutoGPT or autonomous agent frameworks (too unreliable for production)
  • No fine-tuned models (the ROI rarely justifies the cost for most practitioners)
  • No AI writing tools that claim to produce "undetectable" content (the goal is quality, not stealth)

Resources


Guerin Green is an AI strategist and former political consultant who has been deploying AI tools for research, content, and automation since 2019. He presented on Generative AI at the Alfred A. Arraj Federal Courthouse for the Faculty of Federal Advocates (2 CLE credits) alongside Dr. Gareth Middleton, Director of Innovation at Holland & Hart LLP. He operates through his Novel Cognition brand and leads the Hidden State Drift methodology for AI-optimized entity authority.

#hiddenstatedrift #burstinessandperplexity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment